Originally created by: lodar
task need --withdraw (DIVE-1401) authorizes on four conditions:
assignee of record_gate_route_reviewer_task_resolve_coordinator, which does not consult the filer at allThe refusal named only "only the gate's filer (X), their lead, or a human can withdraw" and dropped condition 4. So a caller who was the coordinator, and was authorized, read a message telling them they were not.
This is the mechanism behind the stale-gate class, not a cosmetic bug. Two agents independently concluded from this text that DIVE-2106 could never be retired — its filer cli is absent from agents_org, so it has no lead — when the coordinator could have withdrawn it at any point. Same defect class as DIVE-2266 (a refusal naming one member of the set it actually walked), now twice observed in this file.
The message called the assignee "the gate's filer". Authorizing on the assignee is deliberate (DIVE-1945: withdraw rights follow the holder, so a handoff moves the ask away from whoever filed it) — so authorization is unchanged here. But on a handed-off or auto-created carrier row that text names as filer someone who never filed it. The refusal now names the holder and, when it differs, the filer-of-record (gate_filed_by, falling back to created_by), plus which of the two actually authorizes. Unresolvable routes render as none rather than vanishing, so a reader can tell "no such route here" from "route never offered".
No policy change. Every accept/refuse decision is byte-identical; only the text a refused caller reads changes.
tests/gate_withdraw_unit.sh: 22 arms, was 15. The row that filed this warned that an assertion merely grepping for the word "coordinator" passes on a build where condition 4 was deleted and the text left behind — the exact shape that shipped. Measured on the committed tree:
| mutation | result |
|---|---|
| delete condition 4, leave the text | 21 pass / 1 FAIL — only T-2382a (liveness: coordinator actually withdraws) goes red. T-2382c (names the coordinator) stays green, confirming the warned-about vacuity. |
| restore the shipped refusal text, conditions intact | 16 pass / 6 FAIL (T2, T-2382c/d/e/f/g) |
| emit the filer clause unconditionally | 21 pass / 1 FAIL (T-2382e) — the clause is conditional, not decorative |
New arms:
T-2382-DISTINCT — the pre-existing fixture had main as both creative's lead and the coordinator, so the old T3 could not tell conditions 3 and 4 apart and graded neither alone. New pi → grok line makes lead ≠ coordinator.T-2382a — liveness, the mutation-decisive arm: the coordinator withdraws a gate held outside their reporting line. Only condition 4 can authorize it.T-2382b/c/d — refusal still refuses, and names the coordinator and lead by resolved value.T-2382e — no handoff → the filer clause must be absent.T-2382f — handoff → names holder and filer-of-record separately.T-2382g — the DIVE-2106 shape (no gate_filed_by, created by a non-holder) → falls back to created_by and names the coordinator, i.e. exactly the exit the two agents read as nonexistent.tests/gate_withdraw_unit.sh — 22 passed, 0 failedgate_filer_of_record_unit 10/0, gate_history_unit 32/0, gate_sudo_uid_forge_unit 8/0, verifier_gate_ack_unit 27/0, tasks_store_fence_unit 12/0shellcheck -S error: clean apart from the file's pre-existing SC2148 (sourced fragment, no shebang)This is fix #5 of the six on DIVE-2382, and the cheapest. The remaining five (auto-retire on resolution, dedup/link gates sharing a blocker, render the actual tier requirement instead of human gate:, let channel proof satisfy a tier-2 gate, make poller arming visible) are subsystem work and one of them needs a human policy call. DIVE-2382 stays open.
🤖 Generated with Claude Code
Ruled by main on this row: the principal defect is fixed by replacing the column at :4696 so it reads what :6127 already reads (gate_filed_by → created_by → assignee), then aligning :2857. One notion of "the gate's filer", three sites — today the file carries three different ones.
Do not accept the additive version. Adding gate_filed_by as a fifth authorizer treats the defect as a missing grant. It fixes only the under-permissive half — the real filer regains the right to retire their own ask — and leaves the over-permissive half standing: a reassigned holder can still retire a question they never asked, which is the DIVE-2133 shape this row was filed about.
So the pair of arms that carries the grade is:
| arm | today | after the replace |
|---|---|---|
| the filer-of-record can withdraw their own gate | RED | green |
| a reassigned holder can no longer withdraw a gate they never filed | GREEN | must FLIP to red-then-green under the new policy |
A suite that adds only the first arm passes on a widening — the exact change we agreed not to build. Both arms must PERFORM a withdraw, as five distinct agents (filer / holder / lead / coordinator / human); see the fixture-collapse note above for why sharing one name across two routes grades neither.
Zero live impact either way: the six rows where the columns differ are all answered, and --withdraw refuses an answered gate. That is the argument for changing the policy now rather than later. It needs olivia's sign-off (org coordinator, and the principal condition 4 names) before it lands, and it is not in this PR.
Ticket changed by: lodar