chore: version packages

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -1,5 +1,26 @@
 # aicodeman
 
+## 1.11.2
+
+### Patch Changes
+
+- Make Antigravity (`agy`) a first-class CLI everywhere, and stop presenting Gemini CLI as a consumer product now that it is enterprise-only.
+
+  Antigravity was already wired into the session layer, schemas, run-mode menu and remote/Docker command maps, but the surfaces around it were never updated. Gemini keeps full support; Antigravity now sits beside it.
+
+  Fixes:
+  - **Docker cases with `mode: 'antigravity'` were broken.** `docker/agent.Dockerfile` installs its CLIs from npm, and `agy` is not an npm package, so the binary was never in the image and the container died on command-not-found. It now gets its own installer step. The `--dir /usr/local/bin` flag is load-bearing: the installer's default `$HOME/.local/bin` resolves to root's home at build time and would be unreachable by the `agent` user the container runs as. Note the binary is roughly 190MB, making it the largest layer in the image, so rebuild with `node scripts/build-agent-image.mjs` when convenient.
+  - **Welcome screen** gained a "Run Antigravity" action, gated on `agy` being present like the other CLI buttons, styled with the same cyan identity as the toolbar run button and run-mode dot.
+  - **`install.sh`** now detects `agy` (search paths mirroring `antigravity-cli-resolver.ts`), counts it as a satisfying AI CLI so an Antigravity-only box is not told it has none, and recommends it instead of Gemini in the install hints.
+
+  Documentation corrections where it had become factually wrong: `architecture-invariants.md` described `isExternalCliMode()` as opencode/codex/gemini when the code has included antigravity for some time, said "all three modes", and omitted `ANTIGRAVITY_*` from the env-prefix allowlist row; the `agentType` enum in `cron-guide.md`, `SessionMode` in `cron-discovery.md`, and `RemoteCommandMode` in `remote-sessions.md` were all stale.
+
+  Also updated both READMEs (five CLIs, Gemini marked enterprise-only), the `antigravity` npm keyword, and comment drift in eight places. Test coverage added for the new welcome button.
+
+  Antigravity stores its state under `~/.gemini/antigravity-cli/` rather than a `~/.antigravity` directory, so the existing `.gemini` Docker credential seed already covers it. That is now recorded in a code comment so no dead configuration gets added later.
+
+- b982c5d: Keep the brief Response Viewer output inside the same message card and Markdown wrapper used by the full conversation view, so opening the viewer without clicking More preserves the same readable formatting.
+
 ## 1.11.1
 
 ### Patch Changes

CLAUDE.md

@@ -47,7 +47,7 @@ The production server caches static files for 1 year, `immutable` (`maxAge: '1y'
 
 ## COM Shorthand (Deployment)
 
-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 `.github/SECURITY.md`.
+Uses [Semantic Versioning](https://semver.org/) (`MAJOR.MINOR.PATCH`) via `@changesets/cli`. What SemVer actually covers (the CLI, documented env vars, **and the HTTP/SSE API under `/api/v1`**: endpoint paths, response envelope, `errorCode` values and SSE event names are public/stable; on-disk state, internal TS modules, and experimental features are internal/unstable) is defined in `docs/versioning-policy.md`. Third-party integration surfaces are documented in `docs/extending-codeman.md`. Security reporting + known limitations live in `.github/SECURITY.md`.
 
 When user says "COM":
 
@@ -74,7 +74,7 @@ When user says "COM":
 
 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**: 1.11.1 (must match `package.json`)
+**Version**: 1.11.2 (must match `package.json`)
 
 ## Project Overview
 

README.md

@@ -5,7 +5,7 @@
 <h2 align="center">Mission control for AI coding agents</h2>
 
 <p align="center">
-  <em>Claude Code &bull; OpenCode &bull; Codex &bull; Gemini &bull; Terminal - One Dashboard &bull; Any Device</em>
+  <em>Claude Code &bull; OpenCode &bull; Codex &bull; Antigravity &bull; Gemini &bull; Terminal - One Dashboard &bull; Any Device</em>
 </p>
 
 <p align="center">
@@ -27,7 +27,7 @@
   <img src="docs/images/subagent-demo-20260724.gif" alt="Codeman — parallel subagent visualization" width="900">
 </p>
 
-**Codeman** is a self-hosted mission control for AI coding agents. It spawns Claude Code, OpenCode, Codex, or Gemini CLI inside persistent tmux sessions, streams the real terminal to any browser, and keeps agents productive after you walk away: it re-prompts on idle, resumes when a usage limit resets, runs scheduled jobs, and shows every background agent working in real time.
+**Codeman** is a self-hosted mission control for AI coding agents. It spawns Claude Code, OpenCode, Codex, Antigravity, or Gemini CLI inside persistent tmux sessions, streams the real terminal to any browser, and keeps agents productive after you walk away: it re-prompts on idle, resumes when a usage limit resets, runs scheduled jobs, and shows every background agent working in real time.
 
 Get started in one line (macOS & Linux, Windows via WSL):
 
@@ -42,7 +42,7 @@ codeman web
 
 The installer asks before every system change, and re-running the same line updates in place. Full details: [Quick Start - Installation](#quick-start---installation).
 
-- **One dashboard, four CLIs** - run [Claude Code, OpenCode, Codex, or Gemini](#more-features) per session (plus plain shell), locally, [in Docker](#isolated-docker-sessions), or [over SSH](#remote-ssh-sessions)
+- **One dashboard, five CLIs** - run [Claude Code, OpenCode, Codex, Antigravity, or Gemini](#more-features) per session (plus plain shell), locally, [in Docker](#isolated-docker-sessions), or [over SSH](#remote-ssh-sessions)
 - **Truly phone-friendly** - a [touch-optimized terminal](#mobile-optimized-web-ui) with instant local echo, QR login, swipe navigation, and push notifications
 - **Runs while you sleep** - [idle detection + respawn cycling](#respawn-controller) and auto-resume when a subscription limit resets, for 24+ hour unattended runs
 - **See your agents think** - [live floating windows](#live-agent-visualization) for every subagent and teammate, with real-time transcripts
@@ -68,7 +68,7 @@ This installs Node.js and tmux if missing, clones Codeman to `~/.codeman/app`, a
 - **Re-run to update.** The same one-liner updates a finished install in place: local changes in `~/.codeman/app` are stashed (never discarded), and a running service is restarted and verified. If a first install was interrupted, re-running resumes the full setup instead. `install.sh update` and `install.sh uninstall` also exist.
 - **CI / headless:** without a terminal attached, steps that would change your system abort with instructions instead of running silently. Set `CODEMAN_NONINTERACTIVE=1` to approve them for automation.
 
-You'll need at least one AI coding CLI installed — [Claude Code](https://docs.anthropic.com/en/docs/claude-code), [OpenCode](https://opencode.ai), [Codex](https://developers.openai.com/codex/cli), or [Gemini CLI](https://github.com/google-gemini/gemini-cli) (any combination works). The installer detects whichever of the four is present; if none is found, it offers to install Claude Code or OpenCode, or you can skip and install one yourself later. 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), [Codex](https://developers.openai.com/codex/cli), [Antigravity](https://antigravity.google), or [Gemini CLI](https://github.com/google-gemini/gemini-cli) (any combination works; Gemini CLI is enterprise-only since Google's consumer cutover, and Antigravity is its successor). The installer detects whichever of the five is present; if none is found, it offers to install Claude Code or OpenCode, or you can skip and install one yourself later. After install:
 
 ```bash
 codeman web
@@ -151,7 +151,7 @@ launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.codeman.web.plist
 wsl bash -c "curl -fsSL https://getcodeman.com/install | 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), [OpenCode](https://opencode.ai), [Codex](https://developers.openai.com/codex/cli), or [Gemini CLI](https://github.com/google-gemini/gemini-cli)). 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), [Codex](https://developers.openai.com/codex/cli), [Antigravity](https://antigravity.google), or [Gemini CLI](https://github.com/google-gemini/gemini-cli)). After installing, `http://localhost:3000` is accessible from your Windows browser.
 
 </details>
 
@@ -231,7 +231,7 @@ Click **+ New Session** (or **Quick Start**). A session is one AI CLI running in
 | Field                        | What it does                                                                                                        |
 | ---------------------------- | ------------------------------------------------------------------------------------------------------------------- |
 | **Working directory / case** | The folder the agent operates in. A "case" is just a named working dir Codeman remembers.                           |
-| **CLI / run mode**           | `Claude` (default), `OpenCode`, `Codex`, `Gemini`, or `Terminal` (plain shell).                                     |
+| **CLI / run mode**           | `Claude` (default), `OpenCode`, `Codex`, `Antigravity`, `Gemini`, or `Terminal` (plain shell).                       |
 | **Model**                    | Per-session model (App Settings → Claude Model). A soft default — `/model` still works in-session.                  |
 | **Effort / Ultracode**       | Reasoning effort (`low`–`max`) or `ultracode` for dynamic multi-agent workflows. Switchable anytime with `/effort`. |
 
@@ -404,7 +404,7 @@ 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**, **Codex**, or **Gemini** per session; env-var prefixes auto-gate (`CLAUDE_CODE_*` vs `OPENCODE_*` vs `CODEX_*` vs `GEMINI_*`/`GOOGLE_*`). See [`docs/opencode-integration.md`](docs/opencode-integration.md)
+- **Multi-CLI** — run **Claude Code**, **OpenCode**, **Codex**, **Antigravity**, or **Gemini** per session; env-var prefixes auto-gate (`CLAUDE_CODE_*` vs `OPENCODE_*` vs `CODEX_*` vs `ANTIGRAVITY_*` vs `GEMINI_*`/`GOOGLE_*`). See [`docs/opencode-integration.md`](docs/opencode-integration.md)
 - **Docker sessions** — run a case inside an isolated, hardened container. One checkbox on **Create New** spins up a container with sensible defaults and starts the agent inside it; multiple sessions share one per-case container; export a container + its workspace to a portable `.tar.gz` to move it to another machine. See [`docs/docker-cases.md`](docs/docker-cases.md)
 - **Remote SSH sessions** — point a case at another machine and run the agent there inside a durable remote tmux: survives SSH drops, auto-reconnects, and can discover + attach sessions already running on the host. See [`docs/remote-sessions.md`](docs/remote-sessions.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
@@ -426,7 +426,7 @@ Run a case inside its own hardened Docker container instead of directly on your
 - **Resource templates** — expand the checkbox for a **Small / Medium / Large / GPU** preset (memory, CPUs, GPU), or set your own. **Disk is elastic** — storage grows as data flows in, no fixed cap.
 - **Shared per-case container** — many sessions can `docker exec` into the same container; killing one session never tears the container out from under the others.
 - **Hardened by default** — non-root, `--cap-drop ALL`, `no-new-privileges`, PID/memory caps, never `--privileged` or the docker socket; a **sealed** profile (no host credentials, network off) is one toggle away.
-- **Seamless auth, isolated credentials** — your host Claude / Codex / Gemini / OpenCode logins work inside the container out of the box: credentials are seeded (copied) in at launch and onboarding/trust prompts are pre-answered, so no login wizard appears. The container keeps its own copies and never writes back to your host credential stores; only conversation transcripts are shared, and exports never capture secrets.
+- **Seamless auth, isolated credentials** — your host Claude / Codex / Antigravity / Gemini / OpenCode logins work inside the container out of the box: credentials are seeded (copied) in at launch and onboarding/trust prompts are pre-answered, so no login wizard appears. The container keeps its own copies and never writes back to your host credential stores; only conversation transcripts are shared, and exports never capture secrets.
 - **Move it to another machine** — export a container's whole environment (toolchain + workspace) to a portable `.tar.gz`, `docker load` it on the other side, and import it into a fresh case.
 - **Durable** — reconnect after a restart lands back in the same live agent; a container stop/reboot resumes the conversation from the bind-mounted transcript.
 
@@ -612,7 +612,7 @@ These run for **every** request — before auth, even on the default no-password
 
 ### Input, files & headers
 
-- **Schema-validated inputs** — every API body is checked with Zod v4 schemas; a `CLAUDE_CODE_*` / `OPENCODE_*` / `CODEX_*` / `GEMINI_*` / `GOOGLE_*` env-prefix allowlist gates which settings each CLI can receive
+- **Schema-validated inputs** — every API body is checked with Zod v4 schemas; a `CLAUDE_CODE_*` / `OPENCODE_*` / `CODEX_*` / `ANTIGRAVITY_*` / `GEMINI_*` / `GOOGLE_*` 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`
 
@@ -844,7 +844,7 @@ flowchart TB
         end
 
         subgraph External["External"]
-            CLI["AI CLI<br/><small>Claude Code / OpenCode / Codex / Gemini</small>"]
+            CLI["AI CLI<br/><small>Claude Code / OpenCode / Codex / Antigravity / Gemini</small>"]
             BG["Background Agents<br/><small>(Task tool)</small>"]
         end
     end

README.zh-CN.md

@@ -5,7 +5,7 @@
 <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 &bull; Gemini &bull; 终端 —— 统一仪表盘 &bull; 任意设备</em>
+  <em>Claude Code &bull; OpenCode &bull; Codex &bull; Antigravity &bull; Gemini &bull; 终端 —— 统一仪表盘 &bull; 任意设备</em>
 </p>
 
 <p align="center">
@@ -58,7 +58,7 @@ curl -fsSL https://getcodeman.com/install | bash
 - **重跑即更新。** 再次运行同一条命令即可原地更新已完成的安装：`~/.codeman/app` 中的本地改动会被 stash（绝不丢弃），运行中的服务会自动重启并校验。若首次安装中途失败，重跑会继续完成完整的安装流程。也可以使用 `install.sh update` 与 `install.sh uninstall`。
 - **CI / 无终端环境：** 没有终端时，涉及系统改动的步骤会带着说明中止，而不是静默执行；在自动化场景设置 `CODEMAN_NONINTERACTIVE=1` 即可批准这些步骤。
 
-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)、[Codex](https://developers.openai.com/codex/cli) 或 [Gemini CLI](https://github.com/google-gemini/gemini-cli)（任意组合均可）。安装器会自动检测这四个中已安装的任意一个；若一个都没有，会提供安装 Claude Code 或 OpenCode 的选项，也可以选择跳过、稍后自行安装。安装完成后：
+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)、[Codex](https://developers.openai.com/codex/cli)、[Antigravity](https://antigravity.google) 或 [Gemini CLI](https://github.com/google-gemini/gemini-cli)（任意组合均可；自 Google 面向消费者停售后，Gemini CLI 仅限企业版，Antigravity 是其继任者）。安装器会自动检测这五个中已安装的任意一个；若一个都没有，会提供安装 Claude Code 或 OpenCode 的选项，也可以选择跳过、稍后自行安装。安装完成后：
 
 ```bash
 codeman web
@@ -141,7 +141,7 @@ launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.codeman.web.plist
 wsl bash -c "curl -fsSL https://getcodeman.com/install | 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)、[Codex](https://developers.openai.com/codex/cli) 或 [Gemini CLI](https://github.com/google-gemini/gemini-cli)）。安装完成后，即可从 Windows 浏览器访问 `http://localhost:3000`。
+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)、[Codex](https://developers.openai.com/codex/cli)、[Antigravity](https://antigravity.google) 或 [Gemini CLI](https://github.com/google-gemini/gemini-cli)）。安装完成后，即可从 Windows 浏览器访问 `http://localhost:3000`。
 
 </details>
 
@@ -221,7 +221,7 @@ codeman web -H 0.0.0.0            # 绑定局域网 —— 必须设置 CODEMAN_
 | 字段                   | 作用                                                                                        |
 | ---------------------- | ------------------------------------------------------------------------------------------- |
 | **工作目录 / case**    | 智能体操作的文件夹。「case」就是一个 Codeman 记住的命名工作目录。                           |
-| **CLI / 运行模式**     | `Claude`（默认）、`OpenCode`、`Codex`、`Gemini` 或 `Terminal`（普通 shell）。               |
+| **CLI / 运行模式**     | `Claude`（默认）、`OpenCode`、`Codex`、`Antigravity`、`Gemini` 或 `Terminal`（普通 shell）。 |
 | **模型**               | 每会话模型（App Settings → Claude Model）。软默认值 —— 会话内 `/model` 依然有效。           |
 | **Effort / Ultracode** | 推理力度（`low`–`max`），或用 `ultracode` 开启动态多智能体工作流。随时可用 `/effort` 切换。 |
 
@@ -394,7 +394,7 @@ PTY 输出 → 16ms 服务端批处理 → DEC 2026 包裹 → SSE → 客户端
 ## 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**、**Codex** 或 **Gemini**；环境变量前缀自动隔离（`CLAUDE_CODE_*`、`OPENCODE_*`、`CODEX_*` 与 `GEMINI_*`/`GOOGLE_*`）。详见 [`docs/opencode-integration.md`](docs/opencode-integration.md)
+- **多 CLI** —— 每个会话可选 **Claude Code**、**OpenCode**、**Codex**、**Antigravity** 或 **Gemini**；环境变量前缀自动隔离（`CLAUDE_CODE_*`、`OPENCODE_*`、`CODEX_*`、`ANTIGRAVITY_*` 与 `GEMINI_*`/`GOOGLE_*`）。详见 [`docs/opencode-integration.md`](docs/opencode-integration.md)
 - **Docker 会话** —— 在隔离且加固的容器中运行案例。**Create New** 上勾选一个复选框即可用合理的默认值启动容器并在其中启动智能体；同一案例的多个会话共享一个容器；可将容器连同工作区导出为可移植的 `.tar.gz`，迁移到另一台机器。详见 [`docs/docker-cases.md`](docs/docker-cases.md)
 - **远程 SSH 会话**：把案例指向另一台机器，让智能体在那里一个持久的远程 tmux 中运行：SSH 断连不中断任务、自动重连，还能发现并附着主机上已在运行的会话。详见 [`docs/remote-sessions.md`](docs/remote-sessions.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
@@ -416,7 +416,7 @@ PTY 输出 → 16ms 服务端批处理 → DEC 2026 包裹 → SSE → 客户端
 - **资源模板** —— 展开复选框可选 **Small / Medium / Large / GPU** 预设（内存、CPU、GPU），也可以完全自定义。**磁盘是弹性的** —— 存储随数据增长，没有固定上限。
 - **按案例共享容器** —— 多个会话可以 `docker exec` 进同一个容器；结束某个会话绝不会影响其他会话所在的容器。
 - **默认加固** —— 非 root、`--cap-drop ALL`、`no-new-privileges`、PID/内存上限，绝不使用 `--privileged` 或 docker socket；**密封（sealed）** 配置（不注入主机凭据、关闭网络）只需一个开关。
-- **无感认证、凭据隔离** —— 主机上的 Claude / Codex / Gemini / OpenCode 登录在容器内开箱即用：凭据在启动时以只读种子方式复制注入，onboarding/信任提示已预先答复，不会弹出登录向导。容器保留自己的副本，绝不回写主机的凭据存储；跨边界共享的只有对话转录，导出文件也绝不包含机密。
+- **无感认证、凭据隔离** —— 主机上的 Claude / Codex / Antigravity / Gemini / OpenCode 登录在容器内开箱即用：凭据在启动时以只读种子方式复制注入，onboarding/信任提示已预先答复，不会弹出登录向导。容器保留自己的副本，绝不回写主机的凭据存储；跨边界共享的只有对话转录，导出文件也绝不包含机密。
 - **迁移到另一台机器** —— 把容器的完整环境（工具链 + 工作区）导出为可移植的 `.tar.gz`，在另一台机器上导入到新案例即可继续。
 - **持久耐用** —— Codeman 重启后重连会回到同一个存活的智能体；容器停止/重启后则从绑定挂载的转录恢复对话。
 
@@ -602,7 +602,7 @@ Codeman 默认用 `--dangerously-skip-permissions` 启动会话，因此 Web UI
 
 ### 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_*` / `GEMINI_*` / `GOOGLE_*` 环境变量前缀允许列表把控每个 CLI 能接收哪些设置
+- **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_*` / `ANTIGRAVITY_*` / `GEMINI_*` / `GOOGLE_*` 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
 
@@ -834,7 +834,7 @@ flowchart TB
         end
 
         subgraph External["外部"]
-            CLI["AI CLI<br/><small>Claude Code / OpenCode / Codex / Gemini</small>"]
+            CLI["AI CLI<br/><small>Claude Code / OpenCode / Codex / Antigravity / Gemini</small>"]
             BG["后台智能体<br/><small>(Task 工具)</small>"]
         end
     end

docker/agent.Dockerfile

@@ -26,8 +26,8 @@ RUN apt-get update \
       openssh-client \
  && rm -rf /var/lib/apt/lists/*
 
-# The agent CLIs (all four backends Codeman supports). Pinning is left to the
-# rebuild cadence (see docs/docker-cases-plan.md, user-decision 2).
+# The npm-published agent CLIs. Pinning is left to the rebuild cadence (see
+# docs/docker-cases-plan.md, user-decision 2).
 RUN npm install -g \
       @anthropic-ai/claude-code \
       @openai/codex \
@@ -35,6 +35,15 @@ RUN npm install -g \
       opencode-ai \
  && npm cache clean --force
 
+# Antigravity (`agy`) is NOT on npm — Google ships a standalone binary through its
+# own installer, so it needs its own step. `--dir /usr/local/bin` is load-bearing:
+# the installer's default target is `$HOME/.local/bin`, which at build time is
+# root's home and would be unreachable by the `agent` user the container runs as.
+# ⚠️ This binary is ~190MB on its own; it is the single largest layer in the image.
+RUN curl -fsSL https://antigravity.google/cli/install.sh | bash -s -- --dir /usr/local/bin \
+ && chmod 755 /usr/local/bin/agy \
+ && agy --version
+
 # `agent` user (gid 0) with an arbitrary-uid-writable HOME. The uid is
 # auto-assigned (node:22-slim already occupies uid 1000 with its `node` user); at
 # runtime Codeman overrides with `--user <hostUid>:0` on Linux, so the baked uid
@@ -50,7 +59,8 @@ ENV HOME=/home/agent
 # dirs: tokens/settings/config are seeded in as writable copies and each CLI's runtime
 # state (backups, tasks, refreshed tokens) stays container-local, while ONLY the shared
 # transcript/rollout dirs (`.claude/projects`, `.codex/sessions`) are bind-mounted from
-# the host. (gemini/gcloud/opencode are whole seed-copies and need no pre-created dir.)
+# the host. (gemini/gcloud/opencode are whole seed-copies and need no pre-created dir;
+# Antigravity nests its state inside `.gemini/antigravity-cli`, so it rides that seed.)
 RUN useradd -g 0 -m -d /home/agent -s /bin/bash agent \
  && mkdir -p /home/agent/.npm /home/agent/.cache /home/agent/.config /home/agent/.codeman \
       /home/agent/.claude/projects /home/agent/.codex/sessions \

docs/architecture-invariants.md

@@ -18,9 +18,9 @@ Implementation detail extracted from `CLAUDE.md` so that file stays small enough
 
 ## Session launch modes
 
-### External CLI modes (OpenCode, Codex, Gemini)
+### External CLI modes (OpenCode, Codex, Gemini, Antigravity)
 
-**External CLI modes (OpenCode, Codex, Gemini)**: `isExternalCliMode()` in `session.ts` (`mode === 'opencode' || 'codex' || 'gemini'`) 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). All three modes **require tmux — no direct PTY fallback** — because secrets are injected via `tmux setenv` (socket-scoped `${this.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`), Gemini gets `GEMINI_API_KEY`/`GOOGLE_API_KEY`/`GOOGLE_CLOUD_PROJECT`/`GOOGLE_APPLICATION_CREDENTIALS`/`GOOGLE_GENAI_USE_VERTEXAI` etc. (`setGeminiEnvVars`, all 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). Gemini specifics: command built by `buildGeminiCommand()` (`--skip-trust` always, `--approval-mode <default|auto_edit|yolo|plan>` defaulting to `yolo` for parity with Claude's `--dangerously-skip-permissions`, `--model`, `--resume` from the `geminiConfig` payload); availability via `GET /api/gemini/status` — session/quick-start routes fail with `OPERATION_FAILED` + install hint (`npm install -g @google/gemini-cli`) when missing. Codex AND Gemini export `COLORTERM=truecolor` + unset `NO_COLOR` (other modes unset `COLORTERM`); Gemini joins `isAltScreenStripMode()` (Codex/Claude/Gemini are Ink TUIs that repaint inline → strip alt-screen/`3J` so scrollback survives). Codex availability via `GET /api/codex/status`. Frontend: run-mode dropdown → `runCodex()`/`runGemini()` in `session-ui.js` ("Run CX"/"Run GM" labels), App Settings → Codex CLI tab; Respawn/Ralph options are Claude-only, so session options open on the Summary tab for external CLI sessions. ⚠️ `run*()` MUST unwrap the `{success,data}` envelope (`(await res.json()).data.available` / `data.data.sessionId`) — reading the raw shape silently breaks the run. Tests: `test/run-mode-ui.test.ts` + `test/gemini-mode.test.ts` (vm-sandbox harness, no real DOM).
+**External CLI modes (OpenCode, Codex, Gemini, Antigravity)**: `isExternalCliMode()` in `session.ts` (`mode === 'opencode' || 'codex' || 'gemini' || 'antigravity'`) 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). All four modes **require tmux — no direct PTY fallback** — because secrets are injected via `tmux setenv` (socket-scoped `${this.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`), Gemini gets `GEMINI_API_KEY`/`GOOGLE_API_KEY`/`GOOGLE_CLOUD_PROJECT`/`GOOGLE_APPLICATION_CREDENTIALS`/`GOOGLE_GENAI_USE_VERTEXAI` etc. (`setGeminiEnvVars`, all 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). Gemini specifics: command built by `buildGeminiCommand()` (`--skip-trust` always, `--approval-mode <default|auto_edit|yolo|plan>` defaulting to `yolo` for parity with Claude's `--dangerously-skip-permissions`, `--model`, `--resume` from the `geminiConfig` payload); availability via `GET /api/gemini/status` — session/quick-start routes fail with `OPERATION_FAILED` + install hint (`npm install -g @google/gemini-cli`) when missing. Codex AND Gemini export `COLORTERM=truecolor` + unset `NO_COLOR` (other modes unset `COLORTERM`); Gemini joins `isAltScreenStripMode()` (Codex/Claude/Gemini are Ink TUIs that repaint inline → strip alt-screen/`3J` so scrollback survives). Codex availability via `GET /api/codex/status`. Antigravity specifics: command built by `buildAntigravityCommand()` (`--model`, `--conversation <id>` resume, `--dangerously-skip-permissions` from the `antigravityConfig` payload); availability via `GET /api/antigravity/status` — routes fail with `OPERATION_FAILED` + install hint (`curl -fsSL https://antigravity.google/cli/install.sh | bash`) when missing. Unlike the other three it is NOT an npm package (standalone binary, `~/.local/bin/agy`), which is why `docker/agent.Dockerfile` installs it with its own `--dir /usr/local/bin` step rather than in the `npm install -g` line, and why it does NOT join `isAltScreenStripMode()`. Frontend: run-mode dropdown → `runCodex()`/`runGemini()` in `session-ui.js` ("Run CX"/"Run GM" labels), App Settings → Codex CLI tab; Respawn/Ralph options are Claude-only, so session options open on the Summary tab for external CLI sessions. ⚠️ `run*()` MUST unwrap the `{success,data}` envelope (`(await res.json()).data.available` / `data.data.sessionId`) — reading the raw shape silently breaks the run. Tests: `test/run-mode-ui.test.ts` + `test/gemini-mode.test.ts` (vm-sandbox harness, no real DOM).
 
 ### Remote sessions over SSH
 
@@ -216,7 +216,7 @@ Copy goes through `_copyText()` (Clipboard API, then hidden-textarea + `execComm
 | **Rate limit**    | 10 failed auth/IP → 429 (15min decay). QR has separate limiter                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
 | **Hook bypass**   | `/api/hook-event` (and `/api/status-telemetry`, the statusLine exporter) skip Basic auth (localhost-only, schema-validated). When auth is active (`CODEMAN_PASSWORD` set), the loopback bypass requires the per-instance `X-Codeman-Hook-Secret` header **unconditionally** — COD-54 introduced it tunnel-gated; COD-91 (PR #127) made it always-on because Codeman can't detect a user's own loopback reverse proxy (own cloudflared/`tailscale serve`/nginx → 127.0.0.1), closing that residual plain-bypass gap. Hook curls cat the secret file at exec time via `$CODEMAN_HOOK_SECRET_FILE` (session env, `config/hook-secret.ts`); a missing/wrong secret gets 401 and rate-limits in a dedicated bucket (never locks out login). Tunnel enable **refuses** without `CODEMAN_PASSWORD` unless exposure is acknowledged — via `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK=1` (env, COD-55) **or** the per-request `acknowledgeUnauthTunnel:true` action field (1.1.9): the welcome/settings tunnel toggle pops a security confirm dialog and, on confirm, resends with that flag (server logs a loud warning on every passwordless tunnel start; curl/API stay refused without password/env/flag). The flag is an action field, never persisted |
 | **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), `CODEMAN_DOCKER_BRIDGE_HOOKS`=1 (opt-in hooks-only listener on the docker bridge gateway so in-container hooks reach a loopback-bound server; bind IP from `CODEMAN_DOCKER_BRIDGE_HOST` or auto-detect)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        |
-| **Validation**    | Zod schemas, Unicode-aware path allowlist regex, env prefix allowlist (`CLAUDE_CODE_*`/`OPENCODE_*`/`CODEX_*`/`GEMINI_*`/`GOOGLE_*`)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
+| **Validation**    | Zod schemas, Unicode-aware path allowlist regex, env prefix allowlist (`CLAUDE_CODE_*`/`OPENCODE_*`/`CODEX_*`/`GEMINI_*`/`GOOGLE_*`/`ANTIGRAVITY_*`)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
 | **Headers**       | CORS localhost-only, CSP, X-Frame-Options, HSTS if HTTPS                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                       |
 
 ## Performance and limits

docs/cron-discovery.md

@@ -44,9 +44,9 @@ records), kept distinct from the existing `ScheduledRun`.
 
 ## 2. Where agent/session types are defined
 
-- `type SessionMode = 'claude' | 'shell' | 'opencode' | 'codex' | 'gemini'`
+- `type SessionMode = 'claude' | 'shell' | 'opencode' | 'codex' | 'gemini' | 'antigravity'`
   (`src/types/session.ts:43-44`). `shell` covers the brief's "Terminal/custom".
-- CLI availability resolvers in `src/utils/{claude,codex,gemini,opencode}-cli-resolver.ts`.
+- CLI availability resolvers in `src/utils/{claude,codex,gemini,antigravity,opencode}-cli-resolver.ts`.
 - **Integration point:** the job's `agentType` reuses `SessionMode` verbatim.
 
 ## 3. Where input is sent into a session

docs/cron-guide.md

@@ -1,7 +1,7 @@
 # Cron Jobs — User & Operator Guide
 
 Codeman's **Cron** feature lets you save named, recurring jobs that automatically
-spin up a Claude (or shell / OpenCode / Codex / Gemini) session on a schedule and
+spin up a Claude (or shell / OpenCode / Codex / Antigravity / Gemini) session on a schedule and
 feed it a prompt. Think "cron for agent sessions": _"every weekday at 3am, open a
 Claude session in `~/proj` and tell it to update dependencies and open a PR."_
 
@@ -91,7 +91,7 @@ These map 1:1 to `CronJobSchema` (`src/web/schemas.ts`) and the `CronJob` type
 | Field                      | Required    | Values / limits                                          | Notes                                                                                                                                                                                      |
 | -------------------------- | ----------- | -------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
 | `name`                     | ✅          | 1–200 chars                                              | Display name; also used as the created session's name.                                                                                                                                     |
-| `agentType`                | ✅          | `claude` \| `shell` \| `opencode` \| `codex` \| `gemini` | Reuses Codeman's `SessionMode`. `shell` = a plain terminal.                                                                                                                                |
+| `agentType`                | ✅          | `claude` \| `shell` \| `opencode` \| `codex` \| `gemini` \| `antigravity` | Reuses Codeman's `SessionMode`. `shell` = a plain terminal.                                                                                                                                |
 | `workingDir`               | ✅          | valid path (allowlist-validated)                         | Validated at **create/update** (must exist, be a directory, and not resolve into a blocked tree — `/etc`, `/root`, `/proc`, `/sys`, `/dev`, or `/` itself) and again **at fire time**.     |
 | `launchCommand`            | —           | ≤ 2000 chars, single line                                | `shell` mode only: sent as the **first input line** once the shell is up, before the prompt. Ignored for other agent types.                                                                |
 | `promptMode`               | ✅          | `inline_text` \| `prompt_file_path`                      | See §5.                                                                                                                                                                                    |

docs/docker-cases.md

@@ -2,7 +2,7 @@
 
 Run a case inside an **isolated Docker container** instead of directly on the host. Any number of Codeman sessions can share one container (it is scoped to the case, not the session), so a whole project lives in a sandbox with its own network, resource caps, and filesystem, and you can **export the container to move it to another machine**.
 
-Docker mode is a **location overlay on cases**, the direct analog of [remote SSH cases](./remote-hosts.md): where a remote case runs a local tmux pane doing `ssh host` into a durable remote tmux server, a docker case runs a local tmux pane doing `docker exec -it` into a durable **in-container** tmux server. It is not a separate `SessionMode`, so `claude` / `shell` / `opencode` / `codex` / `gemini` all work inside the container.
+Docker mode is a **location overlay on cases**, the direct analog of [remote SSH cases](./remote-hosts.md): where a remote case runs a local tmux pane doing `ssh host` into a durable remote tmux server, a docker case runs a local tmux pane doing `docker exec -it` into a durable **in-container** tmux server. It is not a separate `SessionMode`, so `claude` / `shell` / `opencode` / `codex` / `gemini` / `antigravity` all work inside the container.
 
 ## One-time setup: build the base image
 

docs/extending-codeman.md

@@ -0,0 +1,228 @@
+# Extending Codeman
+
+Codeman has no plugin runtime, and that is a deliberate choice rather than a
+missing feature. A plugin runtime means running third-party code inside a process
+that spawns agents with your credentials, on a server people routinely expose
+over a tunnel or Tailscale. Codeman's security model is one of its reasons to
+exist, so it does not hand that away for an extension mechanism.
+
+Instead there are four seams that already work, from any language, with nothing
+installed:
+
+| You want to | Use | Runs where |
+| --- | --- | --- |
+| Show your own UI inside Codeman | [Web tabs](#seam-1-web-tabs) | Your own process, rendered as a tab |
+| React when an agent needs you | [SSE events](#seam-2-sse-events) | Anywhere that can hold an HTTP connection |
+| Drive Codeman from a script | [HTTP API](#seam-3-http-api-and-cli) or the `codeman` CLI | Anywhere |
+| React inside a Claude session | [Hooks](#seam-4-hooks) | The agent's own machine |
+
+Everything below is covered by the stability promise in
+[`versioning-policy.md`](versioning-policy.md): endpoint paths, the response
+envelope, `errorCode` values, and SSE event names are stable. Additive changes
+(new endpoints, new optional fields, new events) are non-breaking. Breaking
+changes ship under a new prefix (`/api/v2`).
+
+## Before you start
+
+**Base URL.** `http://127.0.0.1:3000` by default. Prefer the versioned prefix
+`/api/v1/...` for anything you publish; the unversioned `/api/...` is an alias.
+
+**Auth.** If `CODEMAN_PASSWORD` is set, send HTTP Basic on every request, or
+authenticate once and keep the `codeman_session` cookie. With no password set,
+Codeman is loopback-only and unauthenticated.
+
+```bash
+curl -u admin:$CODEMAN_PASSWORD http://127.0.0.1:3000/api/v1/sessions
+```
+
+**Envelope.** Every response is `{"success": true, "data": ...}` or
+`{"success": false, "error": "...", "errorCode": "..."}`. Check the HTTP status
+or `body.success`, then read `body.data`. The full `errorCode` to status mapping
+is in [`api-reference.md`](api-reference.md).
+
+## Seam 1: Web tabs
+
+The highest-leverage seam. Any web app you can serve locally becomes a tab beside
+your agent sessions. You write a normal web page; Codeman handles embedding it.
+
+```bash
+curl -u admin:$PASS -X POST http://127.0.0.1:3000/api/v1/webviews \
+  -H 'Content-Type: application/json' \
+  -d '{"name":"My Dashboard","url":"http://127.0.0.1:8787","icon":"📊"}'
+```
+
+Fields: `name` (1 to 60 chars), `url`, and optionally `icon` (a single glyph, max
+8 code units), `embedMode` (`proxy` by default, or `direct`), and `trusted`.
+
+Related endpoints: `GET /api/v1/webviews`, `PATCH /api/v1/webviews/:id`,
+`DELETE /api/v1/webviews/:id`, `POST /api/v1/webviews/probe` (reachability and
+framing check), `POST /api/v1/webviews/:id/open`.
+
+### Why it is proxied
+
+By default your page is served through Codeman's own origin at `/webview/:cap/*`
+rather than framed directly. A direct iframe fails three ways at once: production
+is HTTPS so `http://` targets are blocked as mixed content, many dashboards send
+`X-Frame-Options: DENY`, and Codeman's own `default-src 'self'` CSP blocks
+cross-origin frames. Proxying solves all three without weakening the CSP.
+
+### The two things that will confuse you
+
+A proxied frame is sandboxed and therefore **opaque-origin** unless you set
+`trusted: true`. Two consequences look like bugs in your own app:
+
+1. **Root-absolute URLs built at runtime** (`/assets/x.png` assembled in JS)
+   escape the injected `<base>` tag. Codeman injects a `runtimeUrlShim()` that
+   patches the common DOM sinks, but if you construct URLs in an unusual way,
+   prefer relative paths.
+2. **Same-host `fetch` and `XHR` are CORS-checked with `Origin: null`.** Codeman
+   handles this with `buildProxyCorsHeaders()`, and the proxy is exempt from the
+   global `OPTIONS` short-circuit. If you see "Failed to fetch" while the page
+   itself renders fine, this is the area to look at.
+
+⚠️ `trusted: true` opts out of the sandbox. A proxied page is served from
+Codeman's origin, so `allow-same-origin` lets it read the Codeman page and call
+the API that spawns agents. Only mark your own trusted code.
+
+## Seam 2: SSE events
+
+`GET /api/v1/events` is a Server-Sent Events stream. Each message is
+`event: <name>` plus `data: <json>`. There are 149 event names following a
+`domain:action` convention, registered in `src/web/sse-events.ts`.
+
+The ones most integrations want:
+
+| Event | Meaning |
+| --- | --- |
+| `session:created`, `session:deleted` | A session appeared or went away |
+| `session:idle` | The agent stopped working |
+| `session:completion` | A completion message was detected |
+| `session:exit`, `session:error` | The session ended or failed |
+| `hook:permission_prompt` | The agent is asking for permission |
+| `hook:idle_prompt`, `hook:stop` | The agent is waiting on you, or stopped |
+| `hook:task_completed`, `task:completed` | Work finished |
+| `subagent:discovered`, `subagent:completed` | Background agent lifecycle |
+| `mux:died` | A multiplexer session died unexpectedly |
+| `cron:runCreated`, `cron:runUpdated` | Scheduled job activity |
+
+### Filtering
+
+`?sessions=id1,id2` suppresses only the high-volume `session:terminal` stream for
+sessions you did not list. Lifecycle and metadata events are always delivered, so
+you cannot accidentally filter away the thing you are listening for.
+
+Pass `?clientId=<uuid>` to enable live filter updates through
+`POST /api/v1/events/subscribe` without reconnecting the stream.
+
+### Example: notify when any agent needs you
+
+```js
+const res = await fetch('http://127.0.0.1:3000/api/v1/events', {
+  headers: { Authorization: 'Basic ' + btoa(`admin:${process.env.CODEMAN_PASSWORD}`) },
+});
+const reader = res.body.getReader();
+const decoder = new TextDecoder();
+let buf = '';
+const WANTED = new Set(['hook:permission_prompt', 'hook:idle_prompt', 'session:idle']);
+
+for (;;) {
+  const { value, done } = await reader.read();
+  if (done) break;
+  buf += decoder.decode(value, { stream: true });
+  const frames = buf.split('\n\n');
+  buf = frames.pop() ?? '';
+  for (const frame of frames) {
+    const name = frame.match(/^event: (.+)$/m)?.[1];
+    const data = frame.match(/^data: (.+)$/m)?.[1];
+    if (name && WANTED.has(name)) notify(name, JSON.parse(data ?? '{}'));
+  }
+}
+```
+
+## Seam 3: HTTP API and CLI
+
+Around 199 handlers across 21 route files cover sessions, cases, files, cron,
+respawn, Ralph, the orchestrator, search, and admin. Each route module carries an
+`@fileoverview` describing its endpoints.
+
+The common ones:
+
+```bash
+# List sessions (live + persisted + transcript history, deduped)
+curl -u admin:$PASS http://127.0.0.1:3000/api/v1/sessions/unified
+
+# Create a session
+curl -u admin:$PASS -X POST http://127.0.0.1:3000/api/v1/sessions \
+  -H 'Content-Type: application/json' \
+  -d '{"workingDir":"/home/me/project","mode":"claude"}'
+
+# Send a prompt (single-line only, "\r" submits)
+curl -u admin:$PASS -X POST http://127.0.0.1:3000/api/v1/sessions/$ID/input \
+  -H 'Content-Type: application/json' \
+  -d '{"input":"run the tests","useScreen":true}'
+```
+
+For shell scripting, the `codeman` CLI is the same surface without the HTTP
+plumbing:
+
+```
+codeman session start|stop|list|logs     codeman task add|list|status|remove|clear
+codeman ralph start|stop|status|reset    codeman users add|passwd|list
+codeman status | list | attach <path>    codeman doctor
+```
+
+## Seam 4: Hooks
+
+Claude Code hooks post to `POST /api/v1/hook-event` from inside an agent session.
+Codeman installs its own hooks automatically, but the endpoint is open to yours.
+
+```json
+{ "event": "task_completed", "sessionId": "abc123", "data": { "any": "json" } }
+```
+
+`event` must be one of `permission_prompt`, `elicitation_dialog`, `idle_prompt`,
+`stop`, `teammate_idle`, `task_completed`. Each becomes the matching `hook:*` SSE
+event.
+
+⚠️ This endpoint skips Basic auth so hooks keep working, but when auth is active
+the loopback bypass requires the `X-Codeman-Hook-Secret` header
+(`~/.codeman/hook-secret`) unconditionally.
+
+## Gotchas
+
+Every one of these has cost somebody real time.
+
+- **CORS is localhost-only.** `Access-Control-Allow-Origin` is echoed only for
+  `localhost`, `127.0.0.1`, and `::1`. A browser app on any other origin cannot
+  call the API. Integrate server-side.
+- **A missing `Origin` header is allowed**, which is why curl, CLIs, and hooks
+  work. Cross-site origins are blocked by the CSRF guard.
+- **Reverse-proxy domains are rejected** by the anti-DNS-rebinding Host allowlist
+  unless added via `CODEMAN_ALLOWED_HOSTS=host,.suffix`.
+- **`null` is not `undefined`.** Request schemas use Zod `.optional()`, which
+  accepts `undefined` only. `JSON.stringify({ field: null })` keeps the null on
+  the wire and fails with `INVALID_INPUT`. Omit the key instead. This has caused
+  shipped bugs more than once.
+- **`text/plain` bodies stay raw.** Auto-parsing them as JSON enabled
+  simple-request CSRF, so it is deliberate. Send `application/json`.
+- **Prompts are single-line.** Input is delivered as text plus a separate Enter;
+  a multi-line string breaks the agent's input handling. Send `\r` to submit.
+- **Unwrap the envelope** before reading fields. `data` is not the response body.
+
+## Publishing your integration
+
+There is no registry and no review queue. Add the GitHub topic
+**`codeman-integration`** to your public repository so others can find it, and
+link back to Codeman in your README.
+
+If a real ecosystem of these appears, a manifest format and an install command
+become worth building. Until then, these four seams are the contract, and they
+require nothing of you but HTTP.
+
+## What Codeman deliberately does not have
+
+- **No in-process plugin runtime.** See the reasoning at the top of this page.
+- **No build or startup hooks** for third-party code. Run your own process.
+- **No per-plugin config or state directories.** Manage your own files.
+- **No sandbox**, because there is nothing to sandbox. Your integration is your
+  process, with your permissions, talking HTTP.

docs/remote-sessions.md

@@ -1,7 +1,7 @@
 # Remote Sessions (SSH)
 
 Codeman can run a session's agent on a **remote host over SSH** instead of the
-local machine. The agent (Claude, OpenCode, Codex, Gemini, or a plain shell)
+local machine. The agent (Claude, OpenCode, Codex, Antigravity, Gemini, or a plain shell)
 runs inside a `tmux` server **on the remote host**, so it survives the SSH
 connection dropping; Codeman attaches to it the same way it attaches to a local
 managed session.
@@ -30,7 +30,7 @@ Types live in `src/types/session.ts`; persistence in `src/remote-hosts.ts`.
 | `RemoteHost` (extends `RemoteSshOptions`) | A saved host: `id`, `label`, `host`, `username`, `port?`, `commands?` (per-mode launch command override). |
 | `RemoteCase` | A working directory on a host: `name`, `type: 'remote'`, `hostId`, `remotePath`. |
 | `SessionRemote` (extends `RemoteSshOptions`) | The resolved bundle stamped onto a live session: host coordinates + `remotePath` + `commands`, plus **`owned?`** and **`remoteSessionName?`** (COD-105 — see [Ownership](#ownership-launched-vs-discovered-and-attached-cod-105)). Built by `toSessionRemote(host, case)` (sets `owned: true`) for the launch path, or `toAttachedSessionRemote(host, name, path)` (sets `owned: false`) for the attach path. Both copy the advanced SSH options through so every connection is identical. |
-| `RemoteCommandMode` | `Extract<SessionMode, 'shell' \| 'claude' \| 'opencode' \| 'codex' \| 'gemini'>` — the modes that can run remotely. |
+| `RemoteCommandMode` | `Extract<SessionMode, 'shell' \| 'claude' \| 'opencode' \| 'codex' \| 'gemini' \| 'antigravity'>` — the modes that can run remotely. |
 | `RemoteSessionInfo` (COD-105) | One discovered remote tmux session: `name` (always `codeman-*`), `attached` (a client is connected), `created` (epoch s), `windows`. Returned by `listRemoteCodemanSessions()`. |
 
 Persistence is two flat JSON arrays in the instance data dir:
@@ -115,7 +115,7 @@ Key points:
 - **`exec <cli>`** replaces the pane shell with the agent, so the pane PID *is*
   the agent. The per-mode command comes from `remote.commands?.[mode]` or
   `defaultRemoteCommandForMode(mode)` (`exec claude` / `exec opencode` /
-  `exec codex` / `exec gemini` / `exec bash -l`).
+  `exec codex` / `exec gemini` / `exec agy` / `exec bash -l`).
 - The **whole tmux invocation is a single shell-quoted ssh argument**, and the
   pane command is independently quoted, so a `remotePath` with spaces is safe.
 - Connection options come from the **same `buildSshConnectionArgs(remote)`** as

docs/security-architecture.md

@@ -489,7 +489,7 @@ production layout (`~/.codeman`, `-L codeman`, port 3000).
 Docker cases (1.4.0) run a session inside a per‑case container instead of on the host. The security posture:
 
 - **Hardened create flags, always** — `--cap-drop ALL`, `--security-opt no-new-privileges`, `--pids-limit` (fork‑bomb guard), `--memory` == `--memory-swap` (a real OOM cap), `--init`, and non‑root: `--user <hostUid>:0` on Linux (host uid → workspace files stay host‑owned; GID 0 keeps `$HOME` writable), `--userns=keep-id` on rootless Podman. **Never** `--privileged`, and **never** the docker socket — the pure builder in `docker-hosts.ts` cannot emit them and the schema cannot represent them.
-- **Credentials never enter an image** — the convenient default bind‑mounts host cred dirs (`~/.claude`, `~/.codex`, `~/.gemini`, `~/.config/{gcloud,opencode}`) read‑write. Bind mounts are physically excluded from `docker commit`, so exported images are secret‑free. API‑key CLIs get their key as an exec‑time NAME‑ONLY `--env OPENAI_API_KEY` (no `=value`, no `ps` leak, never committed); a create‑time `-e` for a secret is never used. The **sealed** profile (`mountCredentials:false` + `network:none`) drops the host mounts; full‑image export is then refused (an in‑container login would ride the committed layer) unless a pre‑commit scrub is opted into.
+- **Credentials never enter an image** — the convenient default bind‑mounts host cred dirs (`~/.claude`, `~/.codex`, `~/.gemini` — which also carries Antigravity's `antigravity-cli/` state — and `~/.config/{gcloud,opencode}`) read‑write. Bind mounts are physically excluded from `docker commit`, so exported images are secret‑free. API‑key CLIs get their key as an exec‑time NAME‑ONLY `--env OPENAI_API_KEY` (no `=value`, no `ps` leak, never committed); a create‑time `-e` for a secret is never used. The **sealed** profile (`mountCredentials:false` + `network:none`) drops the host mounts; full‑image export is then refused (an in‑container login would ride the committed layer) unless a pre‑commit scrub is opted into.
 - **Blast radius — accept it explicitly** — the convenient profile mounts an arbitrary host workspace RW plus the host credential dirs RW into a network‑enabled container, so container‑run agent code can read/modify those host trees and reach the network at once. Still a net improvement over today's on‑host `--dangerously-skip-permissions` execution; use the sealed profile for genuinely untrusted work.
 - **Import is untrusted‑bundle‑safe** — `/api/docker-cases/import` validates the manifest + per‑member SHA‑256 before extraction, rejects absolute / `..` tar members (traversal guard), and re‑tags the loaded image into a quarantined namespace so it can never overwrite `codeman/agent:base` or a pre‑existing tag.
 - **Host guard & the bridge‑hooks listener** — in‑container hook callbacks carry `Host: host.docker.internal` / `host.containers.internal`; both are on the always‑on host‑header allowlist (`DOCKER_HOST_GATEWAY_ALIASES`) and resolve to the host only from inside a container netns, so they are not a browser DNS‑rebinding surface. On a loopback‑only server, in‑container hooks are opt‑in via `CODEMAN_DOCKER_BRIDGE_HOOKS=1`, which binds a SECOND listener on the docker bridge gateway serving **only** the hook endpoints (every other path → `403`) into the same hook‑secret‑gated pipeline. The bridge is host‑internal (containers + host), not the LAN, so it does not widen network exposure; the hook secret is bind‑mounted read‑only and referenced by path.

docs/vm-cases-plan.md

@@ -0,0 +1,201 @@
+<!-- Design doc drafted 2026-07-28 from WWDC26 session 224 research. STATUS: PLANNED, NOT IMPLEMENTED. Blocked on macOS 27 "Golden Gate" (beta now, GA expected fall 2026). -->
+
+# VM Cases (macOS Virtualization framework), Implementation Plan
+
+## Status
+
+PLANNED, nothing implemented. This is the design + phased execution plan for a native-macOS VM isolation tier for cases ("the VM subsystem"), modeled on Docker cases (`docs/docker-cases-plan.md`). Testbed prerequisite: a macOS 27 host (see Section 8).
+
+**⚠ DESIGN DIRECTION (owner, 2026-07-29): the subsystem is GUI-first.** Users want real macOS desktops, not headless SSH machines. Guests may be macOS (GUI-only in practice) or Linux (GUI or headless). Key decision 3 below carries the full consequences; anything in this doc that reads as "Linux-first / headless-first" predates this and has been revised.
+
+**2026-07-29: Phase 0 substantially validated on the beta testbed; full Apple-stack reference now lives in [`docs/vm-subsystem-apple-stack.md`](vm-subsystem-apple-stack.md)** (API surfaces, beta bugs, our empirical results, and design implications). Plan-relevant corrections from that work: vmnet's topology/port-forwarding APIs are macOS 26 (only the loopback fix is 27); guest provisioning is macOS-guests-only (Linux stays cloud-init, proven working); DiskImageKit has NO flatten/merge, so the `export` subcommand ships the layer chain (or flattens in-guest) instead of flattening; seed ISOs are base-build-time only, never attached at case runtime; per-case EFI variable stores are mandatory; guest health checks read DHCP leases, never serial/ping.
+
+## 1. Context and motivation
+
+WWDC 2026 session 224 ("Expand the Capabilities of your Virtualization App", https://developer.apple.com/videos/play/wwdc2026/224/) shipped the missing pieces for programmatic, fleet-style VM management on macOS:
+
+- **`VZMacGuestProvisioningOptions`**: automated first-boot setup of a macOS guest (user account, auto-login, SSH enabled) with zero interactive setup.
+- **DiskImageKit**: stacked disk images on the Apple Sparse Image Format (ASIF): a read-only base layer plus cheap per-VM cache/overlay layers. Direct analog of Docker image layers + writable container layer.
+- **vmnet framework**: custom network topologies and port forwarding from the host process.
+- **`VZCustomVirtioDevice`**: custom low-latency host<->guest channels (Linux guests).
+- **AccessoryAccess**: USB passthrough (not relevant to Codeman, out of scope).
+
+Codeman's isolation story today is Docker cases. On macOS, Docker means Docker Desktop / a Linux VM anyway, with weaker fidelity and a heavyweight dependency. The Virtualization framework gives hardware-virtualized per-case sandboxes natively, with a layered-image story that mirrors what `scripts/build-agent-image.mjs` does for Docker. This is the premium native-macOS tier ON TOP of Docker cases, never a replacement (Docker remains the cross-platform story; the Linux prod box cannot use any of this).
+
+## 2. Platform reality (hard constraints)
+
+| Constraint                | Detail                                                                                                                                              |
+| ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Host OS                   | macOS 27 "Golden Gate" required for the new APIs (dev beta since 2026-06-08, public beta since 2026-07-13, GA expected fall 2026)                    |
+| Host hardware             | Apple Silicon only (macOS 27 dropped Intel). Testbed: the owner's dedicated MacBook (Section 8); the M4 Mac mini (macOS 26.4, runs the second Codeman install) stays on stable + untouched |
+| Guest provisioning        | `VZMacGuestProvisioningOptions` needs macOS 27 on BOTH host and guest. Linux guests provision via cloud-init instead                                 |
+| macOS guest concurrency   | **Hard kernel cap: 2 concurrent macOS VMs per host. MEASURED on 27 beta 4 (2026-07-29), not inferred**: the 3rd VM is refused instantly with `VZErrorDomain` code 6 while 39% of RAM is free, so more hardware does NOT raise it. Since macOS GUI guests are the headline use case, this is a real product capacity limit to schedule around and surface in the UI. Linux guests are uncapped (resource-bound only) |
+| Language                  | Virtualization framework is Swift/ObjC only; Node cannot call it. Requires a Swift helper binary (Key decision 2)                                    |
+| Entitlement               | Host process needs `com.apple.security.virtualization`. Fine for a locally built dev binary; distribution needs signing thought (Section 9)          |
+| Nested virtualization     | Linux-guest-only on M3+. A macOS 27 VM cannot dependably host its own guests, so the host-side APIs must be tested on bare-metal 27 (dual-boot)      |
+| CI                        | Cannot run in CI (needs beta macOS on Apple Silicon). Same answer as tmux/docker: no-op all VM IO under `VITEST`, unit-test the pure parts            |
+
+## 3. Goal and user stories
+
+Add "VM cases" to Codeman: a case can point at a per-case virtual machine on a macOS host, and any CLI backend runs inside it over the existing remote-SSH session machinery. A LOCATION OVERLAY on cases, exactly like remote-SSH and Docker cases, NEVER a sixth `SessionMode`.
+
+- As a Mac user, I link a case to a VM so an autonomous run executes behind a hardware virtualization boundary (stronger than Docker's shared kernel) while file viewing, transcripts, and hooks keep working.
+- Per-case VMs are instant and cheap: a shared provisioned base image plus a per-case overlay, not a full image copy per case.
+- Killing a session kills only its in-guest tmux; the VM stays up while sibling sessions remain; case delete tears the VM down.
+- I export a case's VM overlay as a portable artifact (mirror of `docker-exports/`), secrets excluded.
+- On a non-mac host, or a Mac without the helper, the feature is invisible: zero UI, zero probes, zero errors.
+
+Non-goals for the MVP: USB passthrough, custom Virtio channels (Phase 3 candidate), macOS-guest fleets (capped at 2 anyway), Kubernetes-style orchestration, Intel Macs.
+
+## 4. Architecture
+
+```
+Codeman (Node, unchanged session layer)
+  |  JSON over stdout (same pattern as shelling out to docker/tmux)
+  v
+codeman-vm  (Swift package: CLI + per-VM GUI runner app in the console session)
+  |  Virtualization / DiskImageKit / vmnet
+  v
+per-case VM (macOS or Linux)
+      |-- GUI mode: VZVirtualMachineView in a window  --> guest screen sharing --> browser (noVNC)
+      |-- shell:    SSH on vmnet IP                   --> existing remote-SSH tmux machinery
+      ^ VirtioFS: host case dir mounted at the SAME absolute path
+```
+
+Note the runner is a **GUI app in the console user's session**, not a detached daemon: a daemon-launched VM cannot render, which is fatal for macOS guests and for Linux desktop cases.
+
+### Key decision 1: location overlay, not a mode
+
+Identical reasoning to Docker/remote-SSH (see CLAUDE.md): the session layer, respawn, Ralph, recovery, and quick-start plumbing all stay untouched. `SessionMode` stays five-valued. State mirrors the Docker pair: `~/.codeman/vm-hosts.json` + `vm-cases.json`, new `src/vm-hosts.ts` with the storage + pure helpers split.
+
+### Key decision 2: Swift helper CLI (`codeman-vm`)
+
+The framework is Swift-only, so all VM work lives in a SwiftPM package (`packages/codeman-vm/`), a CLI with a stable JSON contract:
+
+- `create-base --guest linux|macos`: build the shared base image. Linux: boot an arm64 cloud image with EFI + cloud-init, install Node 22 + tmux + the four CLIs (same inventory as `docker/agent.Dockerfile`), seal as base ASIF. macOS: IPSW restore + `VZMacGuestProvisioningOptions` (agent user, SSH on), then **desktop-readiness baking**, which is mandatory for GUI guests: suppress the per-user first-login assistant (`com.apple.SetupAssistant` keys + the User Template), enable auto-login (`autoLoginUser` + `/etc/kcpassword`), disable screensaver/lock/display-sleep, and set a static wallpaper (animated "aerials" wallpaper is unusable over remote display). ⚠ Use RAW (not ASIF) for macOS guest disks until the beta's macOS-guest space-reclamation bug is fixed.
+- `create <case>`: DiskImageKit stacked image: shared read-only base + fresh per-case overlay. Near-instant, space-efficient.
+- `start <case>` / `stop` / `status` / `ip`: lifecycle + vmnet NAT; `ip` reports the guest SSH endpoint.
+- `export <case>` / `import`: flatten overlay + workspace tar + manifest, credentials excluded (mirror of docker-export).
+
+A VM dies with its owning process, so `start` spawns a DETACHED per-VM runner process (analog of the detached `scripts/self-update.sh` trick) rather than a monolithic daemon; `status` talks to it over a unix socket in the instance data dir (`dataPath()`, never a hardcoded `~/.codeman` path).
+
+### Key decision 3: multi-guest, and GUI is a first-class mode (REVISED 2026-07-29 by the repo owner)
+
+The subsystem supports both macOS and Linux guests, and a guest runs in one of two **display modes**:
+
+| | macOS guest | Linux guest |
+| --- | --- | --- |
+| **GUI mode** | **the point of the feature**; a real macOS desktop. Mandatory: nothing renders without an attached `VZVirtualMachineView` in an unlocked host session | supported (EFI + virtio-gpu framebuffer) for desktop Linux cases |
+| **Headless mode** | not offered: a macOS guest with no view renders nothing, so a "headless macOS desktop" is a contradiction. SSH-only macOS is possible but is not what this feature is for | supported and cheap; the natural mode for agent/CI work, driven over SSH |
+
+Consequences that flow from GUI being first-class:
+- VM processes are **GUI apps in the console user's session** (LaunchAgent / `launchctl asuser`), never daemons. A daemon-launched VM cannot render.
+- **The host is part of the product surface**: it must auto-login, never lock, never sleep, and keep a live WindowServer. Host lock == every VM's screen goes black, so the screen lock is effectively a global kill switch for every VM display on the machine. The product must own these host settings rather than treat them as user preference.
+- **FileVault conflicts with unattended GUI hosting** and the trade-off must be a deliberate choice: FileVault disables auto-login, so a full-disk-encrypted host needs a human at a keyboard (or a remote screen-sharing session) after every reboot before any VM can render. Options are (a) FileVault on, accept manual login per boot, (b) FileVault off on a dedicated VM host so it boots straight into a rendering session, or (c) FileVault on plus a remote-unlock runbook. Codeman should detect the state and tell the user which one they are in instead of silently serving black screens.
+- **Guests must be desktop-ready, not just booted**: auto-login, no screensaver/lock, and the per-user first-login assistant pre-suppressed at base-image time (`com.apple.SetupAssistant` keys, plus the User Template so later accounts inherit it). Otherwise the user connects to a login prompt or a setup wizard, which is exactly what happened during the first hands-on run.
+- **Capacity is capped for macOS**: at most 2 concurrent macOS VMs per host, confirmed by our own test on 27 beta 4 (3rd refused with `VZErrorDomain` 6 at 39% free RAM; it is a kernel quota, so bigger hardware does not help). Scheduling must queue or evict beyond 2, the UI must explain why, and the scheduler should tolerate the acknowledged slot-leak bug (a slot occupied with nothing running, host-reboot to clear). Linux guests are uncapped and bounded only by host resources, which is the lever for scaling case counts on one machine.
+- **Access is via the guest's own screen**, viewable in a browser through the noVNC chain (see `docs/vm-subsystem-apple-stack.md` §8), so no client-version or client-install requirements land on the user.
+
+Provisioning per guest type: `VZMacGuestProvisioningOptions` for macOS (needs 27-on-27, first-boot-only, and does NOT skip the per-user wizard), cloud-init NoCloud seed ISO for Linux (proven working).
+
+### Key decision 3b: the GUI VM host profile, and supervision that catches black screens
+
+GUI hosting only works if the host is configured for it and supervised. This profile was derived the hard way on the testbed (prototyped there 2026-07-30) and should be what `codeman-vm` installs and verifies:
+
+**Host profile** (the product should own these, not leave them to preference):
+1. **No login barrier.** Either FileVault off + auto-login (a dedicated VM host boots straight into a rendering session, fully unattended), or FileVault on and remote reboots done with `sudo fdesetup authrestart`, where the pre-boot unlock *is* the login so the machine returns already logged in with encryption intact. **`authrestart` is VERIFIED on the testbed (2026-07-30): the host rebooted remotely and came back with a live logged-in console session, FileVault still enabled, no password prompt** — this is the recommended pattern for an encrypted GUI VM host. Plain reboots on a FileVault host always need a human, so Codeman should detect that combination and warn instead of serving black screens.
+2. **Never lock**: lock policy off (needs the account password, so it is a setup step, not a scriptable one) plus `caffeinate -d -i -m -u` re-armed per session.
+3. **Never sleep**: `pmset -a sleep 0 displaysleep 0 disablesleep 1`; a physical display is NOT required (a lid-closed laptop renders fine, only an unlocked session matters). Note OS updates reset these.
+4. **Session-independent control plane**: run VPN/remote access as a system service, never a session app, and keep the access chain (forwards, VNC proxies, web endpoints) in LaunchDaemons so a session restart cannot sever operator access.
+
+**Supervision** must be a **root LaunchDaemon**, not a user LaunchAgent. This is the load-bearing detail: a user agent cannot launch a GUI app into the Aqua session, so its restart attempts fail *silently* (the child dies instantly, leaving an empty log while the supervisor cheerfully reports success). A root daemon can, via `launchctl asuser <uid> sudo -u <user> …`, and those launches persist. Prototyped and verified on the testbed 2026-07-30; a working supervisor runs on a short interval and:
+
+- Restarts the runner when the process is gone **or when its log shows `WindowServer event port death`**, which means it is permanently blind while still looking alive.
+- Defers restarts while the console is at the login window, and launches into whichever session actually exists (resolve the console user with `stat -f %Su /dev/console`, never a hardcoded one).
+- Re-points the guest port-forward whenever the guest's NAT lease changes, which happens on **every guest boot** under plain NAT. A vmnet DHCP reservation for a stable per-case IP is the better long-term answer.
+- **Re-applies host power settings**, because `pmset -a disablesleep 1` does NOT survive a reboot (caught on the supervisor's first run after a real reboot) and OS updates reset it too.
+- Re-arms the keep-awake helper, which dies with its session.
+- Ideally also samples the guest framebuffer for non-black content, since a black screen is the one symptom common to every failure mode here.
+
+`pgrep` alone is worthless for health: every failure mode in this session presented as a healthy process.
+
+### Key decision 4: sessions ride the existing remote-SSH machinery
+
+A provisioned guest is literally an SSH host on a vmnet IP. Session launch = the remote-SSH flow with the host swapped in: durable remote `tmux -L codeman-remote`, session names failing `SAFE_MUX_NAME_PATTERN` on purpose, EVERY ssh command line through `buildSshConnectionArgs()` (command-injection invariant), run flows through `POST /api/quick-start` (never `POST /api/sessions`, which stat-validates `workingDir` locally). What is genuinely new is only lifecycle (create/start/stop/export) and the vm-hosts/vm-cases overlay state.
+
+### Key decision 5: workspace via VirtioFS at the same absolute path
+
+Mirror the Docker bind-mount invariant: the case workspace is a real host directory shared into the guest via VirtioFS and mounted at the SAME absolute path. That keeps file-routes/watchers on real host bytes and makes the in-guest transcript projHash match the host. Without this, transcripts/attachments/file viewer all silently degrade.
+
+### Key decision 6: credentials seeded, hooks bridged
+
+- Credentials are SEEDED (read-only share, copied into the guest once at create), never shared read-write, and excluded from exports: byte-for-byte the Docker cases rule and rationale.
+- Hooks: on the loopback-only prod bind a guest cannot reach `127.0.0.1:3000`. Mirror `CODEMAN_DOCKER_BRIDGE_HOOKS` with a `CODEMAN_VM_BRIDGE_HOOKS` opt-in listener on the vmnet gateway IP; otherwise idle detection falls back to output-based, same as Docker.
+
+### Key decision 7: drift and teardown copy Docker semantics verbatim
+
+Config hash label on the VM (guest type, cpu/mem, share list); a drifted launch is REFUSED, never silently launched stale. One VM per case shared by all sessions; session kill = in-guest tmux kill only; case delete = stop + remove overlay; instance-scoped boot reaper for orphaned runner processes.
+
+## 5. Implementation phases
+
+**Phase 0, testbed (no repo code):** dedicated MacBook on the macOS 27 beta, remotely accessible over the tailnet (setup protocol in Section 8), Xcode 27 beta, then a throwaway Swift script proving the loop: create base -> overlay -> boot -> ssh in. This validates 80% of the design before any Codeman code.
+
+**Phase 1, `codeman-vm` helper:** SwiftPM package, the six subcommands above, JSON contract doc, detached runner + unix-socket status, Linux base image build. Deliverable is testable entirely without Codeman.
+
+**Phase 2, Codeman integration:** types (`VmHost`/`VmCase`/`SessionVm`), `src/vm-hosts.ts` (+ pure helpers: config hash, arg building, endpoint parsing), Zod schemas, `case-routes` link/unlink + listing, `quick-start` vm branch reusing the remote-SSH launch path, `Session` threading + recovery round-trip, `VITEST` no-op layer, unit tests. Feature-detect: darwin + arm64 + helper binary present, else invisible.
+
+**Phase 3, polish:** export/import UI, frontend Create Case "VM" tab + case-picker labels, SSE `vm:*` events, macOS-guest opt-in with cap surfaced, custom-Virtio input channel exploration, CLAUDE.md Key Pattern + `docs/vm-cases.md` + COM.
+
+## 6. Testing
+
+- Pure helpers unit-tested (ports pattern from `docker-hosts.ts`: 26 tests there, aim similar).
+- All helper-invoking IO no-ops under `VITEST` (the `IS_TEST_MODE` pattern in `tmux-manager.ts`).
+- End-to-end verification happens ON the beta MacBook, per the always-end-to-end rule: real base build, real per-case overlay boot, real quick-start into the guest, workspace round-trip through VirtioFS, session-delete keeps VM up, case-delete removes it.
+- CI never runs the real path; the static guards are type-level + unit-level only.
+
+## 7. Risks
+
+1. **Beta API churn**: everything here targets beta SDKs; symbol/behavior changes are likely before fall GA. Mitigation: Phase 0/1 are throwaway-tolerant; no Codeman-side commitment until the helper contract survives a beta cycle.
+2. **New artifact class**: Codeman ships pure TypeScript today; a Swift binary changes build/distribution (build-on-install via `xcrun swift build` on macs with Xcode CLT? prebuilt signed binary per release?). Needs an owner decision; local dev build is fine for the whole beta period.
+3. **Entitlement/signing**: `com.apple.security.virtualization` is trivial for local dev, real for distribution.
+4. **Adoption gating**: users need macOS 27 + Apple Silicon for months after GA. Docker cases remain the default recommendation; VM cases ship dark (feature-detected) with zero cost to everyone else.
+
+## 8. Beta testbed plan: dedicated MacBook (actionable now)
+
+Testbed is a dedicated MacBook the owner sacrifices to the beta (after a full backup). This supersedes the earlier dual-boot-the-Mini idea (git history has it): a dedicated machine means no OS-switching, no downtime for the Mini's live Codeman, and no FileVault pre-boot headaches.
+
+**Sequencing rule that makes it headless: configure ALL remote access on the CURRENT macOS first, THEN upgrade in place.** An in-place beta upgrade preserves Remote Login, Tailscale, user accounts, and auto-login, so there is no Setup Assistant and no post-install physical step. (A fresh install would boot into GUI-only Setup Assistant with no SSH, which on a headless box is a dead end.)
+
+Confirmed hardware (2026-07-28): MacBook, M3, 16 GB RAM, 256 GB disk with ~100 GB free. Verdict: green. M3 = eligible + nested-virt capable; 16 GB = host + 2-3 concurrent Linux guests (macOS guest = one at a time); 100 GB = fits with discipline: install Xcode 27 beta with the macOS platform only (skipping iOS/watchOS/tvOS simulators saves 15-20 GB), and defer any macOS guest base (~30 GB) to an external SSD or until actually needed. Linux guests + sparse ASIF overlays are the comfortable path.
+
+### Pre-upgrade checklist (owner, physical, once)
+
+1. Full backup (Time Machine or clone); the machine should be considered beta-only afterwards.
+2. Tailscale: install, sign into the tailnet, confirm it appears in `tailscale status` from another node.
+3. System Settings -> General -> Sharing: **Remote Login ON** (SSH) and **Screen Sharing ON** (for the rare GUI-only moments: Xcode license, Apple Account dialogs).
+4. **FileVault stays ON** (owner decision 2026-07-28, security over convenience). Consequences: auto-login is unavailable, but FileVault's pre-boot unlock doubles as login, so an unlocked boot still lands in a live GUI session; planned remote reboots go through `sudo fdesetup authrestart` (unlocks for exactly one restart); an UNPLANNED reboot (beta kernel panic, battery drain) parks the machine at the pre-boot screen, no SSH/Tailscale, until the password is typed physically. If the testbed goes silent, suspect this first. Keep it on AC so the battery absorbs power blips.
+5. Beta enrollment (manual): sign into the Apple Account in System Settings; System Settings -> General -> Software Update -> **Beta Updates** -> select the **macOS 27 Developer Beta** (preferred: framework fixes land weeks earlier than public beta; free since 2023 after accepting the agreement once at developer.apple.com; public-beta alternative: enroll at beta.apple.com). Then run the offered upgrade: plugged in, lid open, trusted network.
+6. Send over: tailnet name/IP, username, and a first-login password (key install + lockdown happens remotely right after).
+
+### Post-upgrade setup (remote, over the tailnet)
+
+1. Verify: `sw_vers` reports 27.x, SSH reachable.
+2. Server-ize the laptop: `sudo pmset -a sleep 0 disksleep 0 disablesleep 1` (lid-closed operation without an external display), `womp 1` (wake on network), `sudo systemsetup -setrestartpowerfailure on`. Keep on AC power.
+3. Install the controlling host's SSH key, then disable password auth.
+4. Xcode 27 beta install (the one step needing the owner's Apple Account sign-in once, doable via Screen Sharing from anywhere); `xcode-select`, license accept, verify `swift --version` + the 27 SDK (`xcrun --show-sdk-version`).
+5. Phase 0 prototype loop, all remote from here: Linux guest base image (no 27-on-27 provisioning dependency), DiskImageKit overlay, boot, vmnet NAT, ssh into the guest, run `claude --version` inside.
+6. Only after that loop works: start Phase 1 in `packages/codeman-vm/`.
+
+## 9. Open decisions (owner)
+
+1. Linux base distro/image for the default guest (proposal: Ubuntu 24.04 arm64 cloud image, matching the docker agent image's userland).
+2. Helper distribution for GA: build-on-install vs prebuilt signed binary vs "bring your own Xcode".
+3. Ship dark behind `CODEMAN_VM_CASES=1` for the first release, or feature-detect only?
+4. Export format parity with docker-exports (one manifest schema for both?).
+
+## References
+
+- Session 224: https://developer.apple.com/videos/play/wwdc2026/224/
+- Fleet-angle writeup: https://bitrise.io/blog/post/wwdc26-the-virtualization-framework-updates-that-matter-for-large-mac-fleets
+- Beta timeline: https://www.macworld.com/article/3189014/apple-july-2026-ios-ipados-macos-27-public-betas-tv-arcade-releases.html
+- Internal analogs: `docs/docker-cases-plan.md` (architecture template), `docs/remote-sessions.md` (session transport), `docs/architecture-invariants.md#docker-cases`

docs/vm-subsystem-apple-stack.md

@@ -0,0 +1,281 @@
+<!-- Reference doc for the VM subsystem (Codeman VM cases). Compiled 2026-07-29 from: Apple DocC JSON backend, macOS 27 beta 4 SDK on the testbed, a multi-source web research sweep, and hands-on prototyping on a MacBook Air M3 running macOS 27.0 beta (26A5388g). Companion to vm-cases-plan.md (the Codeman integration plan). -->
+
+# The VM Subsystem: Apple Virtualization Stack Reference (macOS 27 "Golden Gate")
+
+"VM subsystem" is the working name for Codeman's native-macOS VM isolation tier and everything under it. This document is the single place for what the Apple stack actually provides, what we have verified ourselves on the beta, and what is known-broken. The Codeman-side design lives in `docs/vm-cases-plan.md`.
+
+**Research method note:** Apple's HTML doc pages are JS-rendered and come back empty to fetchers. The working route is the DocC JSON backend: `https://developer.apple.com/tutorials/data/documentation/<path>.json` (page content) and `https://developer.apple.com/tutorials/data/index/<framework>` (full symbol tree with per-symbol `beta` flags). Everything below marked "Apple docs" was parsed from that backend directly.
+
+## 1. Component map and minimum OS versions
+
+| Component | What it is | Min host OS | Notes |
+| --- | --- | --- | --- |
+| Virtualization.framework core | VMs, EFI/Linux boot, virtio devices, VirtioFS | macOS 11-13 era | Unchanged basics; our prototype uses nothing newer than macOS 13 APIs except the DiskImageKit bridge |
+| **DiskImageKit** | ASIF + raw disk images, layered stacks | **macOS 27** | Swift-only, no ObjC headers. Section 2 |
+| **Guest provisioning** | First-boot account/SSH setup for macOS guests | **macOS 27 host AND guest** | Mac guests only as of beta 4. Section 3 |
+| vmnet topology/port-forward/DHCP APIs | Custom networks, port forwarding | **macOS 26** (NOT 27) | 27 adds exactly one fix: loopback port forwarding. Section 4 |
+| `VZVmnetNetworkDeviceAttachment` | In-process vmnet attach | macOS 26 | |
+| **`VZCustomVirtioDevice`** family | Custom paravirt devices | **macOS 27** | Linux guests only, custom guest driver required. Section 5 |
+| AccessoryAccess (USB passthrough) | USB claim + attach to VMs | macOS 27 | Requires paid-team provisioning profile, Dock app. Out of scope for Codeman. Section 6 |
+
+Corrections to the WWDC-session framing we started with: vmnet's topology family is a macOS 26 story (129 symbols, zero beta-flagged in 27); provisioning does NOT currently extend beyond macOS guests despite the generic-looking `VZGuestProvisioningOptions` base class; DiskImageKit has no attach/mount API at all (it is a file-format library that hands `DiskImage` objects to Virtualization, no `/dev/diskN`, no root needed, no entitlement documented).
+
+## 2. DiskImageKit (macOS 27, Swift-only)
+
+Public framework, `/System/Library/Frameworks/DiskImageKit.framework`. No ObjC headers; the API surface lives in the `.swiftinterface`. Verified present in the CLT 27 beta 4 SDK, and our prototype compiled against it with plain `swiftc` on the first attempt.
+
+### API surface (complete as of beta 4)
+
+```swift
+class DiskImage {
+  convenience init(creating: some DiskImage.CreationConfiguration) throws
+  convenience init(opening: some OpenConfigurationProtocol) throws
+  func appending(any DiskImage.CreationConfiguration & DiskImage.StackableLayer) throws -> any StackedImage
+  func appending(consuming DiskImage) throws -> any StackedImage   // reattach an existing layer; validates parentUUID
+  func truncate(blockCount: Int) throws                            // stacked: affects top layer; does NOT resize guest fs
+  var blockCount, blockSize, format, layerType, layerUUID, parentUUID, openMode, size, url
+}
+protocol StackedImage: DiskImage { var layers: [DiskImage] }
+struct OpenConfiguration { init(url:mode:); Mode = automatic | readOnly | readWrite }
+// CreationConfiguration statics: .asif(url:blockCount:blockSize:), .asifLayer(url:type:), .raw(url:blockCount:)
+// DiskImage.LayerType: .cache | .overlay | .overlay(blockCount:)
+// DiskImage.BlockSize: .bytes512 | .bytes4096
+// Errors: CorruptedImageError, IncompatibleStackingError(reason), InvalidBlockCountError, UnsupportedFormatError
+```
+
+Bridge into Virtualization is a new beta convenience init on the existing attachment class. Note there is no `readOnly:` parameter; read-only-ness comes from each layer's own `openMode`:
+
+```swift
+VZDiskImageStorageDeviceAttachment(diskImage: stack, cachingMode: .automatic, synchronizationMode: .full)
+```
+
+### Stacking rules (Apple docs, verbatim where quoted)
+
+- ASIF works standalone or stacked. "You can only use RAW images as standalone images or as **base** images in stacked configurations." Upper layers are always ASIF.
+- **One cache layer per stack**, any number of overlays conceptually, "shallow stacks perform better" (WWDC 224). No published max-depth guidance.
+- "Layers are processed from bottom (base) to top. The **topmost layer determines the stack's size and receives all writes**." `.overlay(blockCount:)` therefore also grows the virtual disk.
+- UUID chaining: appending sets the child's `parentUUID` to the parent's `layerUUID`. Raw bases have no UUID. "The layer UUID **changes if the layer is written to**", and reattaching a mismatched layer throws `IncompatibleStackingError`. This is the mechanism that makes a shared read-only base safe.
+- Base sharing across multiple VMs is the stated design intent ("can be shared across multiple VMs"), with the WWDC caveat that per-VM auxiliary files (EFI variable store, macOS auxiliary storage) must be duplicated per VM, never shared.
+- **There is no flatten/merge.** An overlay cannot be merged back into its base (confirmed by Howard Oakley's coverage plus an independent hands-on report). Export/move flows must ship the layer chain, or flatten inside a guest (dd to a fresh attached image).
+
+### Known issues and adoption
+
+- **ASIF space reclamation is broken for macOS guests on the beta** (deleted files never return space, survives reboots). Linux guests reclaim correctly on both raw and ASIF via `fstrim -av`. Single detailed field report, unrefuted. Since the VM subsystem targets macOS guests, the practical rule until this is fixed is: back macOS guest disks with RAW, and revisit ASIF stacking for macOS guests each beta (stacking still works, the disks just never shrink).
+- **Zero shipping adopters anywhere.** tart has a design issue with no activity; nobody has published working DiskImageKit code. Everything must be treated as field-untested (and our own testing bears that out, Section 8).
+- Framework binary grew every beta (588 → 598 across betas 1-4); expect churn until GA.
+- Release notes list no DiskImageKit known issues in any beta, which given the above says more about the notes than the framework.
+
+## 3. Guest provisioning (macOS guests only)
+
+```swift
+class VZGuestProvisioningOptions: NSObject { func validate() throws }   // "use one of its subclasses"
+class VZMacGuestProvisioningOptions: VZGuestProvisioningOptions {
+  var fullName, username, password: String
+  var logsInAutomatically: Bool
+  var enablesRemoteLogin: Bool   // SSH
+}
+// Wiring: VZMacOSVirtualMachineStartOptions.guestProvisioningOptions (Mac-typed)
+//         .setGuestProvisioning(_:) throws  (validating setter)
+```
+
+- **Requires macOS 27 on host AND guest.** Older guests **silently ignore** the options (no error).
+- **First boot after restore only.** Cannot reconfigure an already-provisioned VM; property changes after start are no-ops.
+- The base class is forward-looking scaffolding; its only subclass is Mac. A Linux/cloud-init analogue may come later; do not assume it lands in 27.0. For Linux guests, cloud-init NoCloud seed ISOs remain the provisioning path (proven working, Section 8).
+- Field-verified behavior (third-party hands-on, beta 3): provisioned account gets full admin + sudo; Setup Assistant fully skipped; SSH reachable ~48 s after first boot. **Race**: the account is created late in first boot (~T+54 s), after LaunchDaemons start (~T+33 s), so anything at daemon-level must wait for the account to exist.
+- Open Apple-acknowledged bug: provisioned users are invisible to `CSIdentityQueryExecute()` (FB23716201).
+- IPSW acquisition gotcha for automation: `VZMacOSRestoreImage.latestSupported` tracks the latest *release* (returned 26.5.2), not the installed beta; beta IPSWs must be fetched from the seed CDN explicitly.
+
+## 4. vmnet: a macOS 26 feature set, one macOS 27 fix
+
+Everything interesting shipped in macOS 26: `vmnet_network_create`, `vmnet_network_configuration_create`, `..._add_port_forwarding_rule`, `..._add_dhcp_reservation`, subnet/prefix/MTU/external-interface setters, NAT44/NAT66/DHCP/DNS-proxy/RA disables, plus serialization (`vmnet_network_copy_serialization` / `_create_with_serialization`) for handing networks across processes. `VZVmnetNetworkDeviceAttachment` is macOS 26.
+
+macOS 27's only change (beta 4 release notes, verbatim): "The vmnet port forwarding APIs now support port forwarding when communicating over loopback." That closes the old gap where the host could not reach its own forwarded ports via 127.0.0.1 (confirmed working by the original bug reporter). Directly relevant to Codeman's loopback-bound production server talking to per-case guests.
+
+Gotchas:
+- vmnet networks are **not persisted**; they die with the owning process. Persist settings yourself and recreate (or serialize across processes).
+- The `com.apple.vm.networking` entitlement is still restricted ("contact your Apple representative", though DTS says most requests are approved). The plain `VZNATNetworkDeviceAttachment` needs no special entitlement and is what our prototype uses.
+- Ecosystem signal: tart's maintainer is not adopting in-process vmnet (prefers their separate-process softnet), so field testing of these APIs is thin.
+
+## 5. VZCustomVirtioDevice (macOS 27, Linux guests only)
+
+14 new types (`VZCustomVirtioDevice(+Configuration/Delegate/Provider)`, `VZVirtioQueue(+Element)`, `VZVirtioFeatureSet`, shared-memory-region types, `VZGuestMemoryMapping`), wired via `VZVirtualMachineConfiguration.customVirtioDevices`. Mandatory for guest discovery: `deviceID`, `pciClassID`, `pciSubclassID`, `virtioQueueCount`. You must write the Linux guest driver (Virtio spec 1.3/1.4). Threading contract: the framework calls the device/delegate on a serial queue (`deviceQueue`, defaulting to the VM's queue). Zero public adopters. For the VM subsystem this is a Phase 3+ option for a low-latency host-guest channel; SSH over NAT is proven and sufficient for now.
+
+## 6. Signing and entitlements
+
+- **Core loop (VZ + DiskImageKit + provisioning): ad-hoc signing with only `com.apple.security.virtualization` suffices.** Verified by us on beta 4 (plain `codesign --entitlements ... -s -` on a `swiftc` binary) and independently by third parties on beta 3. DiskImageKit documents no entitlement at all.
+- **Over-entitling is the actual trap.** Adding `com.apple.application-identifier`/team-identifier keys without an embedded provisioning profile hangs the process before `main` (watchdog kill); shipping `com.apple.vm.networking` unauthorized gets AMFI SIGKILL at exec (exit 137, no crash report, even for `--version`). Keep the entitlements plist to exactly the one key.
+- **USB passthrough breaks the ad-hoc story**: `com.apple.developer.accessory-access.usb` is profile-restricted (any paid team, no ad-hoc), additionally requires `com.apple.security.device.usb`, and `AAUSBAccessoryManager` presents UI, so it wants a Dock app, not a headless CLI. Out of scope for Codeman.
+- No Xcode required for any of the above: the CLT beta (~500 MB via `softwareupdate`) carries the full macOS 27 SDK including DiskImageKit and compiles/signs everything.
+
+## 7. Ecosystem state (July 2026)
+
+- **tart is now `openai/tart`** (moved from cirruslabs, mid-2026) and **relicensed to FSL-1.1-ALv2** (no longer permissive). Provisioning support shipped in 2.33.0. Old cirruslabs URLs and license assumptions are stale.
+- VirtualBuddy shipped provisioning ("Skip Setup Assistant") in 2.2 betas; had to add account-detail validation and a workaround installer for the cross-version bug below.
+- lima is deliberately waiting for GA before touching macOS 27 APIs.
+- **Code-Hex/vz (Go bindings) is dormant** (no commits since Feb 2026, no macOS 27 APIs), so the entire Go ecosystem (podman-machine, colima) currently has no path to these APIs. Swift is the only realistic binding today, which validates the VM subsystem's Swift-helper design.
+- Useful pattern if ever supporting older SDKs: resolve new classes via `NSClassFromString` at runtime (no link-time dependency), fail gracefully when absent.
+- **Cross-version restore bug**: installing a macOS 27 guest from IPSW on a macOS 26 host fails at 77-78% (`VZErrorDomain 10007`); fixed in 26.6b3 + Xcode 27b4 era, with a nasty MobileDevice.pkg trap (installing it from Xcode 27 beta on a 26 host requires a full macOS reinstall to undo). Not relevant to our 27-host testbed, very relevant to anyone on a 26 host.
+
+## 8. Our empirical results (beta 4, 26A5388g, MacBook Air M3, 2026-07-29)
+
+Prototype tooling, all in `~/vm-lab/` on the testbed, compiled with CLT-only `swiftc` and ad-hoc signed with the single virtualization entitlement:
+
+| Tool | Purpose |
+| --- | --- |
+| `vzboot.swift` | Linux guest: EFI boot + virtio disk/net/entropy + NAT + optional cloud-init seed ISO + serial on stdio |
+| `vzstack.swift` | Same, but boots a DiskImageKit stack (read-only raw base + ASIF overlay) |
+| `vzmac.swift` | macOS guest: `install` (IPSW restore into a bundle) and `run` (boot, `--provision` for first-boot account/SSH) |
+| `vzmacgui.swift` | macOS guest in a real window via `VZVirtualMachineView` (required for the guest to render at all) |
+| `setup-seed.sh` | Builds a cloud-init NoCloud seed ISO with `hdiutil makehybrid` (volume label `cidata`) |
+| `vncproxy.py` | RFB proxy that advertises only security type 2, so version-skewed/browser clients can authenticate |
+| noVNC + `websockify` | Browser access; `websockify --web noVNC-<ver> 0.0.0.0:<port> 127.0.0.1:<proxy>` |
+| `vmwatchdog.sh` + `vmaccess.sh` | Supervision: root LaunchDaemon that restarts a blind/dead runner, re-points the forward, re-applies `pmset`, re-arms keep-awake; plus a keeper for the proxy/web endpoints |
+
+Host-side diagnostics written during this work (in the session scratchpad, not on the testbed): `vnclogin.py` (Apple DH auth + session open, distinguishes "credentials rejected" from "authorized but session refused"), `vncshot.py` (decodes the raw framebuffer to PNG and reports non-black pixel counts, plus optional synthetic wake input), `relay.py` (plain TCP relay used to bridge a tailnet peer to a LAN-only host), `sshpw.py` (pty-driven password SSH for the one-time key bootstrap into a freshly provisioned guest).
+
+### Proven working
+
+1. **Boot**: Debian 12 arm64 cloud images (nocloud and genericcloud variants) boot under `VZEFIBootLoader` + `VZGenericPlatformConfiguration`.
+2. **Networking**: `VZNATNetworkDeviceAttachment` gives the guest a `192.168.64.x` DHCP lease from the host's bootpd (leases visible in `/var/db/dhcpd_leases`, bridge is `bridge100`).
+3. **cloud-init provisioning**: NoCloud seed ISO (built with `hdiutil makehybrid -iso -joliet -default-volume-name cidata`) created a `codeman` user with SSH key + passwordless sudo on first boot; `ssh codeman@<lease-ip>` from the host works with key auth.
+4. **DiskImageKit stack mechanics**: opening a raw base `.readOnly`, appending an ASIF overlay (`ASIFCreationConfiguration.layer(url:type:.overlay)`), attaching via `init(diskImage:)`, and booting it. The overlay received ~44 MB of boot-time writes while the **base file's SHA-256 stayed bit-identical**, which is the write-isolation property the whole per-case design rests on.
+5. **Reattach**: reopening an existing overlay and `appending(consuming:)` onto the same base passes UUID validation.
+6. **macOS guest install (added later the same day)**: `VZMacOSInstaller` restore of the 27.0 IPSW (26A5388g, fetched from the seed CDN via appledb; same build as host) into a sparse 64 GiB raw disk + auxiliary storage: INSTALL-OK on the first attempt, ~25 minutes.
+7. **Headless guest provisioning WORKS**: `VZMacGuestProvisioningOptions` via `setGuestProvisioning` (username, password, `enablesRemoteLogin`, `logsInAutomatically=false`) produced, with zero GUI interaction: an account with full admin (groups include `80(admin)`, `com.apple.access_ssh`), Remote Login on from first boot, port 22 reachable ~140 s after first-boot start, hostname auto-derived from the account ("Codemans-Virtual-Machine"). SSH password auth is on by default, so the bootstrap path is: pty-driven password login once to install `authorized_keys`, key auth thereafter. Note the provisioned account's sudo is NOT passwordless (`echo <pass> | sudo -S ...`), and provisioning is first-boot-only (later boots take no options and just boot).
+8. **Slot-leak bug NOT reproduced on 26A5388g**: a guest-initiated `shutdown -h now` fired `guestDidStop` cleanly and an immediate relaunch started fine (SSH-ready again in ~75 s), so FB22967193 (VM slot leaked on guest-initiated shutdown, host reboot to recover) did not manifest after one cycle. Either fixed in beta 4 or needs more cycles to trigger.
+
+### Unstable / under investigation (beta-quality territory)
+
+Boot reliability degraded over a ~15-VM session on one host boot, ending with reproducible silent hangs (VM process alive, 0% CPU, no DHCP, no ARP, nothing on serial):
+
+- A genericcloud base that had been booted read-write once (cloud-init first boot) subsequently hung on every boot **with the seed ISO still attached**, while booting **without** the seed succeeded, then later runs failed in both configurations. The seed correlation is strong but was observed while host state was already suspect, so it needs a retest from a clean baseline.
+- The first stack-boot "success" that later wedged turned out (via DHCP lease timestamp arithmetic) never to have reached the network at all; its overlay growth was pre-network boot writes.
+- Working hypothesis, matching a class of acknowledged beta bugs (e.g. the VM-slot counter that leaks on guest-initiated shutdown, FB22967193, where only a host reboot recovers): accumulated hypervisor/vmnet state on the host degrades boots. Requires a host reboot + a disciplined retest matrix to confirm.
+
+### Display rendering: the single most important operational finding
+
+**A VZ macOS guest renders nothing unless a `VZVirtualMachineView` is attached AND the host session is actually drawing.** Verified byte-for-byte: the guest's own screen sharing serves an all-zero framebuffer (0 non-black bytes across 400 KB samples, with a sane pixel format: `rmax/gmax/bmax = 255`, shifts 16/8/0), in-guest `screencapture` fails with "could not create image from display", and no `IODisplayWrangler` shows up in the guest's `ioreg`. Three distinct states all produce black:
+
+1. **Headless** (VM run with no view attached).
+2. **View attached, host session locked.** The lock screen suspends drawing and the guest's virtual GPU produces no frames.
+3. **View attached, but the app lost its WindowServer connection** (see the incident below): black permanently until the app is restarted.
+
+**Consequence for the VM subsystem: rendering is a first-class requirement, not an optional extra (owner decision 2026-07-29).** The product serves GUI desktops: mandatory for macOS guests, optional-but-supported for Linux guests (which can also run headless over SSH). Any VM in GUI mode must be launched by an app that attaches a `VZVirtualMachineView`, from inside a host GUI session that is logged in and unlocked. That makes the following non-negotiable parts of the design, not workarounds:
+
+- VMs run as **GUI apps in the console user's session** (launched via a LaunchAgent or `launchctl asuser`), never as daemons.
+- The **host must auto-login and never lock or sleep**; a locked host is equivalent to a powered-off display for every VM on it.
+- The **guest must auto-login, never lock, and have its first-login assistant pre-suppressed**, or the "desktop" a user connects to is a password prompt or a setup wizard.
+- A VM app that loses its WindowServer connection is **permanently blind** and must be restarted; supervision has to detect that, not just check that the process is alive.
+- The **2-concurrent-macOS-VM cap** becomes a real capacity limit for the product, so it must be surfaced in the UI and tested (still untested worldwide as of this writing).
+
+### Incident 2026-07-29: `killall -HUP loginwindow` (never do this on a remote Mac)
+
+Applying a wallpaper change on the testbed with `killall -HUP loginwindow` restarted the host's login session. Three consequences:
+
+1. **The Mac dropped off the tailnet entirely.** Tailscale's App Store build is a GUI app living in the user session, so killing the session killed the VPN; remote access was gone until someone logged in. Recovery came from a second machine on the same LAN: it could still SSH in, and then relay ports back over the tailnet (a plain TCP relay on a tailnet-connected LAN peer is a good out-of-band path worth keeping ready).
+2. **The VM app lost its WindowServer connection** (`HIToolbox: received notification of WindowServer event port death`) while surviving as a process. Every later black screen traced to this, and nothing guest-side could fix it; only restarting the app restored rendering.
+3. The session's `caffeinate` died, so the host resumed auto-locking.
+
+Rule: on a remote Mac, never run session-level commands (`killall -HUP loginwindow`, `pkill -u <user>`, logout, fast user switching). `killall WallpaperAgent` alone is session-safe. Before any such command, enumerate what depends on that session: VPN, VM processes, port forwards, keep-awake helpers.
+
+### Keeping host and guest usable unattended
+
+- **Host**: `caffeinate -d -i -m -u` prevents display sleep but does NOT override the lock policy. "Require password after screen saver begins or display is turned off → Never" must be set in System Settings; it needs the account password, so a passwordless-sudo shell cannot script it, and turning it off does NOT dismiss a lock that is already engaged (one more unlock is always needed). `pmset -a disablesleep 1` keeps a lid-closed laptop awake but **does not survive a reboot**, and OS updates reset it too, so a supervisor should re-apply it rather than assume it sticks.
+- **Rebooting an encrypted host**: use `sudo fdesetup authrestart`. FileVault's pre-boot unlock doubles as the login, so the machine returns with a **live logged-in console session** and encryption intact, no password prompt, and supervision can then bring the VMs back by itself. Verified 2026-07-30. A plain `reboot` parks at the lock screen and blacks out every VM until a human logs in.
+- **Guest**: set `autoLoginUser` plus a valid `/etc/kcpassword` (XOR-obfuscated password file, key `7D 89 52 23 D2 BC DE A3`, payload zero-padded to a multiple of 12). `sysadminctl -autologin` fails with `SACSetAutoLoginPassword error:22` on provisioned accounts, and a fresh guest has no Python, so generate the bytes on the controlling host and copy them in. Then `pmset -a displaysleep 0 sleep 0 disablesleep 1`, `defaults -currentHost write com.apple.screensaver idleTime 0`, `defaults write com.apple.screensaver askForPassword 0`, and `caffeinate` inside the guest. ⚠ `autoLoginUser` was observed being wiped by failed `sysadminctl -autologin` attempts; verify it after each boot until stable.
+- **Wallpaper**: animated "aerials" wallpaper is brutal over VNC. The provider lives in `~/Library/Application Support/com.apple.wallpaper/Store/Index.plist` under several keys (`AllSpacesAndDisplays:Desktop`, `:Idle`, and `SystemDefault:*` which is what the login/lock screen uses). Switch each `Provider` to `com.apple.wallpaper.choice.solid-color` with PlistBuddy and restart `WallpaperAgent`. The login-window copy is cached and only refreshes on a later login cycle.
+
+### Remote GUI/SSH access to a guest (recipe, verified 2026-07-29)
+
+The guest lives on the host-private NAT bridge, so remote access is guest-service + host-forward:
+
+1. **In the macOS guest** (over ssh), use ONE mechanism, fully activated. The reliable form is Remote Management in a single kickstart call:
+   ```
+   sudo .../RemoteManagement/ARDAgent.app/Contents/Resources/kickstart \
+     -activate -configure -access -on \
+     -clientopts -setvnclegacy -vnclegacy yes -setvncpw -vncpw <8-char-pw> \
+     -allowAccessFor -allUsers -privs -all -restart -agent -menu
+   ```
+   ⚠ **Half-configured states authenticate but refuse the session.** Loading `com.apple.screensharing` while Remote Management is deactivated (or vice versa) produces an Apple-client error that names the wrong culprit: *"Screen Sharing is not permitted on <host>. Disable and re-enable Screen Sharing or Remote Management in System Settings"*. A raw-protocol client can still authenticate AND open a framebuffer in that state, so protocol-level tests pass while every Apple client fails. The remedy is exactly what the dialog says, done over ssh: `launchctl unload -w …screensharing.plist`, `kickstart -deactivate -configure -access -off`, `pkill screensharingd`, then the single activate call above.
+   Notes: `launchctl enable system/com.apple.screensharing` fails with "Could not find service" on this build; `load -w` is the plain-Screen-Sharing path if you deliberately want it instead of Remote Management. Apple clients negotiate `RSA-SRP` (auth type 33) and the guest logs `Authentication: SUCCEEDED :: User Name: … :: Type: RSA-SRP` on success, which is the definitive server-side confirmation.
+2. **On the host**: a gateway port-forward makes the guest's 5900 reachable from the whole tailnet without per-client tunnels: self-authorize the host's own key, then `ssh -N -g -L 0.0.0.0:5901:<guest-ip>:5900 <user>@localhost` (nohup'd).
+   ⚠⚠ **NEVER forward on host port 5900.** If the host has Screen Sharing enabled (our testbed does, from the pre-upgrade checklist), launchd already owns 5900 socket-activated. The `ssh -L` bind then fails with "Address already in use" **while the tunnel process keeps running**, so every symptom of success is present (process alive, port answers, real RFB banner) yet **every connection reaches the HOST's login window, not the guest**. This cost us an hour: guest credentials failed against the host's screensharingd, which reads exactly like broken guest auth, and we chased the (real, but irrelevant) provisioned-account identity bug. Diagnostics that would have caught it instantly: `sudo lsof -nP -iTCP:5900 -sTCP:LISTEN` showing `launchd` rather than `ssh`, or the guest's own logs showing NO auth attempts during a failed login. Always use a distinct host port and verify with `lsof` that the forward owns it.
+   ⚠ `-g` binds all interfaces, so the forward is also visible on the host's LAN; the VNC layer still requires the account or VNC password. ⚠ The forward pins the guest IP, which changes per boot under plain NAT; re-point it after a guest reboot (the proper fix is a vmnet DHCP reservation, macOS 26 API, once we move off plain `VZNATNetworkDeviceAttachment`).
+   Verified working: with the forward on 5901, both a provisioned account and a `sysadminctl`-created one authenticate successfully (RFB `SecurityResult` = 0) against the guest. The guest offers security types `[30, 33, 36, 2, 35]`, i.e. Apple DH/SRP **plus classic type 2**, so non-Apple VNC clients work with the legacy password once ARD's `-setvnclegacy` is set. (The host's screensharingd, by contrast, offered no type 2, which is itself a tell that you are talking to the wrong machine.)
+3. **SSH from any tailnet device**: `ssh -J <host-user>@<host> codeman@<guest-ip>` (jump through the host), after adding the connecting machine's key to the guest's `authorized_keys`.
+
+**Client-version incompatibility (macOS 27 servers vs older Screen Sharing clients)**: an older Mac's Screen Sharing client fails Apple's `RSA-SRP` handshake against macOS 27 servers, logging `Authentication: FAILED :: User Name: <user> :: Type: RSA-SRP` server-side, while a macOS 27 client authenticates against the same servers without issue. This was verified against BOTH a macOS 27 guest and a macOS 27 host with the operator's own account, so it is a client-side version skew, not configuration, and no server-side change fixes it. Same family as the documented "macOS 26 host cannot install a 27 guest" bug. Practical workaround: bypass Apple auth entirely with classic VNC auth (security type 2), which macOS offers only when Remote Management legacy VNC is enabled. Two ways to consume it: any third-party VNC client, or a browser via noVNC.
+
+**Browser-based access chain (zero client install, version-proof)**, all hosted on the Mac:
+```
+browser --HTTP/WS--> websockify (+ noVNC static files)
+                       --> type-2-only proxy   # rewrites the server's security-type list to [2]
+                         --> ssh -L forward     # loopback hop; see the Local Network note below
+                           --> guest:5900
+```
+Notes learned the hard way: (a) **never bind the forward on host port 5900** (see the launchd warning above); (b) a Python proxy cannot reach the guest subnet directly because macOS **Local Network privacy** denies headless CLI binaries, surfacing as `No route to host`, so point the proxy at a loopback `ssh -L` forward instead (Apple-signed `ssh` is unaffected); (c) noVNC needs `?resize=scale` or Scaling Mode → Local Scaling, otherwise a Retina host screen (2940x1912) is unusable in a browser window; (d) noVNC speaks security type 2 only, which is exactly why the proxy rewrite is needed.
+
+**Debugging technique that settled all of this**: a ~80-line Python RFB client (scratchpad `vnclogin.py`) that implements Apple DH auth (security type 30) and continues through `ClientInit`/`ServerInit`. It reports the server's `SecurityResult` plus the framebuffer size and desktop name, which separates "credentials rejected" from "authorized but session refused" without any GUI client. Pair it with `log stream --predicate 'process == "screensharingd"'` inside the guest, and drive a REAL Apple client headlessly from the host with `sudo launchctl asuser <uid> sudo -u <user> osascript -e 'tell application "Screen Sharing" to open location "vnc://user:pass@host:port"'`, verifying the result via `lsof -nP -iTCP -a -p <pid>` (an ESTABLISHED socket to the target) since `screencapture` fails on a lid-closed laptop ("could not create image from display"). Tailscale was never implicated: both the raw client and Apple's client work over the tailnet address once the guest service is fully activated.
+
+### Hard-won operational lessons (write these into any tooling)
+
+- **Silent serial is normal, not failure.** Debian's GRUB/kernel log to the graphics console; nothing attaches a getty to hvc0 by default. The reliable boot signal is the DHCP lease (or passive `tcpdump -i bridge100`), never the serial port and never a quick ping (BSD ping's first packet often dies to ARP latency; passive capture showed "dead" guests alive).
+- **DHCP lease entries carry truth**: `name=` shows the guest hostname, and the lease timestamps order events; stale entries linger, so compare timestamps before attributing a lease to a boot.
+- **Never boot a base image read-write.** Every RW boot mutates it (dhclient lease cache, journal, cloud-init state) and destroys experiment reproducibility, exactly why the production design only ever boots bases under overlays. Provision INTO the base once at base-build time, or provision per-case overlays with the seed, then detach the seed.
+- **A killed SSH client does not kill a remote `nohup`'d VM**, and the survivor holds the EFI variable store lock: "The EFI variable store is already in use" (`VZErrorDomain 50002`) means a zombie VM process, `pkill` it.
+- **EFI variable stores are per-VM state.** Fresh stores boot reliably; reuse across different VM instances is at minimum suspect on this beta (Apple's own guidance for cloned VMs is one store per VM). Cheap policy: one store per case, created with the overlay, deleted with it.
+- **Downloads from cloud.debian.org mirrors truncate silently**; always verify byte count against origin `Content-Length` and resume with `curl -C -`.
+- The remote host's default shell is zsh: `=` -prefixed words (`echo ===`) explode via zsh's `=cmd` expansion; keep separators zsh-safe in automation.
+
+### The 2-concurrent-macOS-VM cap: TESTED AND CONFIRMED on macOS 27 beta 4 (2026-07-29)
+
+We measured it, which as far as we can tell nobody had published for macOS 27. Method: `cp -c -R` the guest bundle (APFS clonefile, instant and **zero additional disk**), regenerate the machine identifier per clone (`VZMacMachineIdentifier()` written to `machine.id`; the hardware model is reused), then launch VMs until one is refused.
+
+Result: VM #1 (8 GB, GUI) and VM #2 (4 GB, headless) ran concurrently without complaint. VM #3 was refused **instantly** at `vm.start`:
+
+```
+VZErrorDomain Code=6 "The maximum supported number of active virtual machines has been reached."
+NSLocalizedFailure = "The number of virtual machines exceeds the limit."
+```
+
+**This is a licensing/kernel quota, not a resource limit**: the refusal came with **39% of system memory free** on a 16 GB host, and adding RAM or CPU cannot raise it. It matches the pre-27 behavior (`hv_apple_isa_vm_quota`), so nothing changed in 27 despite the framework's other additions. Linux guests are unaffected and are bounded only by host resources.
+
+Design consequences: macOS-guest capacity per host is **hard-capped at 2**, so a GUI-macOS-per-case product must schedule around it (queue, evict idle VMs, or scale across hosts) and surface it in the UI. Also relevant: the acknowledged slot-leak bug (a guest-initiated shutdown failing to release a slot, recoverable only by host reboot) is far more damaging under a cap of 2 than it sounds; we did not reproduce it on beta 4, but any scheduler should treat "slot appears used but nothing is running" as a real state.
+
+### Not yet tested
+- Cache layers (`LayerType.cache`), `.overlay(blockCount:)` disk growth, stack depth performance, VirtioFS + stack combination, `truncate`, ASIF disks for macOS guests (raw used so far; ASIF has the reclamation bug).
+- One more scripting lesson from this session: inner `ssh` calls inside a piped `sh -s` script MUST use `-n`, or they consume the remainder of the script from stdin and it silently never runs.
+
+### Session timeline (what was actually established, 2026-07-29)
+
+Linux path: base image download (with resume, mirrors truncate) → `vzboot` compiles against the beta SDK first try → EFI boot → NAT DHCP lease → cloud-init seed provisions a user with the host's SSH key → `ssh` into the guest works → DiskImageKit stack boots with an ASIF overlay taking all writes while the base stays SHA-identical. Later Linux boots became unreliable on an un-rebooted host (silent hangs, 0% CPU, no DHCP); a clean-baseline retest is still pending.
+
+macOS path: seed-CDN IPSW (matched to the host build) → `VZMacOSInstaller` restore, ~25 min, first try → first boot with `VZMacGuestProvisioningOptions` creates an admin account with Remote Login on, no interaction needed, SSH reachable ~140 s later → key bootstrap over a one-time password login → guest shutdown/relaunch clean (the slot-leak bug did not reproduce) → GUI access fought through a port collision, a client-version incompatibility, the rendering dependency, and a self-inflicted session kill, ending with a browser-based path plus a guest hardened to auto-login and never lock.
+
+**Lifecycle verified (stop → start), 2026-07-30**: an in-guest `shutdown -h now` fires `guestDidStop` and the runner app exits on its own; relaunching from the same bundle boots the guest in ~2 minutes straight into an auto-logged-in desktop, and the VM slot is released cleanly (an immediate restart works, so the slot-leak bug did not bite). Two operational notes: the guest takes a **new NAT lease on every boot**, so any port-forward must be re-pointed (or use a vmnet DHCP reservation), and a host reboot resets `pmset -a disablesleep`.
+
+⚠ **Provisioning does NOT skip the per-user first-login assistant.** `VZMacGuestProvisioningOptions` skips the initial Setup Assistant (account creation, region, Apple Account) so the machine is immediately reachable, but the first time anyone actually logs into a desktop, macOS still presents its per-user wizard (Apple Intelligence, Siri, privacy, appearance, Touch ID). The operator hit exactly this. For a GUI-first product this MUST be pre-suppressed during base-image creation by writing `com.apple.SetupAssistant` keys for every account that will log in, and into `/System/Library/User Template/English.lproj/Library/Preferences/` so accounts created later inherit it.
+
+⚠ **A partial key list is worse than none**, because the wizard simply shows the panes you missed and the operator has to click through them again after every fresh login (we hit this twice). The set that finally silenced macOS 27 beta 4: `DidSeeCloudSetup`, `DidSeeSiriSetup`, `DidSeePrivacy`, `DidSeeAppearanceSetup`, `DidSeeTouchIDSetup`, `DidSeeAvatarSetup`, `DidSeeScreenTime`, `DidSeeApplePaySetup`, `DidSeeSafariImport`, `DidSeeAccessibility`, **`DidSeeActivationLock`, `DidSeeAppStore`, `DidSeeLockdownMode`** (the three easy to miss), plus the Express-Settings flags **`SkipExpressSettingsUpdating`** and **`SkipFirstLoginOptimization`**, and the version markers `LastSeenCloudProductVersion` / `LastSeenBuddyBuildVersion` / `PreviousSystemVersion` / `PreviousBuildVersion` matching the guest build. Verify afterwards by reading the domain back and checking that no `DidSee*` key is still `0`. Note these keys change between macOS releases, so base-image creation should re-verify per OS version rather than trust a hardcoded list.
+
+## 9. Design implications for Codeman's VM subsystem
+
+0. **GUI is a first-class mode, and for macOS guests it is the whole point (owner decision, 2026-07-29).** The subsystem serves real desktops, not only headless SSH boxes. macOS guests are GUI-only in practice (nothing renders without an attached view). Linux guests are supported in BOTH modes: GUI when the case wants a desktop, headless-over-SSH when it wants a cheap agent sandbox. The costs of the GUI path are in §8 "Display rendering": VMs as GUI apps in a live session, a host that never locks, guests that auto-login with their first-login wizard pre-suppressed, and the macOS concurrency cap as a real capacity limit.
+1. **The macOS-specific liabilities are accepted costs, not reasons to avoid macOS guests**: provisioning is macOS-only and first-boot-only, ASIF space reclamation is broken for macOS guests on the beta (use RAW disks for macOS guests until fixed), and the 2-VM cap applies. Plan around each: RAW-backed macOS disks, provisioning baked into base-image creation, and capacity limits surfaced in the UI.
+2. **Base immutability is not just hygiene, it is load-bearing**: DiskImageKit's UUID invalidation plus our sha-stability proof make a read-only shared base per image-generation the core artifact. Bases are built once (seed attached), then only ever opened `.readOnly` under per-case overlays.
+3. **Seed ISOs are a base-build-time tool only.** Never attach a seed to a routine case boot (correlated with boot hangs on the beta, and semantically wrong anyway since cloud-init already ran).
+4. **Per-case files**: overlay ASIF + EFI variable store live and die together with the case.
+5. **Export = ship the layer chain** (base ref + overlay + manifest), not flatten; there is no flatten API. In-guest `dd` to a fresh image is the fallback for a true single-file export.
+6. **Health checking must be lease/API based**, not serial/ping based, and Codeman's `codeman-vm status` should read `/var/db/dhcpd_leases` (or use vmnet DHCP reservations for deterministic per-case IPs, a macOS 26 API).
+7. **Run `fstrim` periodically in Linux guests** (or mount with discard) so overlays stay sparse.
+8. **Entitlements plist stays minimal** (exactly `com.apple.security.virtualization`) to dodge the AMFI/watchdog traps.
+9. **Expect beta churn**: pin findings to build numbers (this doc: 26A5388g) and retest each beta; the framework binaries changed every beta so far.
+10. **A macOS guest is only "ready" when its desktop is ready**, which is a stricter bar than "the VM booted". Readiness means: VM app running with a live WindowServer connection, guest auto-logged-in (not at a login or lock screen), first-login assistant suppressed, and the guest's screen sharing serving a non-black framebuffer. Health checks should sample the framebuffer for non-black content, because every failure mode in this session (headless run, locked host, dead WindowServer, locked guest, setup wizard) presents as a perfectly healthy-looking process with a black or useless screen.
+10b. **Supervision must run as a root LaunchDaemon.** A user LaunchAgent cannot launch a GUI app into the Aqua session; its restarts fail silently (child dies instantly, empty log, supervisor reports success). Root + `launchctl asuser <uid> sudo -u <user> …` works and the launched process persists. This bit us on the first supervisor implementation and is easy to repeat.
+
+11. **Remote-access plumbing belongs in the helper CLI, not in ad-hoc shell**: a `codeman-vm` implementation should own port selection (never 5900), forward lifecycle across guest IP changes (or better, vmnet DHCP reservations for stable per-case IPs), and a documented browser path, because every failure in this session came from hand-rolled plumbing rather than from the Virtualization APIs themselves.
+12. **Never let control-plane connectivity depend on a GUI session** on a remote Mac host: prefer a Tailscale system service over the App Store app, and keep a LAN-adjacent peer able to relay as an out-of-band recovery path.
+
+## Sources
+
+Apple DocC JSON backend (diskimagekit, virtualization, vmnet trees; macOS 27 release notes) | WWDC26 session 224 https://developer.apple.com/videos/play/wwdc2026/224/ | eclecticlight.co ASIF/virtualization coverage | developer.apple.com/forums threads 839343 (CSIdentity bug), 830118 (cross-version restore), 830119 (VM-slot leak), 830383 (VM cap), 834822 + 831902 (USB entitlements), 822658 (vmnet loopback) | openai/tart issues 1261/1263/1268/1269/1285 | Spooky-Labs provisioning design doc | VirtualBuddy 2.2 release notes | lima-vm discussions | our own test transcripts on the testbed (`~/vm-lab/*.log`, this repo's session)

docs/web-tabs.md

@@ -1,7 +1,7 @@
 # Web Tabs (dashboards as Codeman tabs)
 
 Open any dashboard you run, Grafana, Uptime Kuma, Portainer, a status page on port
-4000, as a tab beside your Claude/Codex/Gemini sessions. Codeman becomes one mission
+4000, as a tab beside your Claude/Codex/Antigravity sessions. Codeman becomes one mission
 control instead of Codeman plus a pile of browser tabs.
 
 ## Using it

install.sh

@@ -116,6 +116,14 @@ GEMINI_SEARCH_PATHS=(
     "$HOME/bin/gemini"
 )
 
+# Antigravity CLI search paths (from src/utils/antigravity-cli-resolver.ts)
+ANTIGRAVITY_SEARCH_PATHS=(
+    "$HOME/.local/bin/agy"
+    "$HOME/.antigravity/bin/agy"
+    "/usr/local/bin/agy"
+    "$HOME/bin/agy"
+)
+
 # ============================================================================
 # Color Output
 # ============================================================================
@@ -493,6 +501,34 @@ get_gemini_path() {
     done
 }
 
+check_antigravity() {
+    if command -v agy &>/dev/null; then
+        return 0
+    fi
+
+    for path in "${ANTIGRAVITY_SEARCH_PATHS[@]}"; do
+        if [[ -x "$path" ]]; then
+            return 0
+        fi
+    done
+
+    return 1
+}
+
+get_antigravity_path() {
+    if command -v agy &>/dev/null; then
+        command -v agy
+        return
+    fi
+
+    for path in "${ANTIGRAVITY_SEARCH_PATHS[@]}"; do
+        if [[ -x "$path" ]]; then
+            echo "$path"
+            return
+        fi
+    done
+}
+
 check_cloudflared() {
     # Check ~/.local/bin first (matches tunnel-manager.ts resolution order)
     if [[ -x "$HOME/.local/bin/cloudflared" ]]; then
@@ -1993,11 +2029,12 @@ main() {
         fi
     fi
 
-    # AI CLI (Codeman drives one of: Claude Code, OpenCode, Codex, Gemini)
+    # AI CLI (Codeman drives one of: Claude Code, OpenCode, Codex, Gemini, Antigravity)
     local has_claude=false
     local has_opencode=false
     local has_codex=false
     local has_gemini=false
+    local has_antigravity=false
 
     info "Checking AI CLI tools..."
     if check_claude; then
@@ -2016,17 +2053,21 @@ main() {
         has_gemini=true
         success "Gemini CLI found at $(get_gemini_path)"
     fi
+    if check_antigravity; then
+        has_antigravity=true
+        success "Antigravity CLI found at $(get_antigravity_path)"
+    fi
 
-    if [[ "$has_claude" == "false" && "$has_opencode" == "false" && "$has_codex" == "false" && "$has_gemini" == "false" ]]; then
+    if [[ "$has_claude" == "false" && "$has_opencode" == "false" && "$has_codex" == "false" && "$has_gemini" == "false" && "$has_antigravity" == "false" ]]; then
         echo ""
-        warn "No AI CLI found. Codeman needs at least one: Claude Code, OpenCode, Codex, or Gemini."
+        warn "No AI CLI found. Codeman needs at least one: Claude Code, OpenCode, Codex, Antigravity, or Gemini."
         headless_guard "install an AI CLI (curl | bash from its vendor)"
         echo ""
         echo -e "  ${BOLD}Which AI CLI would you like to install?${NC}"
         echo -e "    ${CYAN}1)${NC} Claude Code  (Anthropic)"
         echo -e "    ${CYAN}2)${NC} OpenCode     (open-source)"
         echo -e "    ${CYAN}3)${NC} Both"
-        echo -e "    ${CYAN}4)${NC} Skip         (I'll install one myself, e.g. Codex or Gemini)"
+        echo -e "    ${CYAN}4)${NC} Skip         (I'll install one myself, e.g. Codex or Antigravity)"
         echo ""
 
         local cli_choice=""
@@ -2071,8 +2112,8 @@ main() {
 
         if [[ "$cli_choice" == "4" ]]; then
             warn "Skipping AI CLI install. Codeman will run, but sessions need a CLI to drive."
-            info "Install one later, e.g.: npm install -g @openai/codex        (Codex)"
-            info "                    or: npm install -g @google/gemini-cli   (Gemini)"
+            info "Install one later, e.g.: npm install -g @openai/codex                          (Codex)"
+            info "                    or: curl -fsSL https://antigravity.google/cli/install.sh | bash  (Antigravity)"
         elif [[ "$has_claude" == "false" ]] && [[ "$has_opencode" == "false" ]]; then
             die "The selected AI CLI failed to install. Install one manually and re-run the installer."
         fi
@@ -2372,12 +2413,12 @@ main() {
     echo -e "    https://github.com/Ark0N/Codeman"
     echo ""
 
-    if ! check_claude && ! check_opencode && ! check_codex && ! check_gemini; then
+    if ! check_claude && ! check_opencode && ! check_codex && ! check_gemini && ! check_antigravity; then
         echo -e "  ${YELLOW}${BOLD}Reminder:${NC} Install at least one AI CLI to start using Codeman:"
-        echo -e "    ${CYAN}curl -fsSL https://claude.ai/install.sh | bash${NC}  # Claude Code"
-        echo -e "    ${CYAN}curl -fsSL https://opencode.ai/install | bash${NC}   # OpenCode"
-        echo -e "    ${CYAN}npm install -g @openai/codex${NC}                    # Codex"
-        echo -e "    ${CYAN}npm install -g @google/gemini-cli${NC}               # Gemini"
+        echo -e "    ${CYAN}curl -fsSL https://claude.ai/install.sh | bash${NC}                # Claude Code"
+        echo -e "    ${CYAN}curl -fsSL https://opencode.ai/install | bash${NC}                 # OpenCode"
+        echo -e "    ${CYAN}npm install -g @openai/codex${NC}                                  # Codex"
+        echo -e "    ${CYAN}curl -fsSL https://antigravity.google/cli/install.sh | bash${NC}   # Antigravity"
         echo ""
     fi
 

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "aicodeman",
-  "version": "1.11.1",
+  "version": "1.11.2",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "aicodeman",
-      "version": "1.11.1",
+      "version": "1.11.2",
       "hasInstallScript": true,
       "license": "MIT",
       "workspaces": [

package.json

@@ -1,6 +1,6 @@
 {
   "name": "aicodeman",
-  "version": "1.11.1",
+  "version": "1.11.2",
   "description": "Mission control for AI coding agents - run 20 autonomous agents with real-time monitoring and session persistence",
   "type": "module",
   "main": "dist/index.js",
@@ -55,6 +55,7 @@
     "anthropic",
     "opencode",
     "codex",
+    "antigravity",
     "gemini-cli",
     "ai-agents",
     "agent",

packages/xterm-zerolag-input/README.md

@@ -16,7 +16,7 @@
 
 > ### Made for [**Codeman**](https://getcodeman.com)
 >
-> This overlay is the local echo engine of [**Codeman**](https://github.com/Ark0N/Codeman), mission control for AI coding agents: run and monitor a dozen Claude Code, Codex, OpenCode and Gemini sessions at once, watch their subagents work in live floating windows, let them run autonomously overnight, and drive all of it from your phone.
+> This overlay is the local echo engine of [**Codeman**](https://github.com/Ark0N/Codeman), mission control for AI coding agents: run and monitor a dozen Claude Code, Codex, OpenCode and Antigravity sessions at once, watch their subagents work in live floating windows, let them run autonomously overnight, and drive all of it from your phone.
 >
 > That last part is why this library exists. The demo below is a real Codeman session on two phones.
 

src/docker-hosts.ts

@@ -596,6 +596,9 @@ interface CredStorePolicy {
 
 const CRED_STORES: CredStorePolicy[] = [
   { rel: '.codex', shareDirs: ['sessions'], shareFiles: ['history.jsonl'], seedFiles: ['auth.json', 'config.toml'] },
+  // Also covers Antigravity: `agy` nests its whole state (auth `jetski_state.pbtxt`,
+  // `conversations/`, `knowledge/`) under `~/.gemini/antigravity-cli/`, so it needs no
+  // entry of its own. There is no `~/.antigravity` credential dir to add.
   { rel: '.gemini', seedWhole: true },
   { rel: '.config/gcloud', seedWhole: true },
   { rel: '.config/opencode', seedWhole: true },

src/tmux-manager.ts

@@ -982,7 +982,7 @@ export function dockerTmuxSessionName(sessionId: string): string {
 const RESUME_ID_SAFE = /^[A-Za-z0-9._-]+$/;
 
 /**
- * Append the CLI-specific resume flag to a pane command (codex/gemini). Only fires
+ * Append the CLI-specific resume flag to a pane command (codex/gemini/antigravity). Only fires
  * when the in-container tmux is RE-CREATED (`new-session -A` makes the flag inert
  * on a live reattach), i.e. exactly when the previous live agent was lost and we
  * want to resume the conversation from the bind-mounted transcript. Claude mode

src/utils/claude-cli-resolver.ts

@@ -28,7 +28,7 @@ let _claudeDir: string | null = null;
 
 /**
  * Returns true if the Claude CLI binary can be located (via `which` or one of
- * the common install directories). Mirrors `isGeminiAvailable`/`isOpenCodeAvailable`/
+ * the common install directories). Mirrors `isGeminiAvailable`/`isAntigravityAvailable`/`isOpenCodeAvailable`/
  * `isCodexAvailable` in the sibling resolvers.
  */
 export function isClaudeAvailable(): boolean {

src/utils/cloudflared-resolver.ts

@@ -1,7 +1,7 @@
 /**
  * @fileoverview Resolve the `cloudflared` binary across common install paths.
  *
- * Mirrors the CLI resolvers (gemini-cli-resolver.ts et al), for the same reason
+ * Mirrors the CLI resolvers (antigravity-cli-resolver.ts et al), for the same reason
  * they exist: the welcome screen should not offer a button whose only possible
  * outcome is an error toast.
  *

src/web/public/app.js

@@ -1923,6 +1923,37 @@ class CodemanApp {
     }
   }
 
+  /** Build one response-viewer message so the brief and full views share markup and CSS. */
+  _buildResponseViewerMessage(text, role, agentLabel) {
+    const div = document.createElement('div');
+    const isUser = role === 'user';
+    div.className = 'rv-message ' + (isUser ? 'rv-msg-user' : 'rv-msg-assistant');
+
+    const roleBadge = document.createElement('div');
+    roleBadge.className = 'rv-role ' + (isUser ? 'rv-role-user' : 'rv-role-assistant');
+    roleBadge.textContent = isUser ? 'You' : agentLabel;
+    div.appendChild(roleBadge);
+
+    const renderedText = document.createElement('div');
+    renderedText.className = 'rv-text';
+    renderedText.innerHTML = this._renderMarkdown(text);
+    div.appendChild(renderedText);
+    return div;
+  }
+
+  _getResponseViewerAgentLabel() {
+    const mode = this.sessions.get(this.activeSessionId)?.mode;
+    return mode === 'codex'
+      ? 'Codex'
+      : mode === 'gemini'
+        ? 'Gemini'
+        : mode === 'antigravity'
+          ? 'Antigravity'
+          : mode === 'opencode'
+            ? 'OpenCode'
+            : 'Claude';
+  }
+
   async toggleResponseViewer() {
     const viewer = document.getElementById('responseViewer');
     const backdrop = document.getElementById('responseViewerBackdrop');
@@ -1945,7 +1976,7 @@ class CodemanApp {
       // Source 2: Terminal buffer fallback — strip ANSI, drop Claude CLI chrome.
       // Claude + shell only: _cleanTerminalBuffer knows Claude CLI's output, and
       // shell sessions have no transcript source at all; for TUI modes
-      // (codex/opencode/gemini) it yields repaint garbage, so a clear
+      // (codex/opencode/gemini/antigravity) it yields repaint garbage, so a clear
       // placeholder beats a messy screen dump there.
       const sessionMode = this.sessions.get(this.activeSessionId)?.mode || 'claude';
       if (!lastResponse && (sessionMode === 'claude' || sessionMode === 'shell')) {
@@ -1958,7 +1989,11 @@ class CodemanApp {
 
       const body = document.getElementById('responseViewerBody');
       if (lastResponse) {
-        body.innerHTML = this._renderMarkdown(lastResponse);
+        // Keep the brief view inside the same message wrapper as the full
+        // conversation view. The wrapper supplies the card, role badge and
+        // descendant markdown styles that direct body children do not get.
+        body.innerHTML = '';
+        body.appendChild(this._buildResponseViewerMessage(lastResponse, 'assistant', this._getResponseViewerAgentLabel()));
         this._bindResponseViewerInteractions(body);
       } else {
         body.textContent =
@@ -1998,26 +2033,10 @@ class CodemanApp {
       }
 
       // Render conversation thread
-      const mode = this.sessions.get(this.activeSessionId)?.mode;
-      const agentLabel =
-        mode === 'codex' ? 'Codex' : mode === 'gemini' ? 'Gemini' : mode === 'antigravity' ? 'Antigravity' : mode === 'opencode' ? 'OpenCode' : 'Claude';
+      const agentLabel = this._getResponseViewerAgentLabel();
       body.innerHTML = '';
       for (const msg of messages) {
-        const div = document.createElement('div');
-        const isUser = msg.role === 'user';
-        div.className = 'rv-message ' + (isUser ? 'rv-msg-user' : 'rv-msg-assistant');
-
-        const role = document.createElement('div');
-        role.className = 'rv-role ' + (isUser ? 'rv-role-user' : 'rv-role-assistant');
-        role.textContent = isUser ? 'You' : agentLabel;
-        div.appendChild(role);
-
-        const text = document.createElement('div');
-        text.className = 'rv-text';
-        text.innerHTML = this._renderMarkdown(msg.text);
-        div.appendChild(text);
-
-        body.appendChild(div);
+        body.appendChild(this._buildResponseViewerMessage(msg.text, msg.role, agentLabel));
       }
       this._bindResponseViewerInteractions(body);
 
@@ -4262,7 +4281,7 @@ class CodemanApp {
       // (viewport + scrollback + colors) for an instant first paint. For codex
       // this is also a correctness fix — its byte-stream replay shows only the
       // latest TUI frame (the idle welcome banner) because codex doesn't include
-      // earlier conversation in its current redraw. For claude/opencode/gemini
+      // earlier conversation in its current redraw. For claude/opencode/gemini/antigravity
       // the replay is already complete, so the snapshot is purely a faster,
       // scroll-preserving first paint before the canonical fetch reconciles.
       //

src/web/public/index.html

@@ -323,6 +323,10 @@
               <svg width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><polygon points="5 3 19 12 5 21 5 3"/></svg>
               Run OpenCode
             </button>
+            <button class="welcome-btn welcome-btn-antigravity" id="welcomeAntigravityBtn" style="display: none;" onclick="app.setRunMode('antigravity'); app.runAntigravity()">
+              <svg width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><polygon points="5 3 19 12 5 21 5 3"/></svg>
+              Run Antigravity
+            </button>
             <button class="welcome-btn welcome-btn-gemini" id="welcomeGeminiBtn" style="display: none;" onclick="app.setRunMode('gemini'); app.runGemini()">
               <svg width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><polygon points="5 3 19 12 5 21 5 3"/></svg>
               Run Gemini
@@ -2192,7 +2196,7 @@
             <div class="form-row">
               <label>Image</label>
               <input type="text" id="dockerImage" placeholder="codeman/agent:base" autocomplete="off" autocapitalize="off" spellcheck="false">
-              <span class="form-hint">Build it once with <code>node scripts/build-agent-image.mjs</code>. Contains node + claude/codex/gemini + tmux.</span>
+              <span class="form-hint">Build it once with <code>node scripts/build-agent-image.mjs</code>. Contains node + claude/codex/gemini/opencode/agy + tmux.</span>
             </div>
             <div class="form-row">
               <label>Network</label>

src/web/public/session-ui.js

@@ -1,5 +1,5 @@
 /**
- * @fileoverview Quick start (case loading, session spawning for Claude/Shell/OpenCode/Codex/Gemini),
+ * @fileoverview Quick start (case loading, session spawning for Claude/Shell/OpenCode/Codex/Gemini/Antigravity),
  * session options modal (per-session settings, color picker, rename),
  * session options tabs (Ralph config tab), case settings (CRUD, links),
  * create case modal, and mobile case picker.

src/web/public/settings-ui.js

@@ -748,6 +748,7 @@ Object.assign(CodemanApp.prototype, {
     const buttons = [
       ['welcomeClaudeBtn', 'claude'],
       ['welcomeOpencodeBtn', 'opencode'],
+      ['welcomeAntigravityBtn', 'antigravity'],
       ['welcomeGeminiBtn', 'gemini'],
       // Not a run mode, same reasoning: offering a Cloudflare Tunnel on a box
       // without cloudflared can only ever produce "cloudflared not found".

src/web/public/styles.css

@@ -3305,6 +3305,23 @@ body.touch-device .terminal-container .xterm .xterm-helper-textarea {
   transform: translateY(-1px);
 }
 
+/* Antigravity: cyan identity, matching .btn-toolbar.btn-run.mode-antigravity and
+   .run-mode-dot.antigravity so the welcome action reads as the same backend. */
+.welcome-btn-antigravity {
+  background: linear-gradient(135deg, #0b2b33 0%, #0e7490 55%, #0891b2 100%);
+  border-color: rgba(34, 211, 238, 0.4);
+  color: #cffafe;
+  box-shadow: 0 2px 8px rgba(34, 211, 238, 0.16), inset 0 1px 0 rgba(255, 255, 255, 0.06);
+}
+
+.welcome-btn-antigravity:hover {
+  background: linear-gradient(135deg, #124450 0%, #0891b2 55%, #06b6d4 100%);
+  box-shadow: 0 4px 20px rgba(34, 211, 238, 0.3), 0 0 40px rgba(8, 145, 178, 0.12), inset 0 1px 0 rgba(255, 255, 255, 0.08);
+  border-color: rgba(103, 232, 249, 0.5);
+  color: #ecfeff;
+  transform: translateY(-1px);
+}
+
 .welcome-btn-gemini {
   background: linear-gradient(135deg, #10243f 0%, #174ea6 55%, #4f46e5 100%);
   border-color: rgba(96, 165, 250, 0.4);

src/web/public/terminal-ui.js

@@ -1408,7 +1408,7 @@ Object.assign(CodemanApp.prototype, {
     }
     titleSpan.appendChild(document.createTextNode(s.name || s.firstPrompt || shortDir));
 
-    // Badge row: mode (claude/codex/opencode/gemini/shell) + a LIVE pill.
+    // Badge row: mode (claude/codex/opencode/gemini/antigravity/shell) + a LIVE pill.
     const badgeRow = document.createElement('div');
     badgeRow.className = 'history-item-badges';
     if (s.mode) {

src/web/routes/session-routes.ts

@@ -415,7 +415,7 @@ export function registerSessionRoutes(
     //
     // For keys the caller is actively setting, strip any stale disk entry a prior
     // Codeman version may have written. Scope limited to:
-    //   - Claude mode (OpenCode/Codex/Gemini don't read .claude/settings.local.json)
+    //   - Claude mode (OpenCode/Codex/Gemini/Antigravity don't read .claude/settings.local.json)
     //   - workingDir inside CASES_DIR / the per-user case space (Codeman's managed
     //     territory — we never mutate .claude/settings.local.json in arbitrary user
     //     repos that POST /api/sessions can target, as those may have hand-authored
@@ -2027,7 +2027,7 @@ export function registerSessionRoutes(
       if (!host) return createErrorResponse(ApiErrorCode.NOT_FOUND, 'Remote host not found');
 
       // Per-session config that is applied to the LOCAL tmux/CLI wrapper (env vars via
-      // tmux setenv, effort/model CLI args, codex/gemini/opencode config) does NOT
+      // tmux setenv, effort/model CLI args, codex/gemini/antigravity/opencode config) does NOT
       // cross ssh, so it would silently no-op. Reject rather than pretend it worked —
       // remote command/env customization goes through the per-host command override.
       if (
@@ -2364,7 +2364,7 @@ export function registerSessionRoutes(
         });
         ctx.broadcast(SseEvent.SessionInteractive, { id: session.id, mode: 'shell' });
       } else {
-        // 'claude', 'opencode', 'codex', and 'gemini' modes use startInteractive()
+        // 'claude', 'opencode', 'codex', 'gemini', and 'antigravity' modes use startInteractive()
         await session.startInteractive();
         getLifecycleLog().log({
           event: 'started',

src/web/schemas.ts

@@ -516,7 +516,7 @@ export const DockerHostSchema = z.object({
   mountCredentials: z.boolean().optional(),
   hooksEnabled: z.boolean().optional(),
   resumeOnStart: z.boolean().optional(),
-  commands: RemoteCommandOverridesSchema, // same shell/claude/opencode/codex/gemini shape
+  commands: RemoteCommandOverridesSchema, // same shell/claude/opencode/codex/gemini/antigravity shape
   extraCreateArgs: z
     .array(
       z

test/frontend-public-tooling.test.ts

@@ -20,6 +20,15 @@ describe('frontend public asset tooling', () => {
     expect(appJs.includes(0)).toBe(false);
   });
 
+  it('uses the same message wrapper for brief and full response views', () => {
+    const appJs = readFileSync(resolve(repoRoot, 'src/web/public/app.js'), 'utf8');
+
+    expect(appJs).toContain("body.appendChild(this._buildResponseViewerMessage(lastResponse, 'assistant'");
+    expect(appJs).toContain('body.appendChild(this._buildResponseViewerMessage(msg.text, msg.role, agentLabel));');
+    expect(appJs).toContain("div.className = 'rv-message ' + (isUser ? 'rv-msg-user' : 'rv-msg-assistant');");
+    expect(appJs).toContain("renderedText.className = 'rv-text';");
+  });
+
   it('runs the public asset check script', () => {
     expect(() => {
       execFileSync('npm', ['run', 'check:public-assets', '--silent'], {

test/run-mode-ui.test.ts

@@ -351,7 +351,13 @@ describe('Codex quick start settings', () => {
     function loadUi(flags: Record<string, boolean> | undefined) {
       const CodemanApp = function CodemanApp(this: any) {};
       const welcomeBtns: Record<string, { style: { display: string } }> = {};
-      for (const id of ['welcomeClaudeBtn', 'welcomeOpencodeBtn', 'welcomeGeminiBtn', 'welcomeTunnelBtn']) {
+      for (const id of [
+        'welcomeClaudeBtn',
+        'welcomeOpencodeBtn',
+        'welcomeAntigravityBtn',
+        'welcomeGeminiBtn',
+        'welcomeTunnelBtn',
+      ]) {
         welcomeBtns[id] = { style: { display: 'PRISTINE' } };
       }
       const modeBtns: Record<string, { style: { display: string } }> = {};
@@ -394,6 +400,7 @@ describe('Codex quick start settings', () => {
       app.applyWelcomeCliVisibility();
       expect(welcomeBtns.welcomeClaudeBtn.style.display).toBe('flex');
       expect(welcomeBtns.welcomeOpencodeBtn.style.display).toBe('none');
+      expect(welcomeBtns.welcomeAntigravityBtn.style.display).toBe('none');
       expect(welcomeBtns.welcomeGeminiBtn.style.display).toBe('none');
       // #200 originally DELETED the tunnel button and its QR outright; it is gated
       // on cloudflared instead, so a box that has cloudflared keeps the feature.
@@ -402,6 +409,12 @@ describe('Codex quick start settings', () => {
       const withTunnel = loadUi({ ...ALL_OFF, cloudflared: true });
       withTunnel.app.applyWelcomeCliVisibility();
       expect(withTunnel.welcomeBtns.welcomeTunnelBtn.style.display).toBe('flex');
+
+      // Antigravity is a first-class welcome action, gated on `agy` like the rest.
+      const withAgy = loadUi({ ...ALL_OFF, antigravity: true });
+      withAgy.app.applyWelcomeCliVisibility();
+      expect(withAgy.welcomeBtns.welcomeAntigravityBtn.style.display).toBe('flex');
+      expect(withAgy.welcomeBtns.welcomeClaudeBtn.style.display).toBe('none');
     });
 
     it('gates every run mode in the dropdown, antigravity included, and never shell', () => {