- name
- optimize-agent-capabilities
- description
- Read-only audit of Skill, Plugin, MCP, hook, and rules visibility across coding-agent clients. Use when the user requests capability inventory, routing analysis, or a repeat audit after client or extension updates. Never change files or client configuration.
# Optimize Agent Capabilities
Treat the repository's L0-L4 documents and templates as the source of truth. This Skill is an
optional execution entry, not a second routing specification.
## Workflow
1. Run the read-only inventory:
```bash
node <skill-dir>/scripts/audit.mjs --json
```
2. Read only the matching profile in `profiles/`.
3. Classify every detected client:
- `audit-only`: report findings and a client-specific plan; do not edit client configuration.
- `unknown`: use the generic profile and do not infer configuration keys or paths.
4. Compare findings with the shared L0-L4 model and propose the smallest useful routing plan.
5. Report detected roots, counts, support boundaries, unsupported assumptions, and next steps.
## Safety
- This integration is read-only. Do not install, uninstall, enable, disable, or rewrite anything.
- Do not edit extension caches directly.
- Do not claim support from format similarity alone; client discovery and invocation controls differ.
- Preserve explicit invocation wherever the client supports it.
- Unknown clients receive a plan, not guessed mutations.
## Extending Clients
Add a profile first. A future PR may propose a write-capable adapter only when all five operations
are deterministic and tested: `detect`, `inventory`, `plan`, `apply`, and `verify`. See
`references/client-adapters.md`.
在 GitHub 查看