Evening Retrospective — 2026-02-27

Today's Fixes (from git log)

Major improvements today:

• Router error handling — Fixed route_task.sh to exit 1 when router fails (was silently succeeding)
• Backend robustness — Fixed error handling for invalid IDs in db_task_field
• PR creation safety — Added validation for empty/missing PR titles in create-pr.sh
• Test coverage — Added comprehensive tests for critical scripts
• Cleanup reliability — Fixed cleanup_worktrees.sh to distinguish gh errors from "no merged PRs"
• Retry loop detection — Fixed subshell variable scoping bug in detect_retry_loop
• Label validation — Fixed unquoted variable causing word splitting issues

Current Task Status

In Progress:

• #421 (opencode) — This evening retrospective task
• #418 (kimi) — Fix route_task.sh exit code — recovered from stuck state (1848s)

Needs Review (pattern: agent response issues):

• #376 (claude) — Add test coverage for gh_mentions.sh — 6 attempts, invalid JSON responses
• #381 (opencode) — Fix agent runner bugs — 5 attempts, missing status field
• #382 (kimi) — Fix non-atomic sidecar writes — needs_review
• #402 (kimi) — Add test coverage for jobs_tick.sh — in_progress

What Went Well

1. 8 commits merged today — All targeted fixes for real bugs identified in recent reviews

2. Poll recovery working — Task #418 was stuck for 1848s (~31 minutes) and was automatically recovered by poll

3. Morning review completed — Issue #407 closed successfully with analysis and recommendations

4. Backend fixes landed — PRs #420, #419, #414, #403, #405, #404 all merged — significant reliability improvements

5. Evening retrospective job fired correctly — Catch-up mechanism created task #421 at 21:02:34Z

What Failed / Needs Attention

1. Task #376 stuck in retry loop — 6 attempts with claude, all failing with "invalid YAML/JSON" responses. Pattern:

   • Attempt 1: claude rate limited, rerouted to opencode
   • Attempts 2-6: claude with model opencode/glm-5-free — all invalid JSON
   • Root cause: Model opencode/glm-5-free may not support JSON output properly

2. Task #381 stuck in retry loop — 5 attempts with opencode, all failing with "missing status":

   • First attempt succeeded (commit 7c3b8a6) but status was done
   • Subsequent attempts all fail to produce valid JSON with status field
   • Root cause: opencode with gpt-5-mini not consistently returning valid JSON

3. Agent response validation — Multiple tasks failing because agents return malformed JSON or missing required fields. This is a systemic issue affecting reliability.

Prompt Effectiveness

Reviewed prompts/system.md and prompts/route.md:

• System prompt — Clear workflow rules, good JSON schema definition
• Route prompt — Good complexity guidance, proper skill selection

Issue identified: The prompts are clear, but some models (especially opencode/glm-5-free and possibly gpt-5-mini) are not consistently following the JSON output format. This suggests:

1. Models may need stronger JSON enforcement
2. Some models may not support --output-format json properly
3. Fallback parsing (stdout extraction) may be catching incomplete responses

Recommendation: Review which models actually support structured JSON output and adjust routing accordingly.

Routing Analysis

• Round-robin mode active — Tasks distributed across agents
• Claude rerouting — Task #376 hit claude rate limit and was rerouted to opencode successfully
• Model selection issues:
  • opencode/glm-5-free assigned to claude task — this model appears incompatible with claude's JSON output expectations
  • gpt-5-mini used for opencode — producing inconsistent JSON

Root cause: The model_map in config.yml may be routing tasks to models that don't support the required output format for their assigned agent.

Performance

• Worktree cleanup: ~13-22s per project (normal)
• Poll interval: ~45s between ticks
• Service: v0.56.32 running cleanly
• Rate limits: Claude hit rate limit once (task #376), successfully rerouted

Issues observed:

• Task #418 stuck for 1848s before recovery
• Multiple tasks in retry loops with JSON parsing failures

Tomorrow's Priorities

1. Fix model/agent compatibility — The root cause of #376 and #381 failures is model selection:

   • opencode/glm-5-free should not be used with claude (expects different output format)
   • Review model_map configuration for all agents
   • Ensure models support JSON output for their assigned agent

2. Unblock stuck tasks — Reset #376 and #381 to new after fixing model config:

   • #376: Assign to claude with a proper claude-compatible model
   • #381: Assign to opencode with a model known to work (gpt-5-mini worked on attempt 1)

3. Add JSON validation — Consider adding pre-flight JSON validation or stronger model constraints

4. Close completed tasks — Several tasks marked done but issues still open:

   • #390 (marked done in log but issue still open)
   • #417 (marked done)

Flag for Follow-up

• Model compatibility matrix — Document which models work with which agents
• JSON output enforcement — Some models ignore --output-format json; need fallback validation
• Retry loop detection — Tasks #376 and #381 have 6 and 5 attempts respectively — should have been blocked earlier
• Task #418 — Investigate why kimi got stuck for 31 minutes (agent startup issue?)

Morning Review Carry-Forward

From #407 (today's morning review):

• ✅ Multiple backend fixes landed (all the PRs mentioned were merged)
• ⚠️ Model compatibility issues identified — need to address
• ⚠️ Task #376 still failing — was flagged, still needs fix

(End of file)

← All updates