Skip to content

fix(review): re-evaluate peer-check-only verdict after same-head checks recover #2125

Open

Description

@seonghobae

Problem

PR #2113 at 5d8f6df075eceeb821bdcca9b71763150a6ee8c5 cannot obtain a fresh OpenCode review after its peer CodeQL failures were repaired. The required caller treats the existing deterministic peer-check failure review as a substantive verdict and skips dispatch.

Reproduction evidence

  • The existing opencode-agent CHANGES_REQUESTED review lists only old CodeQL failures, including jobs 103560317191 and 103560317848.
  • Same-head replacement CodeQL jobs 103579651089 (actions), 103579650724 (python), and dispatch 103579992970 succeeded.
  • Rerun of Required OpenCode Review 34695759446 completed bootstrap, admission, coverage-source-tree and coverage-evidence successfully.
  • Job 103581933696 at 2026-09-12 16:11:58.248864 UTC logged Current-head substantive OpenCode verdict already exists; scheduler wake skipped. It then reported CHANGES_REQUESTED while succeeding as a job.
  • The final opencode-review failure still points to old receiver run 34698522096. A successful caller job is not a new review or approval.

Required repair and acceptance

Repair the canonical review admission predicate. Distinguish actual substantive code findings from deterministic peer-check or model-unavailable fallback outcomes. Permit governed same-head reevaluation when authenticated peer-check evidence changes. Preserve rejection of closed, draft, stale-head and unauthorized requests; do not dismiss reviews, fabricate approval, add paid fallback or weaken merge protection.

Add a regression covering a peer-check-only CHANGES_REQUESTED review followed by successful same-head peer checks, plus an actual substantive finding that must retain its deduplication behavior. Verify the new receiver run and formal current-head review on #2113. No identical caller rerun is useful before the predicate is repaired.

This blocks the MLLO cross-organization admission lane and subsequent protected-main deployment. The central owner task has been notified; coordinate its writer before editing. This issue is diagnosis, not a claim of repair or deployment.

Activity

  1. self-assigned this
    on Oct 4, 2026
  2. seonghobae commented on Oct 4, 2026

    @seonghobae
    ContributorAuthor

    Hermes Coordinator pickup 2026-10-04 18:54 KST: existing PR#2126 assigned seonghobae; no competing2126 worktree writer found. Isolated existinglane /Users/seonghobae/orca/workspaces/dot-github/hermes-pr2126 ordinarily merge-forwarded protectedmain37b10243 with local b00e91d (not pushed). Scope: shared exact historical coverage-fallback classifier consumed by receipt+scheduler; trust author/head/complete fixed producer envelope/run-attempt identity; unknown/mixed/source findings remain BLOCK, newer fallback cannot revive oldapproval, failedpeers never merge. Actual consumer specimen BandScope#1176 review5301203359/run35908696823; deployed Cargo repair#2360 is distinct. Real central scheduler37192132705 reproduces BLOCK. Project#1 issue item added/setInProgress. Actorallowlists/permissions/credentials/protection unchanged; no duplicate PR, no direct personal privileged dispatch, no fake status/dismissal. Disjoint owner brief sent to live central Orca session; lack of response is not acceptance. Remotehead recheck and independent review before any existinglane push.

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

Metadata

Metadata

Assignees

Labels

bugSomething isn't workingpriority: highHigh-priority or P1 work

Type

No type

Projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions