test(custom-model): make the harness smoke test dynamic, verify all 9 CLIs end-to-end

Rewrites scripts/test-local-llm-harnesses.mjs -> .ts to read the live CLI
registry (enabledClis()) and call the real production
buildCustomModelInjection()/applyConfigDirInjection() instead of keeping a
second hand-maintained copy of every CLI's env/config shape. A future
registry change (new CLI, edited env var, fixed config template) is now
picked up automatically with zero edits to this script; only the one-shot
invocation flags (info the registry genuinely doesn't model) stay in a
small hand-maintained ONE_SHOT table, and a registry CLI with no entry
there reports UNKNOWN rather than being silently skipped.

Extracted src/custom-model-injection-apply.ts (applyConfigDirInjection/
removeConfigDir) so the production route and this script share one
implementation instead of two.

Full end-to-end run against a real llama-swap server, inside a
codeman/agent:llm-test Docker image with all 9 CLI binaries:

- claude, opencode, pi, grok, omp: PASS, real "hello world" replies
- codex: confirmed FAIL for a real protocol reason, not a bug — it only
  speaks the Responses API since Feb 2026, which llama.cpp/llama-swap
  don't implement
- gemini: confirmed FAIL, unresolved after real investigation — an
  undocumented GATEWAY AuthType gemini-cli selects once
  GOOGLE_GEMINI_BASE_URL is set rejects every auth-key format/override
  tried
- deepseek: reaches the server (env vars are read) but gets a consistent
  HTTP_404; root cause not identified, documented as best-effort/unknown
- antigravity: SKIP, no known mechanism (unchanged)

Two real bugs found and fixed along the way (grok, pi/omp registry
entries in stock.ts): grok's original recipe (env vars) was flat-out
wrong, not just unverified — the real mechanism is a config.toml
[model.<name>] block redirected via GROK_HOME. pi/omp's PI_CONFIG_DIR
does nothing for either (grepped pi's entire bundled source — the string
appears nowhere); the real redirect is the child process's own HOME, and
both need `models` as an array of {id} objects, not an object keyed by
id (silently loaded zero models otherwise).

deployment_plan.md, PR.md, docs/custom-model-endpoints.md, and CLAUDE.md
updated with the final confidence table reflecting all of the above.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017HqNWfmtBU2KN29SvSVWB3

CLAUDE.md

@@ -225,7 +225,7 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 **DeepSeek web UI** (`POST`/`GET`/`DELETE /api/deepseek/web`, `deepseek-web-server.ts`): the Run menu's "DeepSeek web UI..." entry supervises ONE background `dsh web` child process, deliberately **NOT a shell session**. The session version worked and was still wrong in use: it put a terminal tab on screen next to the web tab the user actually asked for, every single time, and nothing about a long-lived HTTP server needs to be a tab. ⚠️ What a session gave for free now has to be paid for explicitly, and every piece is load-bearing: **exactly one** server (a second click REUSES it rather than racing it for a port, which two sessions structurally could not do), **restarted when the browser authority changes** (`--trusted-host` fences dsh's `/api` against the browser authority, and a Codeman reachable at both loopback and a tailnet name has two, so whoever asks last wins: the asker is by definition the origin about to load the page), **killed on shutdown** (`stopDeepSeekWeb()` in the server teardown, because the child is detached so its whole plugin tree can be signalled at once, which also means it would OUTLIVE Codeman and hold its port against the next start), and **failures returned to the caller**, since with no tab there is nowhere for a stack trace to land. ⚠️ The port search starts at dsh's own default 3080 and walks 40, never fixed: that default is precisely the port most likely to be taken already by the user's own `dsh web`, and hardcoding it killed this feature with EADDRINUSE once. Free-port detection BINDS rather than connects (a connect probe cannot tell "free" from "listening but not answering yet"), so it is racy by nature and the caller still waits for the server to really answer before reporting success. ⚠️ Both `POST` and `DELETE` sit at the **same privilege bar as the profile installer** (`canUsernameRunPrivilegedCommands`) even though the action reads as "open a page": booting a dsh profile executes the plugin code in it, and the server is a single shared instance, so stopping it in multi-user mode takes it out from under other users' tabs.
 
-**Custom Model Endpoint Profiles** (opt-in, `customModelEndpointsEnabled`, SYNCED, default OFF; `docs/custom-model-endpoints.md`, design doc `deployment_plan.md`): points a session at a user-configured custom OpenAI-compatible endpoint — local (llama.cpp, DGX Spark, Strix Halo) or cloud (Azure AI Foundry, OpenRouter) — instead of its harness's native cloud backend. Endpoints are a read/write-array store (`custom-model-hosts.ts`, `~/.codeman/custom-model-hosts.json`) discovered via `GET <baseUrl>/v1/models`; `CustomModelHost.authStyle` (default `'both'`) sends BOTH `Authorization: Bearer` and `api-key` headers on discovery since cloud gateways (Azure) and local servers (llama.cpp) disagree on the convention and there is no way to know in advance which one a given endpoint wants. ⚠️ The actual per-CLI redirect is `capabilities.customModelInjection` on the CLI registry (four kinds: `env` for claude/gemini/grok/deepseek, `configContentEnv` reusing opencode's existing `OPENCODE_CONFIG_CONTENT`, `configDir` for codex/pi/omp — writes an isolated per-session config file, NEVER the user's real `~/.codex`/`~/.pi` config — and `unsupported` for antigravity, which has no known mechanism), computed by the pure `custom-model-injection.ts` (mirrors `session-cli-builder.ts`'s no-IO discipline). ⚠️ Applying a selection **restarts the session's CLI process in place** via `Session.restartCli()` — a de-restricted `reattachRemote()` reusing the same `respawn-pane -k` primitive local/remote respawns already share — because every one of these harnesses reads its endpoint config at process start, never per-turn, so there is no live hot-swap; `Session.setCustomModel()` undoes the PREVIOUS selection's env keys (and deletes its old `configDir`) before merging the new ones in, so switching endpoints or clearing back to native cloud never leaves a stale key behind. ⚠️ **Security**: every env var this feature can redirect (`ANTHROPIC_BASE_URL`, `GOOGLE_GEMINI_BASE_URL`, `CODEX_HOME`, `PI_CONFIG_DIR`, `OPENCODE_CONFIG_CONTENT`, etc.) is in that CLI's `privilegedEnvKeys` — several of these were reachable via the generic `envOverrides` field's prefix allowlist BEFORE this feature existed (the env allowlist is global and prefix-based, not per-CLI-scoped), so building this surfaced and closed a pre-existing gap rather than opening a new one. `ANTHROPIC_*` is deliberately NOT in claude's `allowedPrefixes` at all — Anthropic-traffic redirection can only happen through this feature's own admin-configured, SSRF-guarded route, never a plain client-supplied `envOverrides`. Confidence is per-CLI: claude/opencode/codex are hand-verified against a real llama.cpp/llama-swap server; gemini/pi/grok/deepseek/omp have their ONE-SHOT INVOCATION flags confirmed against real installed binaries' `--help` output, but their custom-endpoint env/config conventions remain unverified — see the confidence table in `deployment_plan.md`. The standalone `scripts/test-local-llm-harnesses.mjs` (reads a gitignored `scripts/local-llm-test.config.json`, template `.example.json` tracked) smoke-tests real CLI binaries against a real endpoint outside the web UI entirely, independent of tmux/sessions.
+**Custom Model Endpoint Profiles** (opt-in, `customModelEndpointsEnabled`, SYNCED, default OFF; `docs/custom-model-endpoints.md`, design doc `deployment_plan.md`): points a session at a user-configured custom OpenAI-compatible endpoint — local (llama.cpp, DGX Spark, Strix Halo) or cloud (Azure AI Foundry, OpenRouter) — instead of its harness's native cloud backend. Endpoints are a read/write-array store (`custom-model-hosts.ts`, `~/.codeman/custom-model-hosts.json`) discovered via `GET <baseUrl>/v1/models`; `CustomModelHost.authStyle` is `'bearer'` (default, `Authorization: Bearer`) or `'api-key'` (Azure's convention) — **never both**, live-tested against a real server: sending both headers on one request reliably hangs it indefinitely, reproduced 3×. ⚠️ The actual per-CLI redirect is `capabilities.customModelInjection` on the CLI registry (four kinds: `env` for claude/gemini/deepseek, `configContentEnv` reusing opencode's existing `OPENCODE_CONFIG_CONTENT`, `configDir` for codex/pi/grok/omp — writes an isolated per-session config file, NEVER the user's real `~/.codex`/`~/.pi`/`~/.omp`/grok config — and `unsupported` for antigravity, which has no known mechanism), computed by the pure `custom-model-injection.ts` (mirrors `session-cli-builder.ts`'s no-IO discipline). ⚠️ `PI_CONFIG_DIR` does NOTHING for pi or omp (grepped pi's entire bundled JS source — the string appears nowhere); both hardcode `~/.pi/agent/models.json` / `~/.omp/agent/models.yml` with no dedicated override, so the real redirect for both is the child process's own **`HOME`**, and both need `models` as an ARRAY of `{id}` objects (an object keyed by id silently loads zero models). Grok's real mechanism turned out to be a `config.toml` `[model.<name>]` block redirected via `GROK_HOME` — its original env-var-based recipe was flat-out wrong (produced "Not signed in" against a real binary), not just unverified. ⚠️ Applying a selection **restarts the session's CLI process in place** via `Session.restartCli()` — a de-restricted `reattachRemote()` reusing the same `respawn-pane -k` primitive local/remote respawns already share — because every one of these harnesses reads its endpoint config at process start, never per-turn, so there is no live hot-swap; `Session.setCustomModel()` undoes the PREVIOUS selection's env keys (and deletes its old `configDir`) before merging the new ones in, so switching endpoints or clearing back to native cloud never leaves a stale key behind. ⚠️ **Security**: every env var this feature can redirect (`ANTHROPIC_BASE_URL`, `GOOGLE_GEMINI_BASE_URL`, `CODEX_HOME`, `GROK_HOME`, `HOME` for pi/omp, `OPENCODE_CONFIG_CONTENT`, etc.) is in that CLI's `privilegedEnvKeys` — several of these were reachable via the generic `envOverrides` field's prefix allowlist BEFORE this feature existed (the env allowlist is global and prefix-based, not per-CLI-scoped), so building this surfaced and closed a pre-existing gap rather than opening a new one. `ANTHROPIC_*` is deliberately NOT in claude's `allowedPrefixes` at all — Anthropic-traffic redirection can only happen through this feature's own admin-configured, SSRF-guarded route, never a plain client-supplied `envOverrides`. **Confidence, verified end-to-end against a real llama-swap server via the DYNAMIC `scripts/test-local-llm-harnesses.ts`** (reads the live CLI registry, so a registry change needs zero script edits): claude/opencode/pi/grok/omp **PASS**; codex config structure is correct but codex only speaks the Responses API since Feb 2026, which llama.cpp/llama-swap don't implement — a confirmed protocol gap, not a bug; gemini fails with `Invalid auth method selected` (an undocumented `GATEWAY` AuthType gemini-cli selects once `GOOGLE_GEMINI_BASE_URL` is set — unresolved after real investigation); deepseek reaches the server but gets a consistent `HTTP_404` (root cause not identified); antigravity has no known mechanism at all. See the confidence table in `deployment_plan.md` for the full detail on each.
 
 **Run launch synchronization**: the Run entrypoint holds an in-flight lock and disables `#runBtn` for the whole launch (≥500ms), so a double click cannot create duplicate sessions with the same `w<n>-<case>` name. `_ensureCreatedSessionVisible()` runs before `selectSession()`, and `_onSessionCreated()` stays an idempotent upsert, so POST-first and SSE-first ordering both produce exactly one rendered tab. ⚠️ **Closing has the mirror-image race and one owner**: `closeSession()` reads `wasActive` BEFORE its `await` and announces the delete via `_closingSessions`, while `_onSessionDeleted` skips the active-session handoff for an id in that set. Both used to read `activeSessionId` after the fact, so the `session_deleted` broadcast for your own delete could null it first and closing the tab you were on landed on the welcome screen instead of the next session, on the same build, depending on timing. The fallback also picks the first order entry that is still in `sessions` (a dead id can linger in `sessionOrder`, same reason Alt+N indexes a live-filtered list). A delete from ANOTHER client still shows the welcome screen, which is the honest answer when what you were looking at was taken away. Tests: `test/session-close-fallback.test.ts`. → [architecture-invariants#run-launch-synchronization](docs/architecture-invariants.md#run-launch-synchronization)