Greptile Summary

This PR adds a timeoutSeconds field to ModelProviderConfig and the matching Zod schema, then reads it in runEmbeddedAttempt to pass a per-provider timeout to the global undici dispatcher — useful for slow local providers like Ollama or LM Studio.

• The schema.help.ts file is missing the \"models.providers.*.timeoutSeconds\" help entry that every other ModelProviderSchema field has, and docs/.generated/config-baseline.sha256 was not regenerated. The pnpm config:docs:check CI gate will fail. Run pnpm config:docs:gen and commit the updated hash file.

Confidence Score: 4/5

Safe to merge after adding the missing help entry and regenerating the config baseline sha256.

One P1 finding: the config doc baseline was not regenerated after modifying ModelProviderSchema, and the help text for the new field is absent from schema.help.ts — this will cause pnpm config:docs:check to fail in CI. The feature logic itself is correct.

src/config/schema.help.ts (missing help entry) and docs/.generated/config-baseline.sha256 (stale hash)

This is a comment left during a code review.
Path: src/config/schema.help.ts
Line: 753

Comment:
**Missing help entry for `models.providers.*.timeoutSeconds`**

Every field in `ModelProviderSchema` has a corresponding entry in `FIELD_HELP` (see the block starting at line 739). Adding `timeoutSeconds` to the Zod schema without adding a matching `"models.providers.*.timeoutSeconds"` entry here will cause `pnpm config:docs:check` to detect baseline drift and fail. The sha256 in `docs/.generated/config-baseline.sha256` is also unchanged from the pre-PR state, confirming that `pnpm config:docs:gen` was not run.

Add the missing help entry after line 753 (after `authHeader`) and then run `pnpm config:docs:gen` to update the hash file:

```ts
  "models.providers.*.authHeader":
    "When true, credentials are sent via the HTTP Authorization header even if alternate auth is possible. Use this only when your provider or proxy explicitly requires Authorization forwarding.",
  "models.providers.*.timeoutSeconds":
    "Per-provider stream timeout in seconds. Overrides the global 30-minute default. Use this for slow local providers (e.g. Ollama, LM Studio) that require longer inference time.",
```

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/run/attempt.ts
Line: 348-349

Comment:
**Unnecessary type cast — `timeoutSeconds` is already in `ModelProviderConfig`**

`params.config?.models?.providers?.[params.provider]` resolves to `ModelProviderConfig | undefined`, and `ModelProviderConfig` now declares `timeoutSeconds?: number`, so the `as { timeoutSeconds?: number } | undefined` cast is redundant and silently bypasses TypeScript's structural checking for that expression. Drop the cast:

```suggestion
  const providerTimeoutSeconds =
    params.config?.models?.providers?.[params.provider]?.timeoutSeconds;
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "feat(models): add per-provider timeoutSe..." | Re-trigger Greptile}

Hey, any updates on this feature?

Codex automated review: keeping this open.

Keep this PR open. Current main still does not implement models.providers.*.timeoutSeconds; the PR remains a relevant implementation candidate, but it should be revised before merge because the current approach configures process-global Undici timeout state and the generated config schema surface is not fully updated.

Best possible solution:

Keep the PR open and adapt it. The best path is to add models.providers.*.timeoutSeconds consistently across the config type, strict Zod schema, help/labels, generated schema artifacts, docs, and tests, then resolve the selected provider's value and pass it through the provider request/transport fetch path as a per-request timeout instead of changing global Undici dispatcher state.

What I checked:

• Current main lacks the config type field: ModelProviderConfig currently includes baseUrl, auth/API/header fields, request, and models, with no timeoutSeconds member. (src/config/types.models.ts:117, 4878d3e059ce)
• Current main rejects the config key: ModelProviderSchema is strict and lists provider properties without timeoutSeconds, so raw config using the proposed key is not accepted on current main. (src/config/zod-schema.core.ts:352, 4878d3e059ce)
• Generated config metadata omits the field: The generated metadata has models.providers.*.authHeader followed by models.providers.*.request; there is no models.providers.*.timeoutSeconds entry for schema/UI consumers. (src/config/schema.base.generated.ts:26490, 4878d3e059ce)
• Runtime still uses run-level global timeout tuning: runEmbeddedAttempt calls configureEmbeddedAttemptHttpRuntime({ timeoutMs: params.timeoutMs }), and that helper calls ensureGlobalUndiciStreamTimeouts with the run timeout floor rather than any provider-specific config. (src/agents/pi-embedded-runner/run/attempt.ts:580, 4878d3e059ce)
• Per-request timeout seam already exists: buildGuardedModelFetch(model, timeoutMs) can pass a timeout to fetchWithSsrFGuard, whose guarded dispatcher path applies that timeout per request; however OpenAI, Anthropic, and Google transport call sites currently call it with only model. (src/agents/provider-transport-fetch.ts:157, 4878d3e059ce)
• Tests cover the seam, not provider config resolution: Existing provider transport tests verify an explicit timeoutMs argument is threaded into fetchWithSsrFGuard, which supports adapting the PR to resolve provider config and pass it through this path. (src/agents/provider-transport-fetch.test.ts:78, 4878d3e059ce)

Remaining risk / open question:

• Merging the PR as currently described could make concurrent provider calls inherit the wrong timeout because it mutates process-global Undici dispatcher state per embedded attempt.
• The config metadata path must include help, labels, generated schema artifacts, and baseline hashes so Control UI and schema consumers can discover the field.
• A complete fix needs focused tests for provider-config timeout resolution and per-request transport wiring, not only type/schema additions.

Codex Review notes: model gpt-5.5, reasoning high; reviewed against 4878d3e059ce.

Thanks @uninhibite-scholar — this was the right config surface.

Superseded on main by 18b76e3. I kept models.providers.*.timeoutSeconds, but wired it through the resolved runtime model and per-request guarded fetch path instead of mutating the process-global Undici dispatcher during an agent attempt. That means provider timeouts now apply to the model HTTP request only, including guarded fetch abort handling plus Undici connect/header/body timeouts.

Validation on the landed fix:

• pnpm test src/agents/pi-embedded-runner/model.test.ts src/agents/provider-transport-fetch.test.ts extensions/ollama/src/stream-runtime.test.ts src/infra/net/ssrf.dispatcher.test.ts src/infra/net/fetch-guard.ssrf.test.ts
• pnpm config:schema:check && pnpm config:docs:check
• pnpm check:changed
• live local Ollama: OPENCLAW_LIVE_TEST=1 OPENCLAW_LIVE_OLLAMA=1 OPENCLAW_LIVE_OLLAMA_MODEL=llama3.2:latest OPENCLAW_LIVE_OLLAMA_WEB_SEARCH=0 pnpm test:live -- extensions/ollama/ollama.live.test.ts

Closing this PR as superseded by the maintainer-owned implementation. Thanks again.