[META] Dev-PM Agent Loop #215

New Issue

shoko · 2026-04-07T14:18:03+02:00

shoko commented

2026-04-07 14:18:03 +02:00

Summary

This is a meta issue tracking the Dev-PM Agent Loop system implementation.

Problem

Currently, when dev agent hits a question/blocker:

Dev agent stops
User manually checks logs
User decides and restarts dev agent with new prompt

Desired behavior:

Dev agent works...
Dev agent hits blocker → asks PM (low latency, local)
PM agent checks context (PR diff, issue, repo rules, etc.)
PM responds with guidance via PR comment (for user docs)
Dev agent continues with guidance
Loop continues until PR ready
Human approves PR to merge

Core Principles

Dev agent always creates PR - never stops at blockers without asking PM
PM agent is async - dev doesn't wait, PM picks up when ready
PM has checkpoints - complexity increases as PR approaches approval
Human always approves - PM approves for review, human merges

Two Logging Systems

System	Purpose
dev-*.log	Raw terminal output for debugging
context JSON	Clean user/assistant messages for analysis

(Logged decisions are for context, not for PM consultation - PM uses live data)

Phases

Phase 1: Dev Asks PM (Low Latency)

Issue: (to be created)
Status: Blocked - needs research
Scope: How dev asks PM, what triggers, mechanism

Phase 2: PM Simple Context Check

Issue: (to be created)
Status: Blocked
Scope: PR diff vs issue vs repo rules

Phase 3: PM Medium Context Check

Issue: (to be created)
Status: Blocked
Scope: Plus related issues + wider repo context

Phase 4: PM Complex Final Check

Issue: (to be created)
Status: Blocked
Scope: Full check before human approval

References

Original misaligned issue: #123 (closed)

## Summary This is a meta issue tracking the Dev-PM Agent Loop system implementation. ## Problem Currently, when dev agent hits a question/blocker: 1. Dev agent stops 2. User manually checks logs 3. User decides and restarts dev agent with new prompt **Desired behavior:** 1. Dev agent works... 2. Dev agent hits blocker → asks PM (low latency, local) 3. PM agent checks context (PR diff, issue, repo rules, etc.) 4. PM responds with guidance via PR comment (for user docs) 5. Dev agent continues with guidance 6. Loop continues until PR ready 7. Human approves PR to merge ## Core Principles - **Dev agent always creates PR** - never stops at blockers without asking PM - **PM agent is async** - dev doesn't wait, PM picks up when ready - **PM has checkpoints** - complexity increases as PR approaches approval - **Human always approves** - PM approves for review, human merges ## Two Logging Systems | System | Purpose | |--------|---------| | dev-*.log | Raw terminal output for debugging | | context JSON | Clean user/assistant messages for analysis | (Logged decisions are for context, not for PM consultation - PM uses live data) ## Phases ### Phase 1: Dev Asks PM (Low Latency) - **Issue:** (to be created) - **Status:** Blocked - needs research - **Scope:** How dev asks PM, what triggers, mechanism ### Phase 2: PM Simple Context Check - **Issue:** (to be created) - **Status:** Blocked - **Scope:** PR diff vs issue vs repo rules ### Phase 3: PM Medium Context Check - **Issue:** (to be created) - **Status:** Blocked - **Scope:** Plus related issues + wider repo context ### Phase 4: PM Complex Final Check - **Issue:** (to be created) - **Status:** Blocked - **Scope:** Full check before human approval ## References - Original misaligned issue: #123 (closed)

shoko added the meta label 2026-04-07 14:18:03 +02:00

shoko referenced this issue

2026-04-07 14:18:16 +02:00

Feature - Dev-PM Agent Loop (Phase 1: Dev Asks PM) #216

shoko referenced this issue

2026-04-07 14:18:23 +02:00

Feature - Dev-PM Agent Loop (Phase 2: PM Simple Context Check) #217

shoko referenced this issue

2026-04-07 14:18:31 +02:00

Feature - Dev-PM Agent Loop (Phase 3: PM Medium Context Check) #218

shoko referenced this issue

2026-04-07 14:18:38 +02:00