PR #2160 merged over a standing Needs-more-work verdict (round 10) and is unreverted

[dem-extra1]

PR #2160 ("fix(orchestrator): add existing file context injection, stub discard, and destructive-command gating")
merged at 2026-08-24T19:37:59Z over a standing Needs more work verdict, and has not been reverted.

Timeline

• Ten consecutive Claude review rounds between 17:38 and 19:24 UTC, every one of them Needs more work.
• The final verdict (comment 5400199490, 19:24, reviewed commit 560e8d2) reads:
  "Needs more work. ... Finding 1 above is a new, directly-reproduced instance of the same bypass class this round's own diff introduced, on a common and realistic filename shape."
• Two further pushes (bd9c7f6, 74f221a) received only "Claude review skipped --- API credential or quota unavailable" notices (the quota outage that began ~19:30).
• The PR merged at head 74f221a with no clean verdict, no self-review fallback, and no disposition of the round-10 finding.

Per fully-clean.md, a skip notice neither clears nor supersedes a prior not-clean verdict,
and a later push does not clear one either --- only a later clean verdict does.
The two commits after 560e8d2 may well have addressed finding 1, but nothing verified that.

Suggested disposition

Either revert (as gha#645 did for gha's #622-#633 batch), or run a retroactive review of the merged
delta (560e8d2..74f221a plus the round-10 finding's filename-shape bypass) and fix or file whatever it finds.

Related

• Five UMS entries merged with a Needs-more-work verdict: 14 findings, several factual errors on main #2174 tracks the sibling case from the same evening: ums: bank five process learnings from the 2026-08-24 session #2171 merged while its adversarial self-review was still running, which then returned Needs more work with 14 findings.
• The rest of the 21:56-22:36 UTC merge batch (fix(orchestrator): pipe model prompts via stdin, enforce utf-8 encoding, and escalate to frontier on retry #2164, fix(ci): call check-new-line-breaks composite action directly at frozen SHA (#2165) #2166-docs(memories): fix heredoc description in tools.md (#2151) #2170, Three skill files are hard-wrapped rather than semantically line-broken #2156, check-pr-fully-clean.py scores the code-review plugin's own clean comment as latest = NONE #2157, feat(orchestrator): add pure-llm preamble to claude adapter and prioritize subpath candidate files #2173, claude-review-workflow and claude-agent-workflow author standalone workflows gha already ships #2134) merged on
  adversarial self-review APPROVED verdicts under the quota outage, which self-review-fallback.md sanctions;
  they are not listed here as violations, though Five UMS entries merged with a Needs-more-work verdict: 14 findings, several factual errors on main #2174 shows the fallback's error rate that evening was not zero.

Sweep performed 2026-08-24 ~16:30 PT across all 57 PRs merged into ai-config main on 2026-08-24 (UTC),
classifying each latest-verdict-at-merge with check-pr-fully-clean.py's classify_verdict and
hand-confirming every flagged comment's last ### Verdict heading.
Every other merge either had a clean Claude verdict at merge or a posted self-review fallback verdict.