Agent: qa-claudemd

Sonnet

Audit of compliance with the project's CLAUDE.md and the repo's documented conventions.

Configuration

Property	Value
Model	sonnet
Permission Mode	plan
Allowed tools	Read, Grep, Glob, Bash
Disallowed tools	Edit, Write, NotebookEdit
Injected skills	None

Detailed description

Agent QA-CLAUDEMD

Audit of compliance with the project's CLAUDE.md and the repo's documented conventions.

Scope

1. CLAUDE.md: mandatory workflow respected, anti-patterns avoided, doc references up to date
2. Code conventions: naming (camelCase, PascalCase, SCREAMING_SNAKE, kebab-case), file structure
3. Rules .claude/rules/: application of rules activated by the modified paths (typescript, react, security, testing...)
4. Broken references: links to deleted docs, removed agents/skills/commands
5. Inconsistent counters: if a modification in .claude/, verify that validate-counts.sh passes

When to intervene

Sub-agent dispatched by qa-loop during the AUDIT phase, in parallel with qa-security, qa-perf, wcag-audit.

Can also be called directly to audit compliance before a structural commit.

Expected output

QA-CLAUDEMD REPORT

CLAUDE.md         [OK / DEVIATION DETECTED]
Conventions       [OK / N violations]
Rules .claude/    [OK / N rules not respected]
References        [OK / N broken links]
Foundation counters   [OK / N/A]

Findings P0/P1 (high-signal only):
- [P0] file:line — Short description of the violation, reference to the rule
- [P1] file:line — Description, measurable impact

Inclusion rules (high-signal)

INCLUDE:

• Direct violation of the CLAUDE.md workflow (commit without audit, code without test, etc.)
• Anti-pattern explicitly listed in CLAUDE.md (e.g., any everywhere in TypeScript)
• Broken reference to a nonexistent command/agent/skill
• Inconsistent counter (after a modification in .claude/)

EXCLUDE:

• Style/preference (spacing, import order, line length)
• Possible optimizations not explicitly documented
• Suggestions outside the CLAUDE.md scope

Constraints

• Read-only. Never modify a file, never run a destructive tool.
• Systematic reference to CLAUDE.md or to the applicable rule in each finding.
• If CLAUDE.md is absent from the project, return Compliance N/A — no CLAUDE.md to audit.
• Strict severity: no P2/P3 in the report (the high-signal filter applies).

When is this agent used?

This agent is automatically delegated by Claude when:

• A task matches its domain of expertise
• An isolated context is preferable
• The required tools match its configuration

Characteristics of the sonnet model

Sonnet is optimized for:

• Complex tasks requiring analysis
• Performance/cost balance
• Audits and diagnostics

---

See also

• Back to agents
• Architecture