Skip to content

No revision path when an upstream maintainer requests changes on a filed PR #60

Closed

Description

@wolfgang-aura

What happened

openai/openai-agents-python#4890 was filed from run 20260906T104815Z-29582c.
A maintainer reviewed it and requested changes. Mailman has no stage for that.

The run's terminal state is READY_FOR_HUMAN_REVIEW, then the operator files
the PR and the record stops. There is no state, no command and no section of
mailman/procedure.md covering "an upstream reviewer asked for changes on a
filed PR". mailman --help lists nothing between prepare-submission and
retrospective that accepts maintainer feedback as input.

Impact

Every maintainer-requested revision is hand-driven: an operator reads the
review, edits the workspace directly and pushes. That loses the two properties
the harness exists to provide.

  • No reviewer pass. The revision that answers a maintainer goes upstream with
    less scrutiny than the original, which had a full primary/reviewer cycle.
  • No record. run.json still says READY_FOR_HUMAN_REVIEW with
    human_decision: null while the PR has moved two states past it. The
    retrospective and the contribution ledger both read from that record.

A revision is also where a run is most likely to go wrong: the maintainer has
named a constraint, and an agent working without the review text in its prompt
will not honour it.

Evidence

Next action

Add a revision path that reuses the existing agent cycle rather than a second
one:

  1. A REVISION_REQUESTED state reachable from a filed run, carrying the
    upstream review text, its author and its permalink.
  2. mailman fetch-review RUN_ID --pr URL to pull the review bodies and the
    inline comments into the run directory, the way fetch-issue pulls an
    issue.
  3. build-prompts includes that text, so primary and reviewer both work
    against the maintainer's stated constraint.
  4. The completion gate requires the revision to answer every requested change
    or record why it does not.

Until this exists, note in SESSION_HANDOFF.md which runs were revised by hand
so the retrospective does not read them as untouched.

Activity

  1. wolfgang-aura commented on Sep 10, 2026

    @wolfgang-aura
    OwnerAuthor

    mailman/maintainer_review.py reads the upstream review into the run, RunStatus.MAINTAINER_CHANGES_REQUESTED gives the run somewhere to move, and check_revision refuses to call the revision done until each requested change has an answer. CLI: mailman maintainer-review. Covered by tests/test_maintainer_review.py. Verified on main at 59da1d6.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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