25 lines
4.5 KiB
Markdown
25 lines
4.5 KiB
Markdown
# SKILL-LAUNCHER-FILBERT-1 (#1498) — verdict
|
|
|
|
Reviewer: Filbert (independent). Author: Darkwing. Conflict check: none; I had no authorship or prior review role in the mismatch repair or the ms-watch replacement.
|
|
|
|
## Verdict: APPROVED
|
|
|
|
Exact candidate: the nine files of snapshot `/tmp/skill-launcher-review-5gh6v4v0` at manifest SHA-256 `c4dd2ad8ec87138fc756aa3e7a3366dcec32fa8c5d1ecfa3736f746457568904`, plus the explicit deletion of `skills/ms-agent-watch/SKILL.md`, against baseline `12ff5da7df91f5343a59999af98b67518c15d6a4`. All nine hashes verified before and after review; snapshot unchanged. The metadata receipt `docs/plans/reviews/2026-09-08_skill-launcher-review.json` is a separate unit and was not treated as candidate source.
|
|
|
|
## Independent verification
|
|
|
|
- Delta confinement: launcher changes exactly one line (skill list `ms-agent-watch` → `ms-watch`); the test file adds only the missing-skill refusal controls and the exact-path assertions; `ms-communications` and `ms-proactive-agent` change one reference line each (rename only); `ms-watch/SKILL.md` is new to git, matching the declared pre-existing on-disk replacement; the three log/status files gain only the recorded additions. The deletion target is a tracked baseline file, so it is a real git deletion, not an untracked-file cleanup.
|
|
- Fail-closed selection: the launcher's existing per-skill presence check now guards `ms-watch`. The new fixture control (inside the per-agent loop, so all three Pi agents) removes the fixture's `skills/ms-watch/SKILL.md`, asserts `--check` fails with `missing skill: ms-watch`, and asserts neither `.pi/state` nor a fake-Pi capture is created; the captured-args assertions verify the exact `--skill …/skills/ms-watch` argument and the absence of the old path, with the ten-skill count still enforced.
|
|
- My own disposable trees (never the drifted live tree):
|
|
- Overlay tree (`git archive 12ff5da7` + nine files + declared deletion): `node --test scripts/test-rocko-launch.mjs scripts/test-darkwing-launch.mjs` — 5 tests, 5 pass, 0 fail, exit 0.
|
|
- Mismatch reproduction (baseline tree with only `skills/ms-agent-watch/SKILL.md` removed, simulating the live drift that produced the owner goal): the Pi suite fails 3 of 4 with `missing skill: ms-agent-watch` for each agent — independently confirming both the coordinator's observed baseline failure and that the mismatch was real, not cosmetic.
|
|
- Replacement-skill accuracy, read against baseline `tools/agent-watch/agent-watch.sh`: explicit sender override via `MOSAIC_AGENT_SEND` with the unreviewed `$HOME/.config/mosaic` fallback warning; socket precedence `--socket` > `MOSAIC_TMUX_SOCKET` > unique discovery with ambiguity refusal; `list`/`status` described as side-effecting (stale notices, messages), matching the CLI's notice logic; user-bus refusal with status exit codes 0/3/6 and no health inference from failed queries; sender exits 0 and 2 treated as terminal with the legacy `delivered` wording explicitly qualified as non-acknowledgement; interval floor 10 s and defaults 30 s/3600 s with `--repeat`/`--quiet-timeout` all matching the CLI. The skill claims no activation and no reliability guarantees, states that loading installs no timers, and scopes live watch operations to assignment authority. Watch runtime behavior itself was not redesigned; historical stale statements elsewhere were not rewritten.
|
|
- Docs and scope: the mismatch plan, CURRENT.md next-action block, SESSIONS line, and BUILD-LOG entry are accurate; registry review is deferred, not canceled; other untracked skills, private/WUI records, and held work are excluded structurally (the candidate touches only its nine files plus the one deletion). Added lines carry no private identifiers; the UUID/home-path hits in the full log files are baseline history already published at `29c1defe` and later commits.
|
|
- `bash -n` and `node --check` pass. No live launcher, timer/systemd command, watch call, model invocation, private read, source edit, or git write was performed by this review.
|
|
|
|
## Conditions
|
|
|
|
1. Commit exactly the nine reviewed files plus the declared `skills/ms-agent-watch/SKILL.md` deletion on `refactor` from this approval; any other byte change requires re-review. Push per the owner goal after commit; remote identity verified fresh afterward.
|
|
2. Register the shared-log entries and this verdict after the pin lifts, as with the #1497 closeout.
|
|
3. The goal is reviewed commit/push of this repair. No live activation, timer arming, restart, or deployment is authorized or implied; operator acceptance of the unrelated skill work is not.
|