Audit your own agent and skill definitions for loop-reliability defects

The after-the-fact half of authoring: a fixed-category reliability sweep, a read-only confirmation pass, and a check against the canonical tools/description spec — never widening a grant to fix a finding about that same grant.

Also usesclaude-plugins-official

This is the after-the-fact half of authoring, run against skills and agents that already exist. confab:confab-agentic-reliability sweeps four fixed categories with verification on by default; confab:agentic-reliability-auditor is deliberately read-only, so it cannot paper over an excessive-tool-grant finding by widening its own grant. plugin-dev:agent-development supplies the canonical tools:/description: spec a finding of "excessive" is judged against, and andon:andon-verify closes the gap between a guard that exists in prose and a guard that actually runs.

Start here

One prompt to start

Sweep our own skills, agents, and workflows for agentic-reliability defects -- unbounded loops, missing escalation, no verify-wiring, tool grants beyond the stated role -- confirm any excessive-grant finding with a strictly read-only auditor, judge it against the canonical tools/description spec, and then prove any fix actually holds rather than just existing in prose.

Needsconfab, plugin-dev, andon

Beats

Run these in order. Each prompt is copy-pasteable straight into Claude Code.

1. confab:confab-agentic-reliabilityalso in 1 other recipe

   • Author a new skill or agent for a plugin

   Four fixed categories — unbounded retries, absent escalation, missing verify-wiring, excessive tool grants — run with verification on by default.

   Run this beat on its own

   sweep our skills, agents, and workflows for agentic-reliability defects — unbounded loops, no escalation path, no verify wiring, tool grants beyond the stated role

2. confab:agentic-reliability-auditor

   Read-only by its own tool grant (Read, Glob, Grep only) — it cannot 'fix' an excessive-tool-grant finding by widening its own grant, which is the exact failure mode this beat exists to avoid.

3. plugin-dev:agent-development

   Supplies the canonical tools:/description: spec the audit is judged against — without it, 'excessive' has no baseline to be excessive relative to.

4. andon:andon-verifyalso in 13 other recipes

   • Audit supply-chain pinning and dependencies
   • Finish and integrate a development branch
   • Investigate a job that reports success but changed nothing
   • Investigate a fix that did not stick
   • Investigate tests that pass while the code is broken
   • Root-cause a bugfix from a tracked issue
   • Build a new feature in an existing codebase
   • Collapse duplication hand-synced across N sites
   • Execute a written plan across parallel workstreams
   • Migrate a return shape or type representation
   • Propagate a vendored artifact to N copies
   • Do oracle engineering and numerical V&V
   • Whole-branch review without re-trusting the branch's own self-assessment

   'The guard exists' and 'the guard runs' are different claims — the distinction this repo has been burned by repeatedly, and the reason a reliability finding needs proof, not just a report.

Worked example

Grounded in — why this beat order is trustworthy

The unresolved self-assess:arch-health-auditor tool-grant anomaly recorded in this repo's own CLAUDE.md — it reports only {Read, Bash} against three different declared tools: formats tried, including in a fresh claude --print process. A live, currently-open case where a declared grant and the runtime grant disagree and no existing beat in this catalog would have caught it before this recipe existed.

Do / Don't

• ✓Sweep the four fixed reliability categories with verification on by default, not just a Find pass.
• ✓Confirm an excessive-tool-grant finding with a strictly read-only auditor -- one that can't paper over the finding by widening its own grant.
• ✓Judge any 'excessive' tool grant against the canonical tools:/description: spec -- without a baseline, excessive has nothing to be excessive relative to.
• ✓Prove a fix actually runs with andon-verify -- 'the guard exists' and 'the guard runs' are different claims this repo has been burned by conflating.

• ×Don't let an agent fix its own excessive-tool-grant finding by widening its own grant -- that's the exact failure mode a read-only auditor exists to avoid.
• ×Don't judge a tool grant as excessive without the canonical spec as a baseline.
• ×Don't treat a reliability finding as resolved just because a guard was written -- confirm it actually runs, the way this repo's own unresolved arch-health-auditor tool-grant anomaly still hasn't been.