Skip to content

A review comment can post only the tail of the review, referencing analysis it never included #710

Description

@d-morrison

On Morrison-Lab/ai-config#2522, the review workflow posted a comment that references analysis it never included.

The whole posted body (1532 characters, comment, run 33159592591) is a follow-up section about check-pr-fully-clean.py's exit status, ending:

No check has failed, and the diff has no content defects (verified above against the real external PR data). My verdict stands unchanged:

There is no "above". No summary, no verification section, no findings list -- the comment opens on the instrument re-run. "My verdict stands unchanged" refers to a verdict that was never posted.

It is not an edit: created_at and updated_at are both 2026-08-28T09:34:40Z, and no other comment or formal review exists on the PR. Every job in the run reports success, including post-review, so nothing is red.

Two things point at the same mechanism -- the posted comment being the tail of the assistant's output rather than the whole review:

  1. The body starts mid-argument, at what reads as the last section of a longer message.
  2. It ends with **Stopping Page**: Clean stopping point reached -- review posted, no findings, no PR state left to act on in this run. That is a session-recap convention from the reviewing agent's own instruction corpus, not review content. Its presence in a CI-posted comment is the signature of a harness taking the final assistant message.

That failure mode is already documented, in the corpus this workflow loads as a plugin: shared/workflow/flag-session-boundaries.md records d-morrison/rme#1081, where installing the corpus made every prose reply collapse to a one-line stopping-point declaration, because the harness posts the last assistant message. This looks like the same mechanism catching a longer tail rather than only the declaration.

Consequence: the verdict is clean, dated, and names the correct head, so check-pr-fully-clean.py passes it -- but a reader cannot see what was reviewed, and the comment's own text advertises that something is missing. Same family as #709: the check color and the SHA are right, and the account of the work is not.

Worth checking whether post-review concatenates the assistant's turn or takes one message, and whether the reviewing prompt should suppress the stopping-point convention, which belongs to interactive sessions rather than to a CI comment.

Activity

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions