Files
stack/agents/darkwing/work/slice1-s2-review/comment.md
T
2026-10-04 23:13:35 -05:00

1.6 KiB

Row 37, S2, round 1: changes (Darkwing)

Candidate candidate-manifest.sha256 dbfd4a07…, 24 files under packages/bus/. Full record: agents/darkwing/work/slice1-s2-review/review-r1.md.

What holds on my machine: the patch applies at fef4b362 and adds only packages/bus/; the manifest checks 24/24; the schema is byte-identical to v3b; the tests pass 36/36 on Node 26 and on Node 24; the S1 contract check passes against S1 round 2; and all 11 mutants are killed against a clean baseline.

Required:

  • R1. The human proof refuses every real terminal here. It refuses when it can't read an ancestor's environ, and Jason's terminals descend from root-owned plasmalogin-helper. systemd --user and sshd have the same problem. Skip only the marker check for such an ancestor, and test with real /proc (pid 1).
  • R2. A revoked holder can claim its role again with its old capability, before the replacement. role.claim should refuse a run with a revoke row.
  • R3. An arbiter resolves its own cross-role decision: cto with task.scope.change via domain: technical, and pm with task.priority.change. Route to human when the arbiter is the raiser.
  • R4. A refused bindLaunch over IPC exits the broker with 2. Reply and keep running.
  • R5. mutations.py runs on a flattened copy where two tests fail before any mutation, so killed is always true. Keep the repository layout and compare against the baseline.

Eight notes, none blocking. Two of them: one approval authorizes the same action without limit (a lead question for Sage), and the per-commit timestamp scan costs 116 ms at 100k events.