diff --git a/.changeset/run-menu-custom-model-picker.md b/.changeset/run-menu-custom-model-picker.md index d20fae7f..4a8b8b66 100644 --- a/.changeset/run-menu-custom-model-picker.md +++ b/.changeset/run-menu-custom-model-picker.md @@ -2,27 +2,34 @@ 'aicodeman': minor --- -**Custom model endpoints: Run-menu picker, and hardening from real llama-swap validation** (#430, follow-up to #393's HTTP-API-only cut). With **Custom model endpoints** on (App Settings → Models) and at least one saved endpoint carrying a discovered model, the Run dropdown grows a **Custom Endpoints** section generated live off the CLI registry's own `capabilities.customModelInjection` — one entry per (harness that can redirect to a custom endpoint, saved endpoint). Picking one launches that harness and applies the endpoint to it; with two or more discovered models a small, scrollable dialog asks which one first, the endpoint's `defaultModelId` marked but never auto-chosen. Endpoints also now re-discover themselves automatically every 5 minutes in the background, one unreachable endpoint never blocking the others. +feat(custom-model): pick a custom endpoint straight from the Run menu -Everything below was found and fixed against a **real llama-swap server**, not just unit tests: +#393 landed the backend for custom model endpoints and left it reachable only over the +HTTP API. This is the rest of it. Turn on Custom model endpoints in App Settings, save +an endpoint, and the Run dropdown grows a Custom Endpoints section built live off the +CLI registry, one entry per harness that can actually redirect plus each endpoint you +saved. Pick one and it launches that harness pointed at your server, asking which model +first when the endpoint has more than one. Endpoints re-discover themselves every five +minutes, and one unreachable endpoint never blocks the others. App Settings gains full +add, edit and delete for endpoints. -- **Session-busy false refusal.** A freshly launched CLI reports itself `busy` for its own startup (spinner, workspace-trust check) well before the apply call would reach it, and the apply route correctly refuses to restart a session mid-turn — indistinguishable from a fresh boot. The picker now waits for the new session to go idle (bounded at 20s, never an error on timeout) before applying. -- **Errors and confirmations you can actually read.** A failed apply's real server-side reason (not a generic message) reaches the toast, and that specific message stays on screen with a close button instead of vanishing on the usual 3s timer. -- **"Both claude.ai and ANTHROPIC_API_KEY set" warning.** A custom-model Claude session now runs with an isolated `CLAUDE_CONFIG_DIR` (empty, no real credentials in it) so the injected API key never coexists with a stored OAuth login — `projects` is symlinked back to the real config dir so the response viewer/subagent windows/Read My Mind keep working. That isolated, otherwise-empty directory has none of a real profile's prior "Detected a custom API key — use it?" approvals either, which would otherwise re-ask on _every_ launch with nobody at a TTY to answer (and silently refuse the key on its own default); the apply step now pre-seeds that exact approval field the same way answering the prompt once by hand would. -- **Context-window overflow.** Claude Code assumes a large default context window for a model id it doesn't recognize and never compacts, so a real local model's much smaller context silently overflowed (confirmed live: a stock ~33.7K-token system prompt against a 16384-token model). Discovery now also learns each model's real context length and applies it as `CLAUDE_CODE_MAX_CONTEXT_TOKENS` — sourced primarily from llama-swap's own `GET /running`, whose `cmd` field carries the launch flags (`--fit-ctx`/`-c`/`--ctx-size`) actually in effect, since `GET /props`'s `n_ctx` was confirmed live to report the model's theoretical/trained maximum rather than the real `--fit-ctx`-shrunk runtime context (a 154112-vs-16384 discrepancy, caught only because the fixed value still overflowed) — `/props` is now a fallback for a plain llama.cpp server with no `/running` at all. -- **Context floor too small for Claude Code to even start.** Fixing the overflow above surfaced a second, unfixable-by-injection failure: Claude Code's own system prompt and tool schemas cost roughly 36.4K tokens on their own (confirmed live via an `in:0 out:0` failure on the very first message), which can exceed a small model's entire real context before any conversation history exists to trim — no `CLAUDE_CODE_MAX_CONTEXT_TOKENS` value fixes that, since it only governs when history gets compacted. Applying such a model now returns a warning (gated on the CLI registry declaring a `contextLengthVar`, so it's a no-op for every other harness) instead of launching straight into a guaranteed first-message failure, and the Run-menu picker shows it as an in-app dialog naming the model, its discovered context and the ~40K safe floor, with the actual fix spelled out: give the model an explicit larger `-c`/`--ctx-size` in llama-swap's config instead of relying on auto-fit, which optimizes for the biggest model that fits rather than the biggest context. "Launch anyway" is still one click away. -- **The real root cause of "it still says opus, not my model."** llama.cpp runs exactly one model at a time; llama-swap unloads and reloads it on demand, which can take anywhere from a few seconds to well over a minute — long enough that a session mid-swap is indistinguishable from one that never left the native backend. Applying a selection now checks llama-swap's own `GET /running` first (feature-detected; a plain llama.cpp/OpenAI-compatible server has no such endpoint and is never checked); if switching would unload a model **another live session is actively using**, the apply is refused with a warning naming that session instead of silently switching, and a confirmation retry proceeds anyway. Either way, a sticky "loading model…" toast now covers the actual swap window until llama-swap reports the target model ready, so a prompt sent mid-swap reads as "loading," never as silence or an answer from whatever was loaded a moment before. +Seven of the harnesses (opencode, Codex, Gemini, Pi, Grok, DeepSeek and OMP) now launch +directly onto the endpoint with no restart at all, where before you watched a native +boot followed immediately by a second one. Claude still launches and then restarts in +place, which its own resume makes far less jarring. -- **Claude's whole first-run sequence, on every single launch.** A fresh, otherwise-empty `CLAUDE_CONFIG_DIR` isn't just missing the API-key approval above — Claude Code treats it as a brand-new profile and replays the theme picker, the security-notes screen, the per-project "trust this folder?" dialog, and (running bypassed) a one-time permissions-bypass warning, every time, confirmed live. None of that shows up again for a real, already-onboarded profile. `customModelInjection`'s new `skipFirstRunPrompts` (claude's entry only) pre-seeds that same "already been through this" state — `hasCompletedOnboarding` and this session's own project trust into the same `.claude.json` the API-key approval merges into, `skipDangerousModePermissionPrompt` into `settings.json` — so a custom-model launch reaches the conversation exactly as fast as a native cloud one, with nobody there to click through a wizard. +Most of this release's work went into things that only show up against a real server, +and each was found that way rather than in tests: a freshly launched CLI reporting +itself busy for its own startup and getting refused; Claude Code assuming a large +context window for a model it does not recognise and silently overflowing a small one; +a model whose real context is below what Claude Code's own system prompt costs, which +no setting can fix and which now warns before launching into a certain failure; and the +big one, llama.cpp running exactly one model at a time, so applying a selection can +unload the model another session is using. That last case now asks first, tells you +which session it affects, and keeps a "loading model" notice on screen for the whole +swap window, so a prompt sent mid-swap reads as loading rather than as an answer from +whatever was loaded a moment ago. A background sweep also catches the reverse: your +session's model being evicted later by somebody else's ordinary use. -Two more, from actually clicking through the swap-confirm and context-warning dialogs live: their z-index sat under the centred status banner, so a dialog could render fully hidden behind "Claude started — switching to llama-swap…"; and their Cancel/confirm buttons stacked instead of sitting side by side (`.btn-toolbar`'s own `display: flex` needs a row-layout parent it never had). Both dialogs now clear the banner and lay their buttons out centred, side by side. - -- **A session's model getting silently swapped out later, not just at launch.** The conflict check above only ever runs at the moment a session is created or a model applied — confirmed live: a second Codex session picking a different model launched with no warning at all, because nothing conflicted at that exact instant, yet it silently evicted the first session's model regardless (llama.cpp runs one model at a time). There was no mechanism to catch a swap caused by a DIFFERENT session's own later, ordinary use. A new periodic sweep (`detectCustomModelSwapDisplacements`, every 20s, one `GET /running` per distinct endpoint with a live custom-model session) now compares each such session's own model against what's actually loaded, and a new `custom-model:swapped-out` SSE event drives a global toast naming the displaced session and what's now loaded instead — so you find out before typing into a session that's about to trigger yet another reload. Notifies once per displacement, clearing once a session's own model is loaded and ready again so a later, genuinely new displacement notifies again. - -- **The loading banner's second line is now the real backend log line, not just a countdown.** llama-swap's `GET /api/events` SSE stream carries the actual `llama-server` process's own stdout (`load_model: loading model ''`, `llama_server: model loaded`, tokenizer warnings, all of it) tagged `source: "upstream"`, distinct from llama-swap's own `source: "proxy"` request-access lines — confirmed live end-to-end through a real forced swap, and it correctly stays on the last thing llama.cpp said once the load goes quiet rather than clearing to blank. ⚠️ This feature's own first cut targeted `GET /logs` instead (the name that suggested it) and shipped a live-tested implementation against it before this live check caught that `/logs` carries ONLY the proxy request log and never once showed a single backend line, even seconds after a real, confirmed swap — corrected before merge, not after. - -Remote (SSH) and Docker sessions are refused for now (400) — their restart reattaches the durable remote/in-container tmux rather than relaunching the agent. - -- **The loading banner's countdown is gone, replaced by a generic disclaimer and a Cancel button.** Its size-scaled expected-time estimate and matching auto-timeout were both a guess dressed up as a fact — real load time depends on hardware this feature has no way to know, and a fixed number could kill a genuinely slow load partway through. The banner now says "this can take a while depending on your hardware and the model size", polls indefinitely, and carries a **Cancel** button that ends the wait and closes the session on the user's own call rather than a guessed deadline. - -**One more, from watching it launch live: opencode, Codex, Gemini, Pi, Grok, DeepSeek and OMP now launch directly on the endpoint, with no restart at all.** Picking one of these seven from the Run-menu picker used to launch natively first, wait for it to settle, then restart it in place with the endpoint applied — a deliberate two-step design, but visibly a native boot immediately followed by a second one, worst on a CLI whose TUI fully reinitializes on a restart (confirmed live on Codex). `POST /api/quick-start` now accepts a `customModel` field and computes the same injection _before_ the session exists, launching straight onto the endpoint the first time — no visible relaunch, and it also runs the same llama-swap conflict check (warns before unloading a model another live session is using) at create time. Claude still uses the original launch-then-restart path for now (its own `--resume`-based restart is far less jarring, and `runClaude()`'s multi-tab and docker-config-drift-retry logic make folding it into the one-shot path separate work). +Remote SSH and Docker sessions are refused for now, since their restart reattaches a +durable tmux rather than relaunching the agent. diff --git a/CLAUDE.md b/CLAUDE.md index 88a7d4b3..75470f58 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -229,7 +229,7 @@ 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 `docs/custom-model-endpoints-plan.md`; full stack — settings-panel CRUD + the Run-menu picker, on top of the backend below): 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 /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.]` 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. ⚠️ **Two launch paths, chosen by mechanism, not preference — see the second paragraph below for why**: opencode/codex/gemini/pi/grok/deepseek/omp apply the selection ONE-SHOT, before the session/process ever exists, with no restart at all; claude alone still applies a selection by **restarting 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. ⚠️ Deleting a key from `_envOverrides` is NOT enough on its own: `tmux setenv` persists at the tmux-session level and is inherited by `respawn-pane` (measured: `setenv FOO bar` survived two successive `respawn-pane -k`), so the retired keys are queued (`_pendingEnvUnsets`) and ride `RespawnPaneOptions.unsetEnvKeys` into `applyEnvOverrides()`, which `setenv -u`s them BEFORE re-applying the live overrides. ⚠️ `restartCli()` kills a WORKING pane, so a CLI whose launch declares a `fallback` chain (claude) gets the live conversation id pinned as `resumeSessionId` for that one respawn: `--session-id ` refuses an id that already has a transcript (`Session ID ... is already in use`), and without the `--resume || --session-id ` shape the docker/remote pane commands already use, applying a model killed the pane and lost the session. ⚠️ pi, omp and grok need the config file AND a `model` launch param (`custom/` for pi/omp, grok's `[model.codeman-custom]` block name): that is the registry's `customModelInjection.launchModel` template, applied onto the respawn options through `legacyConfigField` by `_withCustomModelLaunchModel()`, never by id, and a model id the CLI's `model` token pattern cannot carry is refused with a 400 rather than silently dropped by the argv engine. ⚠️ Remote (SSH) and Docker sessions are REFUSED (400): their `restartCli()` reattaches a durable tmux rather than restarting the agent and the env lands on the local pane, so they used to report `restarted:true` and change nothing. The selection survives a Codeman restart as the disk-only `__customModel` (bookkeeping: env KEYS, config dir, launch model; never the values, which carry the API key and are re-derived from the endpoint store on recovery), the config dir is removed with the session, and every secret-bearing file (`custom-model-hosts.json`, the per-session config dir) is written 0600. ⚠️ **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, and codex only speaks the Responses API since Feb 2026 (`wire_api = "responses"`) — re-verified live against a llama-swap deployment that DOES answer `/v1/responses` (an earlier test's harder failure against a different deployment does not reproduce everywhere): a plain, no-tool-call chat turn gets a real reply, but a real tool-call attempt came back as `agent_message` TEXT (the tool-call JSON printed as the answer) rather than an executable `function_call` item — confirmed via `codex exec --json`'s raw event stream. Tool execution is what makes codex a coding agent, so it remains not usable for real work either way, just with a more precise failure mode than a flat protocol break; 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's originally-reported `HTTP_404` is root-caused and fixed — its bundled `@deepseek-ai/dsh-llm-deepseek` module builds `${DEEPSEEK_BASE_URL}/chat/completions` with no `/v1` of its own (confirmed by installing the real package and reading its source), so a new `appendV1Suffix` flag on its registry entry (alone — claude/gemini must not get it) runs `endpoint.baseUrl` through `withV1Suffix()` before writing it, live-confirmed against llama-swap (`.../chat/completions` 404s, `.../v1/chat/completions` succeeds) though not yet re-run through an actual `dsh` binary, which isn't installable in this environment; antigravity has no known mechanism at all. See the confidence table in `docs/custom-model-endpoints-plan.md` for the full detail on each. ⚠️ **The Run-menu picker generates entries from `window.__codemanCustomModelClis`** (`server.ts`, injected at page render from `enabledClis().filter(kind==='agent' && customModelInjection.kind!=='unsupported')`, JSON-escaped against a literal `` via the exported `escapeScriptJson()` since `label` is a user-`clis.json`-settable string unlike the neighbouring booleans-only `__codemanCliAvailable`), never a hardcoded per-CLI id list in the frontend — the same "no branching on CLI id outside stock.ts" discipline the registry itself enforces. One entry per (capable, INSTALLED CLI, saved endpoint) pair, e.g. "Claude Code (llama.cpp)", filtered through `isCliAvailable()` like the stock entries. Clicking one calls `selectCustomModelEntry(mode, endpointId)` (`session-ui.js`), which re-fetches the endpoint (never trusts anything cached from the dropdown's render — the 5-minute sweep below or a settings edit may have changed it since) and decides the model: exactly one discovered model launches straight away, two or more open `#customModelPickModal` to ask, with `defaultModelId` marked but never auto-chosen (asking exists so ONE launch can deliberately differ from the saved default). Either way the actual launch (`runCustomModelEntry`) routes through `run()` itself via a temporary `_runMode` swap — never `setRunMode()`, which would persist it as the user's new default — rather than a parallel dispatch table, which is what gives a custom-model launch the same `_runInFlight` lock every other Run click gets and means a CLI whose injection recipe lands later needs no update here. It then GETs `/api/sessions/:id/wait?until=idle&timeout=20000` on that session BEFORE applying — measured live, a freshly launched CLI reports itself `busy` for its own startup (boot spinner, workspace-trust check) well before the apply call would otherwise reach it, and the apply route's `isBusy()` guard correctly can't tell that apart from a real turn in progress, so every fresh launch failed with `SESSION_BUSY` until this wait was added. A timeout there is a normal 200 per the wait endpoint's own contract, never an error, so a session still busy after 20s just reaches the apply call anyway and gets that route's own honest error. It then calls `POST /api/sessions/:id/custom-model` on the session `run()` produced, guarded by snapshotting `activeSessionId` before the call and requiring it to have actually changed after — every `run*()` handles its own failure internally and returns normally rather than throwing, so a declined/failed launch must not silently re-point and restart whatever session was already open. ⚠️ The apply call reads the response body itself (`_api()`) rather than `_apiJson()`, which unwraps success but silently discards a failure body — losing the one thing (`error`) that distinguishes "still busy", "not a discovered model", "remote/Docker session" and everything else the route can report; the resulting toast is `type: 'error'` with an explicit `duration: 0` (no auto-dismiss, an explicit close button) at that one call site — not a blanket sticky-error default, which stacked unbounded on `.toast-container` with no cap or eviction — precisely so a message worth diagnosing survives long enough to be read instead of vanishing on the usual 3s timer. Entries are hidden for a remote/docker active case (the apply route refuses both) and for an endpoint with no discovered models at all (nothing to launch with). ⚠️ **Every saved endpoint's models also re-discover themselves automatically**, a `this.cleanup.setInterval` in `server.ts` (`CUSTOM_MODEL_REDISCOVER_INTERVAL_MS`, 5 minutes, off under `testMode` like the Codex plan-usage poll beside it) calling the exported `refreshAllCustomModelHosts()` (`custom-model-routes.ts`) — one endpoint unreachable on a cycle never blocks the others, and a read-modify-write PER HOST (re-reading the store before each splice, keyed by id) means an admin's concurrent edit or delete wins over a sweep that started before it, never the reverse. +**Custom Model Endpoint Profiles** (opt-in, `customModelEndpointsEnabled`, SYNCED, default OFF; `docs/custom-model-endpoints.md`, design doc `docs/custom-model-endpoints-plan.md`; full stack — settings-panel CRUD + the Run-menu picker, on top of the backend below): 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 /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.]` 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. ⚠️ **Two launch paths, chosen by mechanism, not preference — see the second paragraph below for why**: opencode/codex/gemini/pi/grok/deepseek/omp apply the selection ONE-SHOT, before the session/process ever exists, with no restart at all; claude alone still applies a selection by **restarting 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. ⚠️ Deleting a key from `_envOverrides` is NOT enough on its own: `tmux setenv` persists at the tmux-session level and is inherited by `respawn-pane` (measured: `setenv FOO bar` survived two successive `respawn-pane -k`), so the retired keys are queued (`_pendingEnvUnsets`) and ride `RespawnPaneOptions.unsetEnvKeys` into `applyEnvOverrides()`, which `setenv -u`s them BEFORE re-applying the live overrides. ⚠️ `restartCli()` kills a WORKING pane, so a CLI whose launch declares a `fallback` chain (claude) gets the live conversation id pinned as `resumeSessionId` for that one respawn: `--session-id ` refuses an id that already has a transcript (`Session ID ... is already in use`), and without the `--resume || --session-id ` shape the docker/remote pane commands already use, applying a model killed the pane and lost the session. ⚠️ pi, omp and grok need the config file AND a `model` launch param (`custom/` for pi/omp, grok's `[model.codeman-custom]` block name): that is the registry's `customModelInjection.launchModel` template, applied onto the respawn options through `legacyConfigField` by `_withCustomModelLaunchModel()`, never by id, and a model id the CLI's `model` token pattern cannot carry is refused with a 400 rather than silently dropped by the argv engine. ⚠️ Remote (SSH) and Docker sessions are REFUSED (400): their `restartCli()` reattaches a durable tmux rather than restarting the agent and the env lands on the local pane, so they used to report `restarted:true` and change nothing. The selection survives a Codeman restart as the disk-only `__customModel` (bookkeeping: env KEYS, config dir, launch model; never the values, which carry the API key and are re-derived from the endpoint store on recovery), the config dir is removed with the session, and every secret-bearing file (`custom-model-hosts.json`, the per-session config dir) is written 0600. ⚠️ **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, and codex only speaks the Responses API since Feb 2026 (`wire_api = "responses"`) — re-verified live against a llama-swap deployment that DOES answer `/v1/responses` (an earlier test's harder failure against a different deployment does not reproduce everywhere): a plain, no-tool-call chat turn gets a real reply, but a real tool-call attempt came back as `agent_message` TEXT (the tool-call JSON printed as the answer) rather than an executable `function_call` item — confirmed via `codex exec --json`'s raw event stream. Tool execution is what makes codex a coding agent, so it remains not usable for real work either way, just with a more precise failure mode than a flat protocol break; 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's originally-reported `HTTP_404` is root-caused and fixed — its bundled `@deepseek-ai/dsh-llm-deepseek` module builds `${DEEPSEEK_BASE_URL}/chat/completions` with no `/v1` of its own (confirmed by installing the real package and reading its source), so a new `appendV1Suffix` flag on its registry entry (alone — claude/gemini must not get it) runs `endpoint.baseUrl` through `withV1Suffix()` before writing it, live-confirmed against llama-swap (`.../chat/completions` 404s, `.../v1/chat/completions` succeeds) though not yet re-run through an actual `dsh` binary, which isn't installable in this environment; antigravity has no known mechanism at all. See the confidence table in `docs/custom-model-endpoints-plan.md` for the full detail on each. ⚠️ **The Run-menu picker generates entries from `window.__codemanCustomModelClis`** (`server.ts`, injected at page render from `enabledClis().filter(kind==='agent' && customModelInjection.kind!=='unsupported')`, JSON-escaped against a literal `` via the exported `escapeScriptJson()` since `label` is a user-`clis.json`-settable string unlike the neighbouring booleans-only `__codemanCliAvailable`), never a hardcoded per-CLI id list in the frontend — the same "no branching on CLI id outside stock.ts" discipline the registry itself enforces. One entry per (capable, INSTALLED CLI, saved endpoint) pair, e.g. "Claude Code (llama.cpp)", filtered through `isCliAvailable()` like the stock entries. Clicking one calls `selectCustomModelEntry(mode, endpointId)` (`session-ui.js`), which re-fetches the endpoint (never trusts anything cached from the dropdown's render — the 5-minute sweep below or a settings edit may have changed it since) and decides the model: exactly one discovered model launches straight away, two or more open `#customModelPickModal` to ask, with `defaultModelId` marked but never auto-chosen (asking exists so ONE launch can deliberately differ from the saved default). Either way the actual launch (`runCustomModelEntry`) routes through `run()` itself via a temporary `_runMode` swap — never `setRunMode()`, which would persist it as the user's new default — rather than a parallel dispatch table, which is what gives a custom-model launch the same `_runInFlight` lock every other Run click gets and means a CLI whose injection recipe lands later needs no update here. It then GETs `/api/sessions/:id/wait?until=idle&timeout=20000` on that session BEFORE applying — measured live, a freshly launched CLI reports itself `busy` for its own startup (boot spinner, workspace-trust check) well before the apply call would otherwise reach it, and the apply route's `isBusy()` guard correctly can't tell that apart from a real turn in progress, so every fresh launch failed with `SESSION_BUSY` until this wait was added. A timeout there is a normal 200 per the wait endpoint's own contract, never an error, so a session still busy after 20s just reaches the apply call anyway and gets that route's own honest error. It then calls `POST /api/sessions/:id/custom-model` on the session `run()` produced, guarded by snapshotting `activeSessionId` before the call and requiring it to have actually changed after — every `run*()` handles its own failure internally and returns normally rather than throwing, so a declined/failed launch must not silently re-point and restart whatever session was already open. ⚠️ The apply call reads the response body itself (`_api()`) rather than `_apiJson()`, which unwraps success but silently discards a failure body — losing the one thing (`error`) that distinguishes "still busy", "remote/Docker session" and everything else the route can report (⚠️ neither route validates `modelId` against the endpoint's discovered list, deliberately: discovery can be up to 5 minutes stale, so a 400 there would refuse a launch that works — a typo'd id fails on the CLI's own first request instead); the resulting toast is `type: 'error'` with an explicit `duration: 0` (no auto-dismiss, an explicit close button) at that one call site — not a blanket sticky-error default, which stacked unbounded on `.toast-container` with no cap or eviction — precisely so a message worth diagnosing survives long enough to be read instead of vanishing on the usual 3s timer. Entries are hidden for a remote/docker active case (the apply route refuses both) and for an endpoint with no discovered models at all (nothing to launch with). ⚠️ **Every saved endpoint's models also re-discover themselves automatically**, a `this.cleanup.setInterval` in `server.ts` (`CUSTOM_MODEL_REDISCOVER_INTERVAL_MS`, 5 minutes, off under `testMode` like the Codex plan-usage poll beside it) calling the exported `refreshAllCustomModelHosts()` (`custom-model-routes.ts`) — one endpoint unreachable on a cycle never blocks the others, and a read-modify-write PER HOST (re-reading the store before each splice, keyed by id) means an admin's concurrent edit or delete wins over a sweep that started before it, never the reverse. **Everything below landed after the initial backend + picker cut, each confirmed live against a real llama-swap deployment.** ⚠️ **llama-swap runs one model at a time, and switching can disrupt ANOTHER live session** — before applying, both apply routes call llama-swap's own `GET /running` (feature-detected via `getLlamaSwapStatus()`, `custom-model-routes.ts`; a plain llama.cpp/OpenAI-compatible server has no such endpoint and is simply never checked). If a different model is loaded and ready AND another live session's own selection is using it, the apply returns `{requiresConfirmation, currentlyLoadedModel, affectedSessions}` instead of switching silently; retrying with `confirmed: true` skips the check, and switching with nothing else affected proceeds immediately. llama-swap also has no dedicated "switch model" endpoint — the only thing that actually starts a swap is a real inference request naming the model (confirmed live: applying a selection alone never reached llama-swap's own logs, since nothing had asked it to load anything) — so both routes also fire `triggerLlamaSwapLoad()`, a fire-and-forget `POST /v1/chat/completions` with `max_tokens: 1`, whenever the target model isn't already loaded and ready. ⚠️ **That launch-time check cannot catch a swap caused by a DIFFERENT session's LATER, ordinary use** — confirmed live: a second Codex session picking a different model launched with no warning at all (nothing conflicted at that exact instant), yet it silently evicted the first session's model regardless, since llama-swap has no push notification of its own. `detectCustomModelSwapDisplacements()` (`custom-model-routes.ts`) is a separate periodic sweep (`server.ts`, `CUSTOM_MODEL_SWAP_CHECK_INTERVAL_MS` = 20s) that compares each live custom-model session's own `modelId` against what `/running` actually reports loaded, broadcasting a `custom-model:swapped-out` SSE event — shown as a global toast, never tied to the displaced session's own tab, since the whole point is telling the user before they type into it — the first time a mismatch appears, via a caller-owned de-dupe `Set` cleared once that session's own model is loaded and ready again so a later, genuinely new displacement notifies again rather than staying silently un-notified forever after the first one. ⚠️ **Context length is read from the REAL launch command, never `/props`** — `/props?model=`'s `default_generation_settings.n_ctx` was confirmed live to report a `--fit-ctx`-launched backend's theoretical/trained maximum rather than the real runtime-configured size (a measured 154112-vs-16384 discrepancy, caught only because the unfixed value still overflowed), so discovery parses the actual configured size straight out of `/running`'s own `cmd` field instead (`parseCtxFromCmd`: `--fit-ctx ` first, then plain llama.cpp `-c`/`--ctx-size`), falling back to `/props` only when `cmd` states no recognizable flag at all. ⚠️ **Claude alone gets a context-window FLOOR check, on top of the ceiling `contextLengthVar` already fixes** — `exceedsSafeContextFloor()` (gated on the registry declaring `contextLengthVar`, so a no-op for every other CLI by construction) compares a model's discovered context against `CLAUDE_MIN_SAFE_CONTEXT_TOKENS` (40000): confirmed live, twice, that Claude Code's own system prompt and tool schemas cost roughly 36.4K tokens on the very first message, before any conversation history exists to compact, so a smaller real context fails outright regardless of what `CLAUDE_CODE_MAX_CONTEXT_TOKENS` says (that var only controls when HISTORY gets compacted, and there is none yet on message one). Below the floor, the apply returns `{requiresContextWarning, modelId, contextLength, minSafeContextTokens}` instead of launching, shown as an in-app dialog naming the actual fix: give the model an explicit larger `-c`/`--ctx-size` in llama-swap's config instead of relying on `--fit-ctx` auto-fit, which optimizes for the biggest MODEL that fits rather than the biggest CONTEXT. ⚠️ **A fresh, isolated `CLAUDE_CONFIG_DIR` looks like a brand-new Claude Code profile and replays its ENTIRE first-run sequence on every launch** — the theme picker, the security-notes screen, the per-project "trust this folder?" dialog, and (running with a bypass-permissions flag) a one-time warning about it, confirmed live, none of which a real, already-onboarded profile shows again. `skipFirstRunPrompts` (claude's entry only, requires `apiKeyTrustFile` since it reuses the same file) pre-seeds that same "already been through this" state: `hasCompletedOnboarding` and this session's own `projects[workingDir].hasTrustDialogAccepted` merge into the same `.claude.json` the API-key trust file already writes to, and `skipDangerousModePermissionPrompt` merges into `settings.json` (a different file, same corrupt-tolerant merge). ⚠️ **The loading banner shows the REAL backend log line, not a guess, and has no countdown or auto-timeout at all.** `getLatestLlamaSwapLogLine()` holds one `GET /api/events` SSE connection open per endpoint (confirmed live to stay open indefinitely — read past 220KB over 8s with no `done`; idle-closed after 30s via `pruneIdleLlamaSwapLogTails`, same 20s sweep as the swap-displacement check above), parsing `logData` frames and keeping only `source: "upstream"` (the real `llama-server` process's own stdout) lines, never `source: "proxy"` (llama-swap's own request-access log). ⚠️ `GET /logs` — the endpoint this feature's own first cut targeted, since the name suggested it — was confirmed live to carry ONLY the proxy log and never a single backend line, even seconds after a real, verified model swap; caught and corrected by a live check before merge, not after. The banner itself dropped its size-scaled expected-time estimate and matching auto-timeout (a guess dressed up as a fact that could kill a genuinely slow load partway through on slower hardware) for a generic hardware/model-size disclaimer plus a user-driven **Cancel** button (`_showCenterStatus`'s `onCancel` option, a real button distinct from the plain "×" close glyph an `'error'`-type banner gets) that ends the wait and closes the session on the user's own call rather than a guessed deadline. diff --git a/docs/api-reference.md b/docs/api-reference.md index 95193fc3..7aab6409 100644 --- a/docs/api-reference.md +++ b/docs/api-reference.md @@ -625,7 +625,7 @@ authStyle?, defaultModelId? }` creates one. `id` must match `server.ts`), so there is no route for triggering "refresh all" — one endpoint being unreachable on a cycle never blocks the others. - `GET /api/v1/model-endpoints/:id/running-status` -> `{ isLlamaSwap, -running: [{model, state, cmd?}], logLine? }`, read-only, no admin gate +running: [{model, state}], logLine? }`, read-only, no admin gate (any session owner who could already point a session at this endpoint can equally ask what it currently has loaded). `isLlamaSwap` is feature-detected via the endpoint's own `GET /running` — a plain diff --git a/src/web/public/session-ui.js b/src/web/public/session-ui.js index 07d3e1eb..3047f90c 100644 --- a/src/web/public/session-ui.js +++ b/src/web/public/session-ui.js @@ -920,8 +920,8 @@ Object.assign(CodemanApp.prototype, { // _apiJson() (used everywhere else in this file) unwraps a success body to // its `data`, but on failure it swallows the response entirely and returns // null — exactly the `error` text a caller needs to tell "the endpoint is - // unreachable" apart from "the CLI can't be redirected", "not one of the - // discovered models", or "this is a Docker/remote session". Go through the + // unreachable" apart from "the CLI can't be redirected" or "this is a + // Docker/remote session". Go through the // raw response here instead so a failure is diagnosable, not just present. let { ok, data, res } = await this._applyCustomModelToSession(sessionId, endpointId, modelId);