Found in the adversarial review of #120.
Step 5 fans out to about 24 auditor and skeptic agents with no canary, against the canary-first rule
Where: .claude/skills/update-xmech-page/SKILL.md:137
Step 5 recommends 'one read-only auditor per Mech, one for fleet-wide claims, one for cross-references, and an independent skeptic per auditor who re-derives each proposed change'. That is 12 auditors plus 12 skeptics, and the PR describes a second round on top. This is the most expensive part of the run. The skill applies canary-first only to the pipeline scans (step 4). It never says to run one auditor and one skeptic end to end and check their output before fanning out, as the global rule requires for any costly batch.
Failure scenario. A shared defect, such as the wrong root URL for a meta-refresh site, a prompt that lets an auditor write files, or an output format the editor cannot merge, gets repeated across all 24 agents before anyone notices. The whole fan-out's cost is spent, and the rerun costs the same again.
Evidence. SKILL.md:137-141. The global CLAUDE.md rule: 'Never fan out to a batch of expensive jobs without first running exactly one of them, end to end, and verifying it worked.' The PR body: 'Twelve read-only auditors … An independent skeptic re-derived every proposed change, and a second round verified the 60 things the first round missed.' It mentions no canary for this stage.
Verifier (confirmed, low). The finding reproduces. This PR adds the skill. Its step 5 recommends fanning out to 12 read-only auditors plus 12 skeptics, and the PR describes a second round on top of that. That is the most token-expensive stage of the procedure. The skill spells out canary-first only for the pipeline scans in step 4. Step 5 never says to run one auditor and skeptic end to end and check their output and evidence format before fanning out. The heading "Pipeline, canary first" sits only on step 4, so a reader can take the omission in step 5 to mean no canary is needed there. The PR body reports no canary for this stage either.
Severity stays low for three reasons:
- The global CLAUDE.md canary rule still binds any session that runs the skill, so the gap does not override it. It only fails to reinforce it at the point where it matters most.
- The damage is wasted agent cost and a possible rerun. It does not put wrong content on the site, because the skeptics and the single-editor step still filter the changes.
- Some parts of the failure scenario are speculative, such as a prompt that lets an auditor write files. The skill does say the auditors are read-only.
The suggested fix is correct and small: add a canary sentence to step 5 that runs one auditor and its skeptic on a small Mech, checks the claims and evidence format, then fans out.
Suggested fix. Add: run one auditor and its skeptic on one small Mech first, check the claims it produced and its evidence format, then fan out.
Found in the adversarial review of #120.
Step 5 fans out to about 24 auditor and skeptic agents with no canary, against the canary-first rule
Where:
.claude/skills/update-xmech-page/SKILL.md:137Step 5 recommends 'one read-only auditor per Mech, one for fleet-wide claims, one for cross-references, and an independent skeptic per auditor who re-derives each proposed change'. That is 12 auditors plus 12 skeptics, and the PR describes a second round on top. This is the most expensive part of the run. The skill applies canary-first only to the pipeline scans (step 4). It never says to run one auditor and one skeptic end to end and check their output before fanning out, as the global rule requires for any costly batch.
Failure scenario. A shared defect, such as the wrong root URL for a meta-refresh site, a prompt that lets an auditor write files, or an output format the editor cannot merge, gets repeated across all 24 agents before anyone notices. The whole fan-out's cost is spent, and the rerun costs the same again.
Evidence. SKILL.md:137-141. The global CLAUDE.md rule: 'Never fan out to a batch of expensive jobs without first running exactly one of them, end to end, and verifying it worked.' The PR body: 'Twelve read-only auditors … An independent skeptic re-derived every proposed change, and a second round verified the 60 things the first round missed.' It mentions no canary for this stage.
Verifier (confirmed, low). The finding reproduces. This PR adds the skill. Its step 5 recommends fanning out to 12 read-only auditors plus 12 skeptics, and the PR describes a second round on top of that. That is the most token-expensive stage of the procedure. The skill spells out canary-first only for the pipeline scans in step 4. Step 5 never says to run one auditor and skeptic end to end and check their output and evidence format before fanning out. The heading "Pipeline, canary first" sits only on step 4, so a reader can take the omission in step 5 to mean no canary is needed there. The PR body reports no canary for this stage either.
Severity stays low for three reasons:
The suggested fix is correct and small: add a canary sentence to step 5 that runs one auditor and its skeptic on a small Mech, checks the claims and evidence format, then fans out.
Suggested fix. Add: run one auditor and its skeptic on one small Mech first, check the claims it produced and its evidence format, then fan out.