fix(#2786): surface agent API errors when Claude Code exits 0

[fullsend-ai-coder]

Claude Code exits with code 0 on API/infrastructure failures (e.g., invalid_grant, quota exhaustion) while setting is_error:true in the transcript result event. This made infrastructure errors invisible — green workflows with no PR, no comment, and no GHA error annotation.

Two changes in run.go:

1. After exit code 0, parse the tee'd output.jsonl for
   is_error:true result events. If found, override
   lastExitCode to 1 and emit StepWarn so downstream gating
   treats the run as failed.

2. Remove the lastExitCode != 0 gate around
   ParseTranscriptErrors so ::error:: annotations are always
   emitted, regardless of exit code.

Added ParseTranscriptFile to the TranscriptHandler interface to support checking a single JSONL file (output.jsonl) without scanning an entire directory.

Note: make lint could not run (shellcheck-py fails to download in sandbox). go vet and go test both pass.

---

Closes #2786

Post-script verification

• Branch is not main/master (agent/2786-surface-api-errors)
• Secret scan passed (gitleaks — 43388d57a168ed0b5caf97463414003677404fa2..HEAD)
• Pre-commit hooks passed (authoritative run on runner)
• Tests ran inside sandbox

[github-actions]

Site preview

Preview: https://5728bdf3-site.fullsend-ai.workers.dev

Commit: 5f1462de4478a49fabbc65361ffac12c39e6cd37

[waynesun09]

/review

[github-actions]

E2E tests did not run

E2E tests run automatically for org/repo members and collaborators on pull requests.

For other contributors, a maintainer must add the ok-to-test label after the latest push.

See E2E testing guide for details.

[waynesun09]

/review

[qodo-code-review]

Persistent review updated to latest commit 5f1462d

[waynesun09]

LGTM. Reviewed with a 5-agent review squad (Claude, Gemini, Codex), 3 fix-verify iterations.

Core logic is correct: is_error:true in the final transcript result event is a definitive session failure signal (API errors, quota exhaustion, stale OIDC tokens). No false positive risk — only fires on exit code 0 with a terminal error in the transcript.

Fixes applied on top of the original commit:

• transcriptErrorOverride flag scoped to transcript-error overrides only (not all non-zero exits), reset each iteration
• StepWarn prints lastExitCode (effective value) instead of exitCode
• Post-script correctly skipped on transcript errors

codecov/patch failure is expected — runAgent is an orchestrator function that can't be unit-tested. The underlying parseTranscriptFile has thorough test coverage.

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 9:51 PM UTC · Completed 10:00 PM UTC
Commit: 5f1462d · View workflow run →

[fullsend-ai-retro]

Retro: PR #2860 — Surface agent API errors when Claude Code exits 0

Timeline

1. 2026-06-30 16:07 — Issue #2786 filed by waynesun09 with detailed root-cause analysis and proposed fix
2. 2026-06-30 16:10 — First triage run 28458573867 failed
3. 2026-07-01 17:46 — waynesun09 manually re-triggers /fs-triage
4. 2026-07-01 17:49–17:52 — Second triage run 28536784506 succeeded, applied labels (type/bug, priority/high, component/runner, ready-to-code)
5. 2026-07-01 17:52–18:08 — Code agent run 28537091526 succeeded, created PR #2860 with commit deacf00
6. 2026-07-01 20:33 — waynesun09 manually runs /review (triggering qodo-code-review, not the fullsend review agent)
7. 2026-07-01 20:34 — qodo-code-review flags three concerns: exit code reporting inconsistency, post-script gating gap, and interface ripple
8. 2026-07-01 20:34 — waynesun09 pushes fix commit 5f1462d addressing post-script gating and StepWarn cleanup
9. 2026-07-01 20:44 — waynesun09 approves PR after multi-agent review (Claude, Gemini, Codex)
10. 2026-07-01 21:47 — PR merged

What went well

• Triage quality: The second triage run produced an excellent summary that correctly identified the root cause, recommended fixes, and even proposed test cases.
• Code agent output: The code agent implemented the two fixes proposed in the issue (remove exit-code gate on ParseTranscriptErrors, add output.jsonl parsing for is_error:true) and added three well-targeted tests.
• Issue quality: The issue itself was exceptionally well-written with root-cause analysis, code references, and a real-world example.

What needed rework

The human pushed a follow-up commit fixing two gaps in the code agent's implementation:

1. Post-script gating: The code agent overrode lastExitCode to 1 on transcript errors but didn't propagate this to the post-script defer block, which gates on runErr (not lastExitCode). A new transcriptErrorOverride flag was needed to skip post-script execution.
2. StepWarn simplification: The code agent introduced a three-way if/else branch distinguishing "real non-zero exit" from "transcript error override" — the human collapsed this to a clean two-way branch using lastExitCode consistently.

Skipped proposals (covered by existing issues)

• Review agent visibility: The fullsend review agent did not post any comment on this agent-authored PR. #2637 ("Review agent should always leave a visible trace on agent-authored PRs") already covers this gap.
• First triage failure: The initial triage failed without a clear diagnostic. #1283 (surface triage token minting errors) and #1919 (startup_failure silently lost) cover related concerns.

Proposals filed

• Code agent should trace all consumers of modified gating variables before committing (in fullsend-ai/fullsend)