stack/agents/darkwing/work/queue-116/r2/build.md

Row 116 (#1610) round 2: a resolved merge decision lands as a gated local commit

Brief: docs/plans/2026-10-14_merge-path-for-approved-work.md (blob 7a610f379). Round 1's packet is agents/darkwing/work/queue-116/r1/ (82943bf61). Filbert's round 1 verdict was changes, comment 28464, queue rev 1165, for C1. Sage's sage-0207 asks for C1 as tests only and folds in D1. sage-0204 settles the base and tree question from round 1.

BASE: 1223a0f37, unchanged. No candidate path has changed on refactor from 1223a0f37 to a2ee3a80f. Over packages/bus, scripts/test-conductor.sh and the webui pinned-copy test, the only change is still 7eb5ad0cb's requirement count in packages/bus/tests/requirements.test.mjs. The candidate is build.patch against BASE, and candidate-manifest.sha256 lists the same 6 files. repro.sh exports BASE with git archive, applies build.patch and checks every manifest digest: DIGESTS_OK.

What changed in round 2

Two files changed. read.mjs, land-work.sh, fixture.mjs and CONDUCTOR.md are byte-identical to round 1, as Filbert ruled no change there; their digests in candidate-manifest.sha256 equal round 1's.

File	Round 2 change
scripts/test-land-work.sh	six cases for C1, and a canary file in the fixture target's node_modules. 84 cases, was 78
docs/TOOLS.md	D1: the usage line has --attribution, --business and --project, and the prose says what each does. The Maintenance row says 84 cases

C1: one case per condition

Each case is a gen.mjs variant built through fixture.mjs like the rest. Each mutant is listed with the round 2 mutants below and was run against the full suite.

Filbert	Condition in read.mjs	Case	Mutant	Mutant result
f1	the answer's actor_role is null	answer-role: the answer event has actor_role pm. Refuses not-human	r24-answer-role	killed, only this case fails
f1	the answer's actor_run is null	answer-run: the answer event has actor_run rfixture. Refuses not-human	r25-answer-run	killed, only this case fails
f2	the answer's choice equals the close's	answer-choice: the answer says no, the close says yes. Refuses not-human	r26-answer-choice	killed, only this case fails
f3	exactly one answer	two-answers: the same answer event twice. Refuses not-human	r27-one-answer	killed, only this case fails
f4	raise.task_ref agrees with the work ref	raise-task-disagrees: the raise cites vikunja:7/99, the ref is work/vikunja-7-42. Refuses work-ref	r28-raise-task-ref	killed, only this case fails
f5	later counts only later decisions	earlier-merge: a merge decision E on the same ref, raised before D and withdrawn. A dry run of D exits 0	r29-later-only	killed, only this case fails

I split f1 into two cases and two mutants because read.mjs checks actor_role and actor_run as separate conditions, and a mutant that drops one leaves the other in place. r27 changes answers.length === 1 to >= 1, the narrower form of round 1's r10, which also accepted no answer.

The earlier-merge case goes as far as Filbert asked. E is a merge decision with the same task_ref and question, and a withdrawn close. fixture.mjs writes it before D, so its seq is lower. The dry run passes every check and runs the gate, then writes no ref. With r29, later picks E up and the case refuses superseded.

The canary

Filbert's s26 ruling suggested it. The suite's target node_modules was an empty directory, so "the target's node_modules is still there" couldn't see its contents being deleted. Now it holds a file named canary, and the case also checks that the file is there. s27 is the mutant that shows the case can fail: cleanup does rm -rf "$WT/node_modules/", which follows the link and empties the target's directory. s26, which drops the explicit unlink, stays equivalent, as Filbert confirmed on git 2.55.0.

Tests

scripts/test-land-work.sh, 84 cases, all passing (log: test-land-work.log). The round 1 groups are unchanged except these:

Group	Round 2 addition
The decision	answer-role, answer-run, answer-choice and two-answers refuse not-human, and raise-task-disagrees refuses work-ref. 25 variants, was 20
A dry run	earlier-merge exits 0 with a withdrawn earlier merge decision on the same ref
Cleanup	the target's node_modules is still there, with its canary file

Mutants

mutants.py has 56 mutants, the 49 from round 1 plus r24 to r29 and s27. It applies each to the file, runs the whole suite under a 600 second timeout in its own process group, and restores the file from .orig with a sha256 check. I ran the seven new ones first, then all 56 against the 84-case suite. mutants-run.log has both runs and each mutant's failing cases. The full run ended restored: True, and 55 of 56 are killed. s26 survives, and Filbert ruled it equivalent in comment 28464. Four round 1 mutants now fail one more case than they did, each a new case: r10 also fails two-answers, r14 raise-task-disagrees, and r20 and s16 earlier-merge.

Mutant	Change	Result
r01-schema-digest	skip the schema digest compare	killed, 1 of 84 fail
r02-schema-meta	skip the meta version and digest check	killed, 1 of 84 fail
r03-read-only	open read-write	killed, 2 of 84 fail
r04-business-scope	drop business = ? from every query	killed, 1 of 84 fail
r05-one-raise	accept zero or several raises	killed, 1 of 84 fail
r06-not-a-merge	skip the action check	killed, 1 of 84 fail
r07-resolved-op	accept any single close op	killed, 1 of 84 fail
r08-human-by	skip resolved-by == human	killed, 1 of 84 fail
r09-human-via	skip via == cli	killed, 1 of 84 fail
r10-human-answer	accept a missing human answer	killed, 2 of 84 fail
r11-merge-option	skip choice == approvalChoice	killed, 1 of 84 fail
r12-one-work-ref	accept several work refs	killed, 1 of 84 fail
r13-work-ref-shape	loosen the work ref pattern	killed, 1 of 84 fail
r14-task-ref-agrees	skip the task_ref agreement	killed, 2 of 84 fail
r15-target	skip the target check	killed, 1 of 84 fail
r16-later-supersedes	ignore supersedes	killed, 1 of 84 fail
r17-later-task	ignore a later decision on the task	killed, 1 of 84 fail
r18-later-question	ignore a later question naming the ref	killed, 1 of 84 fail
r19-no-verdict	skip the no-verdict refusal	killed, 1 of 84 fail
r20-latest-verdict	take the earliest verdict	killed, 31 of 84 fail
r21-approve	skip verdict == approve	killed, 1 of 84 fail
r22-verdict-task	skip the verdict body's task_ref	killed, 1 of 84 fail
r23-full-oid	accept a short commit	killed, 1 of 84 fail
r24-answer-role	skip the answer's actor_role check	killed, 1 of 84 fail
r25-answer-run	skip the answer's actor_run check	killed, 1 of 84 fail
r26-answer-choice	skip the answer's choice == the close's	killed, 1 of 84 fail
r27-one-answer	accept several answers	killed, 1 of 84 fail
r28-raise-task-ref	drop the raise's task_ref from the agreement	killed, 1 of 84 fail
r29-later-only	later counts earlier decisions too	killed, 1 of 84 fail
s01-base-early	skip the early --base check	killed, 1 of 84 fail
s02-ref-moved	skip ref-moved	killed, 1 of 84 fail
s03-merge-base	ignore a missing merge base	killed, 1 of 84 fail
s04-empty-diff	accept an empty diff	killed, 1 of 84 fail
s05-root-top-file	allow a top-level file	killed, 1 of 84 fail
s06-root-new-dir	allow a new top-level directory	killed, 1 of 84 fail
s07-root-src-v1	allow src/ and v1/	killed, 1 of 84 fail
s08-dirty-changed	skip the changed-path check	killed, 3 of 84 fail
s09-dirty-staged	skip the staged check	killed, 1 of 84 fail
s10-dirty-untracked	skip the untracked check	killed, 1 of 84 fail
s11-dirty-recheck	skip the recheck before the commit	killed, 2 of 84 fail
s12-apply-conflict	ignore an apply failure	killed, 1 of 84 fail
s13-patch-drift	skip the patch-id compare	killed, 1 of 84 fail
s14-gate-node-modules	skip the node_modules check	killed, 1 of 84 fail
s15-gate-step-fails	ignore a failing gate step	killed, 5 of 84 fail
s16-gate-skips-queue	run the queue suite too	killed, 16 of 84 fail
s17-gate-scripts	run no scripts/test-*.sh	killed, 8 of 84 fail
s18-base-moved-cas	update-ref without the old value	killed, 2 of 84 fail
s19-dry-run	ignore --dry-run	killed, 3 of 84 fail
s20-checkout-sync	skip the checkout sync	killed, 1 of 84 fail
s21-subject	a fixed subject	killed, 1 of 84 fail
s22-attribution	drop the attribution line	killed, 2 of 84 fail
s23-cleanup-worktree	keep the worktree	killed, 6 of 84 fail
s24-cleanup-network	keep the Docker network	killed, 1 of 84 fail
s25-cleanup-scratch	keep the scratch directory	killed, 6 of 84 fail
s26-cleanup-link	skip the explicit node_modules unlink	survived, equivalent (Filbert, comment 28464)
s27-cleanup-follows-link	cleanup follows the link and empties the target's node_modules	killed, 1 of 84 fail

Gate

A targeted builder gate under #271 (gate.sh, logs in gate/), the same steps as round 1. Round 1 ran the webui pinned-copy test by hand beside gate.sh; this round gate.sh runs it as its fourth step. Sage's landing gate stays full.

Step	Result
packages/bus node --test tests/*.test.mjs	156 pass, 0 fail
scripts/test-land-work.sh	84 passed, 0 failed
scripts/test-conductor.sh	17 passed, 0 failed
packages/webui node --test tests/pinned-copy.test.mjs	10 pass, 0 fail

GATE_PASSED (gate-run.log). No step failed, so the flake rule had nothing to count. The suite also passed 84 of 84 on its own before the mutants (test-land-work.log), and s26, the one survivor, passed all 84 in the full mutant run.

Replay

Round 1's replay stands, and I didn't rerun it. It ran read.mjs and land-work.sh, which are byte-identical in this candidate, against scratch copies of the live rows. Filbert checked it under sage-0204: at 1c9a4910d the landed diff is f7c8537b3's own diff and the dry run wrote no ref, and at 4bc908991 the tree equals f7c8537b3^{tree}. The commit stayed local in both. The evidence is in agents/darkwing/work/queue-116/r1/ (build.md's Replay section, replay-1c9a/, replay-4bc9/).

Going live

Nothing goes live. The tool is a conductor command run by hand. No service, role or session calls it. Push, the merge to next or main, re-pins and restarts stay out of this row.