Skip to content

feat(models): discover the harness's own model list instead of hardcoding it - #213

Open
davidliuk wants to merge 1 commit into
mainfrom
feat/harness-model-discovery
Open

feat(models): discover the harness's own model list instead of hardcoding it#213
davidliuk wants to merge 1 commit into
mainfrom
feat/harness-model-discovery

Conversation

@davidliuk

Copy link
Copy Markdown
Collaborator

希望可以支持自动识别目前选的 harness 里面可选的模型(比如 codex、claude code 等,他们的模型列表可能会更新新模型,我们希望可以自动用他的列表,不需要我们反复去适配)

The drift is already real

built-in list what the tool actually serves
Codex (codex-cli 0.145.0) gpt-5.6, gpt-5.3-codex, gpt-5.2-codex, gpt-5.2, gpt-5.1-codex-max, o3, o4-mini gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark
OpenRouter ~60 pinned 342 live, including anthropic/claude-opus-5

Note gpt-5.6 vs gpt-5.6-sol — the pinned default is a model the CLI no longer serves.

How it discovers

Provider Source
Codex codex app-server line-delimited JSON-RPC → initialize, then model/list with nextCursor pagination. This is the same catalogue the Codex CLI's own picker reads. Honours CODEX_CLI_PATH.
OpenRouter GET https://openrouter.ai/api/v1/models (public, no key)
Claude / Cursor / Gemini / Nano Built-in list — these CLIs expose no model-listing command today
Local GPU Ollama /api/tags, unchanged

Adding a harness later means adding one function to the DISCOVERERS map.

It can never make things worse

This repo just spent a PR fixing hangs, so the failure modes got the most attention:

  • Never blocks. Any failure — CLI missing, old, logged out, unresponsive — falls back to the built-in list. GET /api/models/:provider never rejects, so the picker always renders.
  • Hard timeout. 15s per probe, and SIGTERM escalates to SIGKILL after 2s so a CLI that traps signals can't outlive the probe holding its pipes open. The kill timer is unref'd.
  • Bounded pagination. Capped at 10 pages so a harness handing back an endless cursor can't spin.
  • Self-healing cache. Success cached 10 min; failure re-tried after 1 min, so the picker recovers on its own once a CLI is installed or logged in.
  • No stale writes. Concurrent callers collapse onto one probe. Each probe carries a generation stamp, so a slow superseded probe can't land last and undo the refresh that replaced it.
  • No stranded preferences. Built-in models the harness no longer serves stay in the picker, marked deprecated. But if the selected model is one of them, the client moves to the harness's own default rather than silently submitting a model the harness will reject.

API

Endpoint Description
GET /api/models/:provider source is discovered or static. ?refresh=1 bypasses the cache.
POST /api/models/:provider/refresh Drop cache and re-probe — for right after a CLI upgrade or login.
GET /api/models/providers Providers this build can probe.

The /model slash command now reads from the same source, so the two lists can't disagree.

Testing

Against the real Codex CLI, end to end through the Express route:

GET /api/models/codex      -> 200 in 824ms   source=discovered  default=gpt-5.6-sol  count=14
GET /api/models/openrouter -> 200 in 555ms   source=discovered  count=342
cached codex route call: 3ms

26 tests drive the real discovery code against a fake harness that speaks the same JSON-RPC over stdio — no CLI install required in CI. Covered: multi-page pagination, endless-cursor bounding, JSON-RPC errors, non-JSON banner output, silent hangs, hidden-model filtering, multi-byte UTF-8 split across stdout chunks, TTL behaviour, concurrency collapse, forced refresh vs. in-flight stale probe, and unknown providers.

Full suite 133 passed; npm run typecheck and npm run build clean.

Review: Codex reviewed this independently and reported 7 issues. Six were real and are fixed here with tests: the discovered default being ignored by the UI, SIGTERM never escalating, forced refresh joining a stale in-flight probe, pagination stopping after two pages, the hook keeping the previous provider's options during a switch, and UTF-8 corruption from per-chunk toString(). The seventh — a claim that discovery could never work without an initialized notification — is contradicted by the live run above, which succeeded without it; the notification is sent anyway, since the documented lifecycle expects it and other versions may enforce it.

Docs

docs/configuration.md gains a Model Discovery section covering sources, caching, timeouts, and the API.

…ding it

The model lists in shared/modelConstants.js are hand-maintained, so every time
a CLI ships a new model somebody has to notice and send a patch — and until
they do the picker offers models that no longer exist while hiding the ones
that do.

That drift is already real. codex-cli 0.145.0 serves gpt-5.6-sol, gpt-5.4-mini
and gpt-5.3-codex-spark; the pinned list still offered gpt-5.6, o3 and o4-mini.
OpenRouter's live catalogue is 342 models including claude-opus-5, against ~60
pinned.

This asks the tool instead:

- Codex via `codex app-server` line-delimited JSON-RPC (initialize, then
  model/list with nextCursor pagination) — the same catalogue the Codex CLI's
  own picker reads.
- OpenRouter via its public /api/v1/models endpoint.
- Everything else keeps its built-in list; those CLIs expose no equivalent.

Discovery is strictly additive and can never make things worse:

- Any failure — CLI missing, old, logged out, unresponsive — falls back to the
  built-in list. The endpoint never rejects, so the picker always renders.
- 15s hard timeout per probe, with SIGTERM escalating to SIGKILL so a CLI that
  traps signals cannot outlive the probe holding its pipes.
- Results cached 10 minutes; failures re-tried after 1 minute so the picker
  recovers on its own once a CLI is installed or logged in.
- Concurrent callers collapse onto one probe. Probes carry a generation stamp
  so a slow superseded probe cannot land last and undo a refresh.
- Models in the built-in list that the harness no longer serves stay in the
  picker, marked deprecated, so a saved preference is never stranded — but if
  the *selected* model is one of them, the client moves to the harness default
  rather than submitting a model the harness will reject.

New: GET /api/models/:provider (?refresh=1), POST /api/models/:provider/refresh,
GET /api/models/providers. The /model slash command now uses the same source.

Verified against the real codex CLI end to end: 824ms cold, 3ms cached, correct
live list and default. 26 tests drive the real code against a fake harness
covering pagination, endless cursors, JSON-RPC errors, non-JSON banner output,
silent hangs, multi-byte UTF-8 split across stdout chunks, cache/TTL behaviour,
concurrency collapse, and stale-probe ordering.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Copilot AI lite review requested due to automatic review settings August 5, 2026 18:56

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR adds server-side model discovery so the UI and /model command can prefer each harness’s live model catalogue (Codex via JSON-RPC; OpenRouter via HTTP) instead of relying solely on compiled-in lists, while keeping safe fallbacks, caching, and timeouts.

Changes:

  • Introduces a discovery subsystem (harnessModelDiscovery) with caching, timeouts, pagination bounding, and fallback-to-static behavior.
  • Adds protected /api/models/* endpoints and updates the /model slash command to read from the same source.
  • Updates the chat model picker to consume discovered model options via a new useHarnessModels hook, and documents the behavior.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 3 comments.

Show a summary per file
File Description
src/components/chat/view/subcomponents/ChatComposer.tsx Uses useHarnessModels to prefer discovered options and rescues retired stored selections to a harness default.
src/components/chat/hooks/useHarnessModels.ts New client hook to fetch /api/models/:provider with caching/refresh behavior.
server/utils/harnessModelDiscovery.js New discovery implementation for Codex/OpenRouter with merge + caching + concurrency collapse.
server/utils/tests/harnessModelDiscovery.test.js Unit/integration-style tests for discovery transport, pagination, timeouts, cache semantics, and edge cases.
server/routes/models.js New Express routes for model discovery, refresh, and provider listing.
server/routes/commands.js Updates /model command to use getModelsForProvider instead of constants.
server/index.js Mounts the new /api/models routes under auth.
server/tests/models-route.test.mjs HTTP-level tests for /api/models/* route behavior.
docs/configuration.md Adds a “Model Discovery” section describing sources, caching/timeouts, and API.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

for (const option of discovered) {
if (!option?.value || seen.has(option.value)) continue;
seen.add(option.value);
merged.push({ value: option.value, label: option.label || option.value });
Comment thread server/routes/commands.js
Comment on lines +196 to +200
const [claude, cursor, codex] = await Promise.all([
getModelsForProvider('claude'),
getModelsForProvider('cursor'),
getModelsForProvider('codex'),
]);
Comment on lines +37 to +61
// Guards against a slow response for a provider the user has already switched
// away from overwriting the current one.
const requestedProviderRef = useRef<string | null>(null);

const refresh = useCallback(() => setRefreshToken((n) => n + 1), []);

useEffect(() => {
if (!provider) {
setOptions(null);
setSource(null);
setDefaultModel(null);
return;
}

let cancelled = false;
requestedProviderRef.current = provider;
setIsLoading(true);
// Clear immediately rather than on response: otherwise the picker keeps
// rendering the previous provider's models for the duration of the request.
setOptions(null);
setSource(null);
setDefaultModel(null);

const query = refreshToken > 0 ? '?refresh=1' : '';
authenticatedFetch(`/api/models/${encodeURIComponent(provider)}${query}`)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants