loopback: Instructions file for Codex

Agent instructions

Working the Loopback queue

Loopback is the feedback bus for this repo. Real product usage — pinned UI feedback, backend failures with response bodies, UX papercuts, feedback on AI features with their run metadata — lands in one queue. Your job is to close loops: fix what real usage surfaced, and write the outcome back so the reporter sees the pin turn green.

Project slug for this repo: loopback.

The loop

Work one item at a time, most severe first:

1. loopback_list_feedback(project="loopback", status="open") — see what users actually hit. Check triaged too if the open queue is empty.
2. loopback_claim_feedback(id, agent=<your name>) — claim as your own agent name: claude-code, codex, gemini, or your CLI's name. If the claim is rejected, another agent holds it — pick a different item. Never force-claim (force=true) without first stating why in a loopback_add_comment.
3. loopback_get_feedback(id) — read ALL captured context before touching code. A pin is an anchor, not a scope: a pin on a form button often carries the backend root cause. Specifically read:
   • network and extra.failed_responses — failing calls with up to 2KB of response body (error codes, hints, stack fragments);
   • console and repro_steps;
   • extra.context — run_id / model / trace_url when the feedback is on an AI or automation feature. Chase the trace, not just the DOM.
4. Comment your root-cause diagnosis via loopback_add_comment before fixing — it is the audit trail that makes the queue trustworthy.
5. Fix it in this repo. Smallest change that addresses what was reported.
6. loopback_link_change(id, repo, branch, commit, pr_url, diff_summary) — the fix must be traceable from the feedback item.
7. loopback_update_status(id, status="fixed", note=...).
8. Verify for real, don't assume: UI issues → drive the running app with your browser tool/MCP; backend or logic → run the tests or hit the endpoint; usage/metric issues → check the metric or replay.
9. Only after verification: loopback_resolve_feedback(id, resolution="verified", note=..., agent=<your name>). Pass agent — this is the write that turns the reporter's pin full green, and without it the trail records the default "agent" instead of you. Use wontfix with a reason when intentionally not fixing. Resolving flips the reporter's pin green — do not claim it until it is true.