🐛 decision path: approve-with-rationale reads as changes_requested, Refs→Fixes rewrite overrides owner decision, 070 deadlock without a new head #14816

Observed 2026-10-28 while resolving the requires_human backlog through the hive's own decision path (dashboard POST /api/repos/{owner}/{repo}/items/{n}/decision, /hive decision comments, GitHub reviews). Four places where the owner's recorded decision did not produce the outcome the owner stated, with evidence. Each is a small, separable fix; none needed a manual merge or close.

1. /hive decision approve <rationale> on a hive-authored PR is treated as changes requested

unparkCommand.BareApproval() (pkg/github/issue_unpark_command.go) only recognises the approval when the comment is literally approve/lgtm/yes. A maintainer who writes /hive decision approve followed by a rationale — the natural way to answer a requires_human question — gets mode=changes_requested, the author agent is kicked with the rationale "verbatim", pushes a cosmetic head, and the swarm re-review + human approval start over.

• 🐛 fix(review): require sign-off on review-fix follow-up commits #14628: 🐛 fix(review): require sign-off on review-fix follow-up commits #14628 (comment) (mode=changes_requested on an approve + rationale; the "fix" was a body edit, no new head)
• 🐛 fix(automerge): bind unmarked owner approvals to the head's observed push time #14758: 🐛 fix(automerge): bind unmarked owner approvals to the head's observed push time #14758 (comment) → agent pushed docs-only 700a640d, invalidating the owner's APPROVED review on ab64a959

Proposed: treat a /hive decision whose first word is approve/approved/lgtm as an approval and the rest as the recorded rationale; only route to the fix lane when the text does not start with an approval word (or add an explicit /hive decision approve: / --reason form and document it in maintainer-commands.md).

2. The issue-link safeguard rewrote an owner-directed Refs #N back to Fixes #N

On #14628 the owner's decision item 1 was "change Closes #14605 to Refs #14605" (#14605 stays open for two further expectations). The agent applied it at 01:15Z; at 01:44Z EnsureIssueStagePRLink → rewriteRefsIssueLinkToClosing (pkg/github/pr_issue_link.go) rewrote it to Fixes #14605 and stamped <!-- hive:issue-link-fixed --> — see the PR's edit history (editor hivecommons-hive, 2026-10-28T01:44:05Z). Had the PR merged, #14605 would have auto-closed against the recorded human decision. bodySaysIssueStaysOpen did not match the wording "… remain open on the issue".

Proposed: never upgrade Refs→Fixes on a PR head that carries a recorded human decision / hive:decision-head marker; and treat a Refs written after a Closes (body edit history, or the agent's own "applied the decision" comment) as intentional. I restored Refs by hand (the marker now prevents a second rewrite).

3. 070-fixing deadlock when the recorded change request needs no new head

A changes-kind HumanChangeRequest on head H keeps HumanChangeRequestStage → 070 until a new head appears (humanChangesRequestedForHead falls back to the standing store entry even with no CHANGES_REQUESTED review). When the requested change is a body/metadata edit, no head ever appears: #14628 sat at stage/070-fixing + stage/stalled for ~2h after the agent reported done. Neither a GitHub APPROVED review on H nor a /hive decision approve comment can release it — the un-park sweep only reads /hive decision on PRs wearing stage/085 (issue_unpark_command.go). The only release was the dashboard direction API with text approve (which records an approval-kind entry that outranks the standing one): #14628 (comment)

Proposed: let a trusted APPROVED review on H (or the agent's "applied, no head change" completion) clear the standing change request; and read /hive decision approve on 070 as well as 085.

4. Trusted CHANGES_REQUESTED review on a contributor PR did not move it off 085 (crux)

hivecommons/crux#22: owner CHANGES_REQUESTED review on head 40de8a11 at 2026-10-28T00:23Z (hivecommons/crux#22 (review) — see reviews tab), re-stated at 02:37Z; at 03:40Z the PR still wears stage/085-needs-decision. Per TrustedHumanChangeRequested it should derive 072/returned-to-contributor and nudge the author. Possibly crux is outside the lane's trusted-merger config or the review is not read for contributor PRs there — worth a look.

What worked (for the record)

close-superseded via the decision API on contributor and hive PRs: #14702→#14695, #14795→#14791, #14779→#14772, #14794→#14788, #14803→#14807 — each closed with hive/superseded and an attributed comment, first try. direction: approve on hive PRs #14628/#14758 recorded mode=approve.