Goal

Add structured LLM stream failure diagnostics to PawWork session exports so the next stream failure can be attributed to the correct boundary instead of requiring per-incident guesswork.

When this task is done, a failed assistant turn should tell us whether the failure most likely came from local cancellation, PawWork's stream watchdog, SDK / transport stream reading, provider / gateway closure, or an unknown boundary with enough evidence to continue investigation.

This is a diagnostic foundation task, not a user-visible behavior fix.

Scope

In scope:

• Extend the existing llm_trace / session export diagnostics with a schema-versioned v2 shape or compatible v1 extension.
• Capture a compact stream phase timeline: request creation, SDK stream returned, watchdog armed, first event, first provider-progress event, last provider-progress event, failure/completion.
• Capture watchdog configuration and state: connectTimeoutMs, streamTimeoutMs, provider-progress state, and timeout/failure phase.
• Capture a sanitized error fingerprint for stream failures: constructor name, error name, message, code, cause name/message/code, and at most a safe stack/module hint.
• Capture abort state at failure time: whether the LLM abort signal was already aborted, plus any available abort provenance.
• Capture safe provider correlation data when available, such as request id / response id / non-sensitive response headers.
• Keep exports safe: do not include auth headers, cookies, prompt text, tool args, raw provider response body, or arbitrary URLs.
• Preserve current runtime behavior. This task should improve diagnosis only.

Out of scope:

• Translating terminated into user-facing copy.
• Changing retry behavior or timeout policy.
• Reworking the full watchdog architecture.
• Recording every stream chunk or provider packet.
• Building a general observability or telemetry platform.
• Provider-specific error taxonomy beyond safe raw fingerprints and phase classification.

Relevant files or context

Related issues / PRs:

• [Bug] Raw terminated from upstream stream close leaks to assistant message without translation #754 — raw terminated leaks from an upstream / transport stream close; current logs prove the symptom but not the failing boundary.
• [Bug] 30s LLM connect timeout aborts OpenAI reasoning streams (post-#729 residual) #755 — 30s first-provider-progress timeout; current counters were useful but lacked timeout policy and phase detail.
• [Bug] First-turn tool execution aborted without recorded source — model stops, user must resend #721 — first-turn Tool execution aborted; PR fix: preserve abort interrupt provenance #710 improved abort provenance and showed the value of targeted diagnostics.
• PR fix: preserve abort interrupt provenance #710 — preserved abort interrupt provenance for exported-session debugging.
• PR fix(config): gate Instance dispose on actual config change #722 — removed a known no-op config write trigger for scope-driven aborts.
• PR fix: defer LLM stream connect timeout to after HTTP request is sent #729 — deferred the LLM stream connect timeout until after the HTTP request is sent.

Likely files:

• packages/opencode/src/session/llm.ts
• packages/opencode/src/session/llm-trace/types.ts
• packages/opencode/src/session/llm-trace/recorder.ts
• packages/opencode/src/session/processor.ts
• packages/opencode/src/session/export.ts
• packages/opencode/src/session/message-v2.ts
• packages/opencode/test/session/llm.test.ts
• packages/opencode/test/session/export.test.ts

Observed gap from #754's second reproduction:

• The export shows UnknownError with data.message = "terminated" and flags.stream_error = true.
• The trace proves provider progress happened before failure.
• The trace does not show whether the abort signal was already aborted, whether a watchdog fired, what raw error class/code/cause was thrown by undici/SDK, or whether provider request correlation is available.

Verification

• Add focused unit tests for stream failure diagnostics on:
  • connect timeout before first provider-progress event;
  • mid-stream external iterator error after provider progress;
  • local abort / interrupt path if currently observable in the test surface.
• Add export tests proving diagnostics are included and sanitized.
• Add regression tests ensuring sensitive data is not exported: auth headers, cookies, raw prompt text, raw tool args, and response bodies.
• Run targeted checks, likely:
  • bun --cwd packages/opencode test test/session/llm.test.ts test/session/export.test.ts --timeout 30000
  • bun --cwd packages/opencode typecheck
  • git diff --check

Execution mode

Investigate and propose a plan first — the agent must post the plan as an issue comment and wait for an explicit "approved" comment before writing code or opening a PR.