audit-agents

name	audit-agents
description	Audit and maintain agent configuration files (.claude/agents, AGENTS.md, CLAUDE.md, agent-memory). Checks for duplicates, outdated info, naming conflicts, wrong technical facts, and files over 200 lines. Fixes all issues and commits.

Audit & Maintain Agent Configuration Files

You are performing a systematic audit and maintenance pass on all agent configuration files. Apply all fixes directly — do not just report issues without fixing them.

Files to Audit

Agent system prompts:

• .claude/agents/*.md — one per agent role

Agent memory files:

• .claude/agent-memory/*/MEMORY.md — persistent cross-session memory per agent

Project guidelines (AGENTS.md / CLAUDE.md):

• AGENTS.md (root)
• frontend/AGENTS.md
• backend/AGENTS.md
• frontend/tests/e2e/AGENTS.md
• frontend/tests/CLAUDE.md
• Any other AGENTS.md or CLAUDE.md files in the repo (Glob **/AGENTS.md and Glob **/CLAUDE.md)

---

Audit Checklist

1. Naming conflicts

• Read frontmatter of every .claude/agents/*.md
• Each name: value must be unique. If two files share a name, the duplicate (less complete one) must be removed or renamed.

2. Stale/wrong technical facts

Compare documented facts against actual source files:

• SSE streaming format: read backend/src/main/java/com/sinsay/service/SseStreamEncoder.java or ChatService.java to confirm the actual event format (text/event-stream, Vercel UI Message Stream {"type":"text-delta"}) and check it matches what agent files say.
• Package names, class names, file paths: spot-check against real files using Glob/Grep.
• Tech stack versions: check backend/pom.xml and frontend/package.json.
• MCP tools referenced (e.g., IntelliJ MCP): verify they are listed in the agent's frontmatter mcpServers. Remove references to MCPs that are not configured.

3. Outdated boilerplate in agent files

• Remove any "MEMORY.md is currently empty" statement from agent system prompts. The memory system auto-injects the real MEMORY.md content — this placeholder is always wrong and misleading.
• The persistent memory instructions (Guidelines, What to save, What NOT to save, Explicit user requests) are already part of Claude Code's built-in behavior when memory: project is set. Do NOT duplicate this ~35-line block in every agent file. Remove it from agent files.

4. Duplicate information across files

• If the same information appears verbatim (or near-verbatim) in both a root AGENTS.md and an agent system prompt, keep it in one place only:
  • Agent system prompts should hold agent-specific workflow and conventions.
  • AGENTS.md files should hold file/location facts, tech stack, and API contracts.
• Do not repeat SSE format specs in both AGENTS.md and agent files if they say the same thing.
• Do not repeat verification commands in both root AGENTS.md and agent files if they are identical.

5. Information already in Claude's built-in system prompt

The following is provided by Claude Code itself and should NOT be repeated in agent files:

• Generic commit rules ("one logical change per commit", "do not push unless asked")
• Generic TDD cycle description (write → fail → implement → pass → refactor)
• Generic Context7 usage instructions (resolve-library-id + query-docs)
• Generic git conventions

Keep only project-specific variants (e.g., commit message format Backend: summary).

6. Missing critical context

Every agent file should contain the minimum context to start working without reading other docs:

• be-developer: package root, SSE format, key service names, how to run tests
• fe-developer: component structure, Chat integration pattern, Vite proxy/build config, Vitest setup
• qa-engineer: test locations, stack (real backend, no mocks), how to run Playwright, image paths

If an agent file is only memory boilerplate with no project context, add the essential context.

7. File length

• Agent system prompts and AGENTS.md files should be under 200 lines.
• If a file exceeds 200 lines, remove duplication and boilerplate first. If still too long, extract secondary detail to a linked file (e.g., docs/ADR/) and reference it.

8. AGENTS.md scope

• Root AGENTS.md: project overview, commands, global workflow rules.
• backend/AGENTS.md: backend-only facts (tech stack, package structure, API contracts, SSE format, testing patterns).
• frontend/AGENTS.md: frontend-only facts (component structure, chat integration, Vite config, testing).
• frontend/tests/e2e/AGENTS.md: E2E-only facts (Playwright setup, image loading, no-mock rule, QA workflow phases).
• Each file should be self-contained for its scope — an agent working in that directory should not need to read parent files for routine tasks.

9. Orphaned memory files

• Check that each directory under .claude/agent-memory/ has a corresponding .claude/agents/*.md file.
• If memory exists in an unexpected location (e.g., backend/.claude/agent-memory/), determine whether it's the correct location or a duplicate. The canonical location is {project-root}/.claude/agent-memory/{agent-name}/MEMORY.md.

---

Fix Strategy

1. Read all files listed above in parallel.
2. Build a list of all issues found with file + line reference.
3. For each issue, apply the fix directly using Edit or Write tools.
4. After all fixes: re-read modified files and confirm no new issues were introduced.
5. Run a final line-count check on all modified files.
6. Commit: Docs: audit and update agent configuration files

---

Output Format

After completing all fixes, produce a concise summary:

## Agent Audit Summary

### Fixed
- [file] issue description → fix applied

### No Action Needed
- [file] reason

### Manual Review Required
- [file] issue that could not be auto-fixed