2026-06-23 · pull-request-review

Agent-to-agent PR comment and fix loop

PR review feedback can become an owned agent todo with fix evidence instead of a loose chat reminder.

The operator can let agents coordinate around PR comments without losing final review visibility or fix evidence.

handoffPR commentclaimed_by
LoopX Agent-to-agent PR comm claimed_by public safe pattern case goal / trigger todo / gate / evidence public outcome
01

Case context

This case shows how PR review feedback can move across multiple agents without losing ownership. The important chain is comment, claim, handoff, fix, validation, and review packet rather than the chat transcript.

The user-facing value is that every feedback item can answer three questions: who owns it, where the fix evidence is, and who still needs to review it. Public evidence spans the event contract, review packet, heartbeat prompt, and validation fixtures.

02

Repository evidence

Proof

Multiple agents can coordinate around PR review comments without losing owner review.

LoopX intervention

claimed_by, handoff gate, review packet, comment/fix loop

1owner per feedback item
3owner/fix/review questions answered
1review-packet handoff
0unowned PR comment reminders
Todo ownership

`event_sourced_state_contract_v0` defines `todo_claimed` as a canonical event for ownership, lease, or `claimed_by`.

Review packet

`loopx/review_packet.py` preserves `claimed_by` in open-todo rendering and handoff ranking so review packets can show ownership.

Side-agent contract

`docs/heartbeat-automation-prompt.md` requires side agents to self-merge only small validated evidence-backed work or create a claimed handoff todo.

CLI smokes

`examples/todo-lifecycle-cli-smoke.py` and `examples/todo-cli-smoke.py` cover claim, handoff successor, side-agent self-merge, review handoff, and same-agent handoff rejection.

03

LoopX behavior

04

What the user sees

The user does not have to shepherd every PR comment manually; the control plane shows who claimed the feedback, where the fix evidence is, and what still needs review.

Multi-agent collaboration stops depending on memory of which agent saw which comment.

05

Repository sources

Evidence boundary. Public-safe pattern case only; no private screenshots, raw chats, internal review notes, local state, credentials, or unpublished artifacts.