Learn Architecture

Scans directory structure, dependencies, and semantic patterns to update project knowledge.

Inputs

Workflow

1. Check Known Issues

• check_known_issues(tool_name="code_search", error_text="")

2. Detect Project & Persona

• Infer project from cwd via config; default persona "developer"

3. Scan Directory Structure

• Walk project path (max 3 levels), skip node_modules, pycache, venv
• Build tree of key dirs: src, lib, app, api, core, services, models, tests, etc.

4. Analyze Dependencies

• Read pyproject.toml, requirements.txt, package.json
• Extract top dependencies

5. Semantic Code Search (if use_vector_search)

• code_search("API endpoint route handler request response", project, limit=10) → api_patterns
• code_search("class model schema database table field", project, limit=10) → model_patterns
• code_search("exception error handling try except raise", project, limit=10) → error_patterns
• code_search("test fixture mock assert pytest unittest", project, limit=10) → test_patterns
• If focus: code_search("{focus} implementation pattern", project, limit=10) → focus_patterns

6. Update Knowledge

• knowledge_update(project, persona, section="architecture.key_modules", content=modules_list)
• knowledge_update(project, persona, section="architecture.dependencies", content=dependencies)

7. Build Result

Output markdown with:

• Key modules found
• Dependencies list
• Patterns discovered (API, models, errors, tests)
• Focus area results if specified

8. Error Handling

• If "index not found": learn_tool_fix("code_search", "index not found", "Vector index not created", "Run skill_run('bootstrap_knowledge')")

9. Log

• memory_session_log("Learned architecture for {project}", "Persona: {persona}")

Key MCP Tools

• code_search — semantic pattern discovery
• knowledge_update — write architecture to knowledge
• check_known_issues, learn_tool_fix — error handling
• memory_session_log — session logging