Skip to content

Ask for review explicitly; don't rely on a vendor's auto-review #143

Description

@yihanzhu

Operator's rule, from evidence on 2026-08-27. Codex's auto-review fired on nearly every push for hours, then went silent for 22 minutes on #141 and #142 — and answered both within 15 seconds of an explicit @codex review.

The rule: whoever changes the code asks for the re-review by name, after a meaningful change. Never rely on a vendor's push-triggered review.

Why it is better than the implicit trigger, beyond reliability:

  • It separates meaningful changes from noise. Five hash rebaselines today did not deserve a re-review; the deadlock fix did. The pusher knows which is which; a push trigger cannot.
  • It pairs request with verdict, which is what makes head-binding sound — a verdict answering a known request on a known head is much harder to confuse with a stale pass.

Where it lands:

  • AGENTS.md — a working rule for any agent: after a meaningful fix, request the review; do not assume one is coming.
  • work/<slug>/plan.md reviewer rules — the lane's own review-on-pr is workflow-triggered and stays deterministic, but any step that wants a Codex verdict must ask for it, and must then wait for a verdict that answers that request.
  • The deferred fix-stage intent inherits it: its "push a fix" step ends by requesting the re-review.

Small docs change; queued behind the merges of #142 and #141 so it does not reset their converged state.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions