agent-audit

---

name: agent-audit description: Audit Claude Code agent definitions for craft quality: delegation triggers, tool and model fit, context economy, reliability contracts. Reviews the agent .md files themselves, not the overall setup. disable-model-invocation: true

Craft review of Claude Code agent definitions. Scope boundary: /cc-review asks "should this agent exist, and is an agent the right primitive?" This skill asks "given it exists, is the definition well-crafted?". Do not re-litigate existence or primitive choice here; hand those findings to /cc-review instead.

Advisory-first: produce the report, then offer to apply fixes. Never edit before the user accepts findings.

Step 0: Refresh knowledge (every run)

WebFetch https://code.claude.com/docs/en/sub-agents and confirm the currently honored frontmatter fields and their valid values. Training-data knowledge of agent frontmatter is stale. Never flag or recommend a field you did not just verify. If the fetch fails, say so and limit frontmatter findings to fields observed working in the repo.

Step 1: Discover and classify

1. Collect agent files from project .claude/agents/ and global ~/.claude/agents/. When run inside the dotfiles repo, use the source of truth instead: roles/ai/files/claude/agents/ plus roles/ai/files/claude/agent-registry.json.
2. Classify each agent:
   • local: under registry local_agents, or no registry entry. Fixes may be applied.
   • upstream-synced: under a registry repos entry. REPORT-ONLY: in-place edits are silently reverted by claude-agent update. Remedies to offer: contribute the fix upstream, or fork to local (copy the file, move its registry entry to local_agents).
3. Note registry ↔ disk drift (registered but missing, on disk but unregistered) as findings.

Step 2: Score each agent

Dimension	Check
Delegation surface	description is written for the router: when to delegate, when NOT to, non-goals naming the sibling agent that owns that seat; "Use PROACTIVELY" only where auto-delegation is truly wanted
Cost: model fit	model: matches the judgment intensity of the work; flag heavy tiers on mechanical tasks and a missing model: where inheriting the caller's model is wrong
Context economy	tools: is a minimal allowlist (omission = inherit everything); every prompt section earns its tokens; a bounded output contract caps what flows back to the caller
Reliability	Bounded self-correction (retry limit, then stop), explicit failure statuses (blocked, needs-decision), honest-reporting rules (never claim an unrun check)
Latency	Independent lookups prompted as parallel tool calls; no prescribed re-reads of content the agent already holds
Frontmatter validity	Only fields verified in Step 0, with valid values; flag silently-ignored fields
Registry hygiene	groups tags come from the controlled vocabulary in the repo CLAUDE.md; drift findings from Step 1

Step 3: Cross-agent checks

• Trigger collisions: two descriptions that fire on the same delegation phrasing.
• Seat overlap: two agents claiming the same responsibility without naming each other as the non-goal.
• Fleet consistency: model tiers and output contracts that differ without a reason.

Step 4: Report

1. One-line fleet health summary, then what each agent already does well. Do not manufacture findings to look thorough.
2. Findings prioritized P0-P2. Each: What (one line, quoting the exact file and frontmatter key or section) / Why / How (the exact change).
3. Per-agent verdict table: agent | local or upstream-synced | top finding | fix path (apply / fork-to-local / contribute upstream).
4. Offer to apply accepted fixes: local agents only, one agent at a time.