[codex] fix resilient coordinator triage output

[VIVAAN-DHAWAN]

Fixes #456

Summary

• tolerate coordinator triage JSON wrapped in markdown fences or surrounding prose
• default missing single-label fields to safe coordinator choices and choose the first allowed label when a model returns extra allowed values, including multiple dispatch labels
• ignore irrelevant labels for unclear/out-of-scope decisions
• guide triage comments toward a warmer maintainer style modeled after recent lefarcen issue responses: acknowledge the reporter, reference concrete issue details, name routing/status, and give one next step

Root cause

The coordinator treated the LLM response as exact raw JSON and rejected otherwise usable decisions when models returned fenced JSON, prose-wrapped JSON, missing label groups, extra labels on non-valid dispositions, or more than one dispatch label. That made triage silently no-op in common agent-output cases.

Validation

• go test ./internal/coordinator/triage -count=10
• go test ./internal/coordinator ./cmd/looperd ./cmd/looper -count=1
• installed the rebuilt local looper/looperd binaries and restarted the daemon
• live coordinator verification triaged 5 new marked issues across 5 repos: looper-live-lefarcen-20260524T101336Z-01: Add a focus-mode study timer VIVAAN-DHAWAN/ai-study-companion#16, looper-live-lefarcen-20260524T101336Z-02: Quiz answer parser fails when response has spaces VIVAAN-DHAWAN/termux-math-practice#3, looper-live-lefarcen-20260524T101336Z-03: Document OPENAI_API_KEY setup for local runs VIVAAN-DHAWAN/readme-ai-generator#1, looper-live-lefarcen-20260524T101336Z-04: Billing screen is broken VIVAAN-DHAWAN/usage-based-billing#5, and looper-live-lefarcen-20260524T101336Z-05: Add a Fortnite skin generator VIVAAN-DHAWAN/github-roaster#1
• live scenarios covered valid feature, valid bug, docs, unclear/needs-info, and out-of-scope/wontfix; each ended with one Looper coordinator comment and the expected routing labels

Notes

I sampled recent lefarcen comments on nexu-io/open-design issues nexu-io/open-design#2815, nexu-io/open-design#2801, nexu-io/open-design#2794, and nexu-io/open-design#2785. The common pattern was warm acknowledgement, concrete issue recap, explicit routing/status, and a clear next step; this PR adds that guidance to the coordinator prompt without changing the comment stamping/disclosure path.

[quangdang46]

🤖 Hermes QA Report | PR #458 — [codex] fix resilient coordinator triage output

Verdict: APPROVED ✅ — Code is well-structured, properly tested, and addresses a concrete silent-noop bug.

Quality Gates

Check	Result
go vet ./...	✅ PASS
go build ./...	✅ PASS
go test ./internal/coordinator/triage/...	✅ PASS (15/15)

Code Review Summary

What this PR does (3 changes in 2 files, +197/−20):

1. Resilient JSON parsing — Adds extractJSON() that strips markdown fences (```json) and prose wrapping before json.Unmarshal, so the coordinator no longer silently no-ops when LLMs return fenced or prose-wrapped JSON output.

2. Graceful defaulting — Replaces strict requireExactlyOne() with requireAllowedOrDefault() that tolerates missing label fields by falling back to sensible defaults (kind → kind/feature, area → area/runtime for bugs, dispatch → dispatch/plan, complexity → complexity/m). Labels from out-of-scope/unclear dispositions are now ignored instead of rejected.

3. Warmer comment style — Added maintainer-style comment guidance to the prompt with reporter @-mention, concrete issue reference, clear next steps. Added area/sweeper to allowed areas.

Positives:

• Properly tested: 133 new lines of tests covering fenced JSON, text-wrapped JSON, missing labels, extra labels, out-of-scope ignore, first-allowed selection, and warm-style prompt guidance. All existing tests preserved.
• Clean design: The extractJSON → requireAllowedOrDefault chain is simple, composable, and easy to verify. No new abstractions introduced.
• Well-scoped: 197 additions focusing only on the triage output parsing layer. No scope creep.
• PR body is substantive: Root cause analysis, validation with live testing on 5 repos, and concrete examples of what broke.

Concerns:

• None. Code is clean, tests are thorough, changes are minimal and well-justified.

Recommendation

Ready to merge. The upstream origin/main has drifted significantly (137 file diff from branch base), so the PR will need a rebase/merge before landing, but the code itself is clean.

Reviewed at 2026-06-21 from Hermes QA cron.