queen: a claim follows the work as it stands now -- a send-back under the ceiling and an accept whose PR closed unmerged free their issue #9111

Part of #9021 (jam removal, lane B: the claimed and completed skips). The diagnosis is below; the rules extend specs/queen/holds.t27, the card that already decides what a dispatch row holds against its issue as it stands now (holds_files_now, see #9033).

Diagnosis

claimed, live Queen 2026-10-11 12:55Z

Sources: the public /queen/public-tasks and /queen/public-task?issue=N drawers, compared with the open-issue list. 89 open issues were claimed; the tick reported 87 at 12:53Z.

• Dead or vanished workers: 0. All running rows were on runners whose task lease was live. The existing reapers already release a silent runner (10 min), an expired lease, and a stalled row (120 min).
• The send-back claim says "the same bee is expected back on it". No bee comes back.
  • A runner deletes the bee's session when its turn ends; the runner logs show "Agent disposed" and "Session deleted" right after "turn finished".
  • Nothing in the runtime resumes a send-back.
  • So the issue sits idle for an hour, and its own files (HS_SEND_BACK, 120 min) then shut its next attempt out from minute 60 to 120.
  • 8 of the 9 were under the retry ceiling.

completed, the 374 issues of 12:47Z

The PR state is that of branch queen-<issue>, read once per 100 issues with a batched GraphQL query.

Why an accepted dispatch has no PR (all 213):

• every one has a remote queen-<issue> branch;
• for 206, the remote head is not the head the Queen accepted (judgedHead on /queen/public-board);
• in 10 of 10 sampled, that accepted commit does not exist on GitHub, and gh pr list --state all --head queen-N is empty.

The accepted heads are still in the Queen's export, and queen-export-push.yml has failed on every scheduled run since 2026-10-08 06:15Z ("could not reach the export: The read operation timed out", 14 runs). tools/queen/publish.py then refuses each branch ("accepted at X, but the branch is at Y"). The other 7 are its named refusals: a workflow file, writes outside the Boundary, or conflicts.

Rules (holds.t27)

• send_back_claim(send_backs, ceiling): under the ceiling a send-back claims nothing (CL_FREE); at the ceiling it stays live for the review valve.
• accept_claim(read, open, merged, closed_unmerged): an accept whose branch has only PRs closed unmerged is freed (CL_FREE).
  • Merged or open keeps it done.
  • No PR keeps it done: the work waits in the export, and freeing it would discard accepted work.
  • A list that could not be read frees nothing.
• held_stage(claim, stage): a freed claim holds no files (HS_DONE), so the next attempt is not shut out by its own row.

The runtime wiring (trios, base queen) goes behind TRIOS_QUEEN_RELEASE_DEAD_CLAIMS and TRIOS_QUEEN_FREE_CLOSED_PR, both off by default.

Success Criteria

• t27c test-report specs/queen/holds.t27: 0 FAIL, 0 vacuous. The new tests name send_back_claim, accept_claim and held_stage and carry runtime asserts.
• Each new rule, broken once, makes test-report FAIL.
• Replayed on the 12:55Z claims and the 12:47Z completed set:
  • send_back_claim frees 8 and leaves the other 81 alone;
  • accept_claim frees 2 and leaves 121 open-PR, 2 merged-PR and 213 no-PR issues alone.

Filed with no Boundary section on purpose: the change is in hand (lane B), so no bee should be dispatched on it.

🤖 Generated with Claude Code

User Scenarios

• Given a dispatch row in sendBack state whose issue is under the retry ceiling, when the Queen evaluates send_back_claim, then the claim is freed (CL_FREE) so the issue can be re-dispatched without waiting for the idle floor.
• Given a dispatch row in completed state whose branch has only PRs closed unmerged, when the Queen evaluates accept_claim, then the claim is freed (CL_FREE) because the accepted work was never merged and the branch holds no open or merged PR.
• Given a freed claim from either rule above, when held_stage is evaluated, then it returns HS_DONE so the next attempt is not shut out by the previous row's hold files.

Requirements

• FR-001: send_back_claim(send_backs, ceiling) MUST return CL_FREE for send-back rows under the ceiling and CL_LIVE for rows at or above the ceiling.
• FR-002: accept_claim(read, open, merged, closed_unmerged) MUST return CL_FREE when read is true, merged is false, open is false, and closed_unmerged is non-empty; MUST return CL_LIVE when merged is true or open is true; MUST return CL_LIVE when read is true and no PRs exist; MUST return CL_LIVE (free nothing) when read is false.
• FR-003: held_stage(claim, stage) MUST return HS_DONE when claim is CL_FREE, regardless of stage.
• FR-004: The new rules MUST be added to specs/queen/holds.t27 and covered by tests that exercise each branch with runtime asserts.
• FR-005: The runtime wiring MUST be gated behind TRIOS_QUEEN_RELEASE_DEAD_CLAIMS (for send-back) and TRIOS_QUEEN_FREE_CLOSED_PR (for accept), both off by default.

Success Criteria

• t27c test-report specs/queen/holds.t27 exits with exit 0, 0 FAIL, 0 vacuous; new tests name send_back_claim, accept_claim, and held_stage and include runtime asserts.
• Breaking any one new rule (e.g., flipping a CL_FREE to CL_LIVE in the spec) makes t27c test-report specs/queen/holds.t27 FAIL.
• Replay on the 12:55Z claimed set: send_back_claim frees exactly 8 of 9 send-back rows (the 8 under the ceiling) and leaves the other 81 claimed rows unchanged.
• Replay on the 12:47Z completed set: accept_claim frees exactly 2 issues (the two with only closed-unmerged PRs) and leaves 121 open-PR, 2 merged-PR, and 213 no-PR issues as CL_LIVE.

Boundary

• specs/queen/holds.t27