stack/agents/filbert/work/queue-a2-review-r1-2026-09-27.md

Queue A2 review, round 1 (#1508, row 9)

Filbert, 2026-09-27. A1 round 1 review: queue-a1-review-r1-2026-09-26.md (sha256 e464be6c).

Verdict

Approved. The review covers build.patch sha256 dc0be7ce74aaaaaf43deebe68af439d1b2b80c77aad6538420e929de35e2f21f on base c9539baa, with build-manifest.sha256 782bcb62… and build.md ef911685…. I found no blocking finding. The notes below don't block, and none of them changes the candidate.

The two patches outside the candidate, queue-md.patch 2ca8f689… and tools-md.patch c53aea1f…, are also fine to apply. The migration map (queue-migration-map.md 012ddce9…) matches QUEUE.md. I think one of its eight choices needs a change; see n1.

I re-hashed every input after the review. All match the hashes Darkwing sent. The canonical working tree matches the manifest 21/21, HEAD is still c9539baa, .git has no mosaic-queue* file and no pre-commit hook, and core.hooksPath is unset.

What I checked

All of this ran in a scratch clone, /tmp/fqa2, with push disabled.

• Manifest and suites. The manifest checks 21/21 after git apply build.patch. node --test passes 122/122. scripts/test-queue.sh passes 24/24 before genesis (live checks skipped) and 26/26 after it.

• The source diff, file by file: io.mjs, lock.mjs, queue.mjs, store.mjs, cli.mjs, scripts/mosaic, the README and the tests.

• Carried-forward items. N5, N7, N8, N10, N11, N12, P1, P2 and P3 are implemented as ruled. N7 is a README line; the rest have tests.

  • N5: tmpfs is allowed only when allowTmpfs === true, and realIo has no such key.
  • N8: \ and < in a cell are refused, and | is escaped. The GFM cell-count test catches a missing escape.
  • N10: a narrowed closes stays narrowed when issues changes, and the receipt says so.
  • N12: an actor mismatch warns on success and on refusal.
  • P1: every log entry's op goes through checkCallerOpId.
  • P2: review is {rounds} only, and the review issue is the last round's.
  • P3: releaseOrWarn is used on both gate paths and in withLock.

• Dispatch. dispatch.test.mjs runs HEAD's scripts/mosaic, kept as the mosaic-pre-a2.sh fixture, next to the new one. It compares argv, cwd, environment and stdin for every non-queue verb, so the only change is the queue branch.

• Mutations. I wrote 12 mutants of my own, separate from Darkwing's, and the suite kills all 12. They cover:

  • the cell check on blockedReason and on brief anchors;
  • a cell with no | escape, and a cell helper that ignores its options;
  • closes intersected with the old issues instead of the new ones;
  • allowTmpfs present on realIo;
  • unlock dropping its warning, and a silent releaseOrWarn;
  • the actor warning firing when --by is absent;
  • the genesis entry's op skipping checkCallerOpId;
  • the review issue taken from the first round, not the last;
  • unquoted arguments in the queue dispatch.

  I didn't mutate the genesis claims or the legacyView slice. The dry run below checks both directly.

• The map against QUEUE.md. I checked all 30 rows against the blob c8e3d34e with my own script, not map-check.mjs:

  • piece and gate are verbatim, and the required flags match;
  • every brief anchor occurs exactly once in HEAD;
  • closes ⊆ issues on every row;
  • highWater is 30, and nothing is retired.

• The genesis dry run. I reran Darkwing's steps 0–8 myself on the candidate plus both patches, committed in the scratch clone.

  • map-check.mjs passes. The hook installs from HEAD (blob abdf14e7, mode 0755), and its canary passes.
  • genesis reports rev 0 genesis 30 rows. queue-commit.sh --genesis commits it.
  • verify --current and render --check pass.
  • legacyView equals the text between the two marker lines of the patched QUEUE.md, byte for byte. mapBlob is the map's blob in that commit.
  • Claims go to rows 5, 7, 9, 11 and 16, exactly the rows in in-progress, in-review or waiting-on-jason.
  • next resumes row 9 for darkwing, 7 for sage and 5 for dewey, and gives filbert, rocko and jason nothing.
  • assign 10 sage commits as rev 1.
  • A hand edit of queue.json staged for commit is refused by the hook, with the reset command in the message.

Non-blocking

• n1. Row 13 can close #1508 before Gate G passes (map choices 5 and 7). Row 13 is the only row that closes #1508. Its gate owner is sage, and it has no after. Under J5, Sage can move it in-review→done without Jason. Piece E's rule is that a done row's issue is closed. So #1508 would close while row 9, Jason's Gate G, is still in progress.
  • Row 13's own work has to finish first, so this isn't close. But the queue is meant to hold this order, not a seat's memory.
  • Fix it with either of these at genesis or just after:
    • after: [9, 10, 11, 12] on row 13, if after is meant to gate done as well as start; or
    • gate owner jason on row 13, or a gate text that says done needs rows 9–12 done.
  • Row 13's gate still reads "first run Monday 2026-09-21", a date that has passed. It's verbatim from QUEUE.md, and the note corrects it (Q7). Worth a set gate with a reason once genesis is in.
• n2. The priority paragraph leaves QUEUE.md. queue-md.patch puts the "Lead: … / Current priority" paragraphs inside the markers. They become part of legacyView and drop out of the file. That fits Q8. AGENTS.md's cadence still says "unless Jason has named an explicit current priority in that queue", though, and the ratified goal order now lives in docs/plans/2026-09-27_goals-review.md. I suggest the header above the markers point to the goals review. Row 10's AGENTS.md pass should reword that cadence line.
• n3. queue-commit.sh still calls node packages/queue/src/cli.mjs directly. A1 said that would change "until A2". Keeping it is right, since it keeps condition 4 simple. The README or the A2 build note should say it stays, so no one reads it as unfinished.
• n4. Row 9's note says "Filbert approved both". It is true once this verdict lands, and genesis runs only after that. No change needed.
• n5. SEMANTICS and VERSION are unchanged. That's correct: no log exists yet, so there is nothing to replay under older rules.
• n6. "(kept narrowed)" can appear when the kept closes now equals the new issues. The receipt is still true. Cosmetic.

I agree with the other map choices:

• row 7 stays (decision 27);
• row 10's owner is coordinator, with assign 10 sage right after genesis (tested in the dry run);
• rows 10 and 12 have no after;
• row 11 is waiting-on-jason, which is accurate: the queue add brief refusal (briefCheck) is already in A1;
• row 16 stays open;
• rows 20–22 are done per decision 29.

For Sage

The candidate is ready to commit. Then apply the two patches, install the hook, run genesis, commit with queue-commit.sh --genesis, and run assign 10 sage, in the order of the dry run. n1 is the one I'd settle at genesis time.