Summary

Add ProviderTransform.streamTimeouts(model) returning { connectTimeoutMs: 120_000 } for reasoning-capable models (gated on model.capabilities.reasoning), and apply it at the two production llm.stream() call sites — session/processor.ts (main response) and session/prompt.ts (title generation) — with helper-first spread order so any caller-provided StreamInput.connectTimeoutMs still wins. The default CONNECT_STREAM_TIMEOUT_MS is untouched; non-reasoning models keep the 30s ceiling.

Why

The 30-second first-progress watchdog in session/llm.ts aborts reasoning-model streams whose first observable provider event arrives later than the ceiling. The incident recorded in #755 shows OpenAI gpt-5.5 on a long session spending 30204ms with only start: 1 and zero content events before the watchdog aborted. PR #729 fixed an earlier residual where the connect timer armed before the HTTP request was actually sent; the residual addressed here is the 30s ceiling itself, which #729's body explicitly deferred. The capability gate keys off model.capabilities.reasoning after verifying the current models.json catalog has correct reasoning=true labels on the gpt-5 family (22/23, only gpt-5.3-chat-latest excluded), all o-series, Claude haiku/sonnet/opus 4.x thinking variants, and the Gemini 2.5+/3.1 series — so no provider-id allowlist fallback is needed at this time.

Related Issue

Refs #755 (short-term path; the full deferred set — SessionRetry.policy.retryable() not classifying local timeouts, watchdog architectural rewrite, and the parallel mid-stream terminated failure mode — is documented in the issue body and left to follow-up PRs).

Human Review Status

Pending

Review Focus

• Helper-first spread order at session/processor.ts:954 and session/prompt.ts:465. processor.ts receives streamInput from process() callers (session/prompt.ts:2020, session/compaction.ts:480); neither sets connectTimeoutMs today, so the spread defaults to the helper value. The order is helper-first specifically so that a future caller wishing to override does not need code changes here.
• Capability-based gate vs model-id-pattern matching. Catalog cross-checked at PR-prep time and labels are reliable across the four target families; a second pair of eyes on whether model.capabilities.reasoning === true is the right axis is welcome.
• The new contract test floor at >= 90_000 in transform.test.ts codifies the policy direction (the lowest value considered for reasoning models during the issue discussion). Worth questioning whether a corresponding upper bound is also needed.
• Title-generation path applies the helper to mdl from provider.getSmallModel(input.providerID) or the title agent's explicit override. The typical small model is non-reasoning so the helper returns {}; only when a user explicitly configures a reasoning small variant for the title agent does the 120s apply, and the failure mode is silent (caught at prompt.ts:482-488, logs and falls back to the default title). The deliberate choice was to avoid a per-mode branch inside the helper.

Risk Notes

• After the 120s ceiling is reached, a first-progress timeout still surfaces as a hard UnknownError because SessionRetry.policy.retryable() does not type-tag local timeouts and does not match the bare error message. The user-experience improvement here is the lower hit rate; no automatic retry is added in this PR. Retry classification is deferred until the parallel mid-stream terminated failure mode is analyzed, so retry can be designed once against both failure shapes rather than speculatively per-shape.
• Explicit connectTimeoutMs: undefined from a caller would clobber the helper value during spread and fall through to the 30s default via llm.ts:434-439. No production path constructs streamInput this way today, but a future caller with conditional override should pass a positive number or omit the field rather than set it to undefined.
• Behavior change is strictly scoped to reasoning-capable models. Non-reasoning models keep the 30s default.
• No visible UI or copy changed; the visible-UI conditional checkbox is left unticked for that reason.
• No platform / packaging / updater / signing / paths / shell / permissions surface was touched; the macOS/Windows conditional checkbox is left unticked for that reason.
• No docs / release notes / dependencies / permissions / credentials / deletion behavior / generated content / local file changes; the related conditional checkbox is left unticked for that reason.

How To Verify

typecheck (bun --cwd packages/opencode run typecheck):                       ok
bun test packages/opencode/test/provider/transform.test.ts (streamTimeouts):  3 pass / 0 fail
bun test packages/opencode/test/session/ packages/opencode/test/provider/:    940 pass / 4 skip / 1 todo / 0 fail
internal cross-review (Claude Opus + Codex high, parallel):                   0 P0 / 0 P1 / 1 P3 both reviewers flagged (policy floor too loose) fixed by tightening test to >= 90_000

Screenshots or Recordings

Not applicable (no UI change).

Checklist

• Type label — this PR carries exactly one of bug, enhancement, task, documentation. Type labels are author-added; the labeler bot does NOT assign them. Add the label in the GitHub UI, then tick this.
• Routing labels — this PR carries at least one of app, ui, platform, harness, ci. The labeler bot assigns these on PR open based on changed paths. Confirm the bot's choice (or override if wrong), then tick this.
• Priority label — this PR carries exactly one of P0, P1, P2, P3. The priority-triage bot suggests one on PR open. Confirm or override, then tick this.
• Human Review Status above is set to Pending, Approved by @<reviewer>, or Not required: <reason> (default is Pending; "not required" is restricted to bot-authored low-risk PRs).
• I linked the related issue, or stated in Summary why there is no issue.
• I described the review focus and any meaningful risks.
• I replaced the example block in How To Verify with the real verification steps and the key result for each.
• I did not introduce unrelated refactors, dependencies, generated files, or file changes beyond the stated scope.
• (conditional) I manually checked visible UI or copy changes when needed, with screenshots or recordings. Leave unticked only if no visible UI or copy changed.
• (conditional) I considered macOS and Windows impact for platform, packaging, updater, signing, paths, shell, or permissions changes. Leave unticked only if no platform/packaging surface was touched.
• (conditional) I called out docs, release notes, dependencies, permissions, credentials, deletion behavior, generated content, or local file changes when relevant. Leave unticked only if none of those surfaces was touched.
• I reviewed the final diff for unrelated changes and suspicious dependency changes.
• I am targeting dev, and my PR title and commit messages use Conventional Commits in English.

Summary by CodeRabbit

• New Features

  • Added intelligent timeout management for AI model streaming with extended timeouts for reasoning-capable models, improving stability during complex inference tasks.
  • Enhanced title generation streams with optimized timeout configuration for better performance on reasoning models.

• Tests

  • Added test coverage for new timeout management functionality, validating behavior across different model types.