Line 686 Edited a Comment With No Snapshot

Line 686 of the ticket wrapper module edited the existing status comment with no snapshot of it first. A cross-review caught it at 11:55 AM.

By then a work group titled “JIRA comment tooling destroys existing comments” had already run from 10:48 AM to 11:25 AM. One commit in that group covered four tickets. It made edit-don’t-stack edit only comments carrying the agent marker, and never the author’s hand-written, review-round index or finding-tagged ones. It also snapshots the prior body to .runs/jira_comment_backups first, and it made the review-round index finder strict on the h3 R<N> index.

The worker’s report on those four tickets landed at 11:25 AM and said “the full gate was not run, as instructed.” Other workers that morning were stuck waiting on a gate. One of them was still waiting at 11:14 AM after 1590.24s. It named the command as python -m pytest -q -m "not slow" -n auto --dist loadfile and reported that nothing was committed or pushed yet.

What the second pair of eyes found

The reviewer was a separate run: codex exec, medium effort, 146.92s. It marked line 686 as a risk. The status command edits the existing status comment, and it did so without a snapshot, even though the commit above had already put a snapshot in front of edit-don’t-stack.

A second review ran at 11:57 AM and flagged a different file. cleanup_feature_branch() at line 604 of the git helpers checks origin/<branch> without fetching it. It was labelled a bug.

The follow-up commit

One commit carried the cross-review follow-ups. Two of its changes were in the comment tooling:

• The status command snapshots before editing.
• The agent marker only counts as a trailing suffix, and quoted markers are stripped.

The same commit also routed the SQL CLI through run_query (HY010).

After it landed, the status command at line 686 snapshots the existing comment before editing it, the way edit-don’t-stack already did.

dxdev LESSON · paste into your AI coding agent
LESSON: When One Code Path Gains A Safety Step, Audit Every Sibling Path That Mutates The Same Data
SOURCE: dxdev.com/blog/edit-in-place-only-touches-its-own-comments
WHAT HAPPENED: A work group titled "JIRA comment tooling destroys existing comments" produced a commit that restricted edit-don't-stack to agent-marked comments and snapshotted the prior body to `.runs/jira_comment_backups` first. The status command edits the same kind of comment, but the commit did not add a snapshot there. A separate `codex exec` review, run at medium effort for 146.92s, flagged line 686 as a risk because of the missing snapshot. A follow-up commit made the status command snapshot before editing. It also tightened the agent marker so it only counts as a trailing suffix and quoted markers are stripped.
THE RULE: When you add a safeguard such as a backup to one code path that mutates data, find every other path that mutates the same data and apply the safeguard there too. An independent review is a good way to catch the sibling paths the original author missed.
CHECK MY CODE, then report PASS or FAIL with file:line for each:
      1. Every code path that edits an existing JIRA comment writes a snapshot to the backup directory before the edit, not just edit-don't-stack.
      2. A test edits a comment through each entry point, including the status command, and asserts a backup file exists with the prior body.
      3. When a safeguard is added in a commit, a search for other callers of the same edit or update function is recorded in the review, and each caller is either covered or explicitly exempted.
THEN PRINT: a table (check, PASS/FAIL, evidence, fix) + a verdict (applies / partially / OUT_OF_SCOPE / no) + the single most important next action.