Goal: Regularly review completed work and suggest process improvements (AGENTS.md updates, new skills/tools).
Current State
- Work happens, gets committed
- No systematic review of what worked/didn't
- AGENTS.md updates are manual and sporadic
- No learning loop
Desired State
- Weekly (or post-epic) automated review
- AI analyzes completed tasks, commits, failures
- Suggests concrete improvements:
- AGENTS.md updates (new conventions, gotchas)
- New skills for common patterns
- New tools for repeated manual steps
- Human approves/rejects suggestions
Design
Trigger Options
1. Cron job (weekly)
2. Post-epic completion hook
3. Failure threshold (N task failures triggers review)
4. Manual ava review-process
Review Process
1. Gather data:
- Completed tasks this period (task list --status=done --since=7d)
- Git commits with messages
- Any task failures or needs-help transitions
- Agent logs if available
2. Analyze patterns:
- What types of tasks succeeded/failed?
- What knowledge was re-discovered?
- What manual steps were repeated?
3. Generate suggestions:
- AGENTS.md diff with new conventions
- Skill definitions for patterns
- Tool specs for automation candidates
4. Present to human:
- Via Telegram (Ava)
- Or generate PR/task for review
Implementation Path
1. Start with manual trigger: ava review-process
2. Add weekly cron once manual works
3. Add post-epic hook
Files to examine
- Omni/Ava.hs, Omni/Agent/Telegram.hs - Ava bot
- AGENTS.md - current conventions
- Omni/Task/Core.hs - task queries
Success criteria
- Can run
ava review-process and get useful suggestions - At least one AGENTS.md improvement identified per week
- Suggestions are actionable (not vague)