Feature - Task Queue with Conviction Level #123

Closed
opened 2026-04-03 11:14:43 +02:00 by shoko · 0 comments
Owner

Summary

Enhance PM agent autonomy with conviction-level decision system.

Depends On

  • #116 (Foundation - Modularize Shell Script)
  • #118 (Foundation - Structured Logging)
  • #119 (Foundation - JSON-based IPC)

Details

PM agent currently works via manual delegation. This feature enables autonomous processing:

  • PM agent reads from ~/.kugetsu/queue.json
  • Evaluates conviction level from multiple signals:
    • PR state (draft, open, approved, changes requested)
    • Feedbacks on PR
    • Issue comments
    • Issue labels/ milestones
    • Repository context (rules, past similar PRs)

Conviction Level System

  • Below threshold → Flag for human decision, don't act
  • Above threshold → PM acts autonomously
  • Threshold → Configurable per repo or globally

Conviction Signals

Research needed to define:

  • What signals indicate need_review vs done?
  • What signals indicate "safe to merge" vs "needs human eyes"?
  • Default threshold values?
  • Per-repo vs global threshold config?

Open Questions (Research Needed)

  1. Conviction level: How to tune thresholds? What values make sense as defaults?
  2. Signal weighting: Should all signals be equal weight, or some more important?
  3. Human override: How does human set conviction level per decision?

0.2.x Only

This is a new feature. It will not be available on 0.1.x.

Acceptance Criteria

  • PM agent can process queue autonomously
  • Conviction level calculated from multiple signals
  • Below threshold = flag for human
  • Above threshold = PM acts
  • Threshold configurable per repo or globally
  • Research issues resolved with documented decisions
## Summary Enhance PM agent autonomy with conviction-level decision system. ## Depends On - #116 (Foundation - Modularize Shell Script) - #118 (Foundation - Structured Logging) - #119 (Foundation - JSON-based IPC) ## Details PM agent currently works via manual delegation. This feature enables autonomous processing: - PM agent reads from `~/.kugetsu/queue.json` - Evaluates conviction level from multiple signals: - PR state (draft, open, approved, changes requested) - Feedbacks on PR - Issue comments - Issue labels/ milestones - Repository context (rules, past similar PRs) ## Conviction Level System - **Below threshold** → Flag for human decision, don't act - **Above threshold** → PM acts autonomously - **Threshold** → Configurable per repo or globally ## Conviction Signals Research needed to define: - What signals indicate `need_review` vs `done`? - What signals indicate "safe to merge" vs "needs human eyes"? - Default threshold values? - Per-repo vs global threshold config? ## Open Questions (Research Needed) 1. **Conviction level**: How to tune thresholds? What values make sense as defaults? 2. **Signal weighting**: Should all signals be equal weight, or some more important? 3. **Human override**: How does human set conviction level per decision? ## 0.2.x Only This is a new feature. It will not be available on 0.1.x. ## Acceptance Criteria - [ ] PM agent can process queue autonomously - [ ] Conviction level calculated from multiple signals - [ ] Below threshold = flag for human - [ ] Above threshold = PM acts - [ ] Threshold configurable per repo or globally - [ ] Research issues resolved with documented decisions
shoko added the 0.2.x-onlyfeature labels 2026-04-03 11:18:22 +02:00
shoko added this to the v0.2.x milestone 2026-04-03 11:18:22 +02:00
shoko added 0.3.x-only and removed 0.2.x-only labels 2026-04-05 07:03:47 +02:00
shoko added the blocked label 2026-04-07 08:55:41 +02:00
shoko closed this issue 2026-04-07 14:17:52 +02:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: shoko/kugetsu#123