The new structured repo doc prompt (5 sections with headers) needs more
room than the old 4-6 sentence prompt. File and relationship docs stay
at 256 tokens. Plumbed max_tokens through _call_llm → _call_ollama/_call_openai.
Repo summaries now generate structured, opinionated overviews with:
- What This Project Does (purpose, not generic description)
- Architecture (entry points, request flow, key abstractions)
- Key Patterns (conventions an agent must match)
- Where Things Live (non-obvious file ownership map)
- Gotchas (surprises, race conditions, naming traps)
Addresses Mike/Dmitry feedback: roll-up summaries must be 'actually usable.'
Old prompt produced 4-6 generic sentences. New prompt produces a real
onboarding doc that reads like a senior engineer wrote it.
- find_path: BFS traversal to trace relationship chains between two files (max N hops)
- get_file_signatures: lightweight context mode returning just function/type names
- db.find_path(): bidirectional BFS with shortest-path tracking
- db.get_all_files(): list all files with docs and staleness
- db.get_file_signatures(): return functions list without full doc payload
Inspired by Octocode's GraphRAG path-finding pattern.
Addresses Mike/Dmitry feedback on usable roll-up summaries.
- Installs MCP config to ~/.claude/mcp.json (merges with existing)
- Appends knowledge graph instructions to ~/.claude/CLAUDE.md
- Run once, works in any repo after that
- LLM_BACKEND=openai routes to /v1/chat/completions
- Default: ollama (unchanged)
- For Kiro gateway: LLM_BACKEND=openai OPENAI_URL=http://192.168.86.11:8000 OPENAI_MODEL=claude-haiku-4
- Updated README with new env vars
- Schema versioned with migrations table
- watch.sh runs db migrations on every pull
- watch.sh uses uv run/sync instead of pip
- Fixed parser.py reference in watch.sh
- parser.py renamed to go_parser.py (avoids Python builtin conflict)
- docgen.py was missing from flat structure
- Added pyproject.toml for uv
- Updated .mcp.json to use uv run
- Updated README for uv workflow
- SQLite backend with file/repo/relationship entities
- tree-sitter Go AST parser for deterministic import detection
- Ollama doc generation with concurrent batch processing
- MCP server (FastMCP) for Claude Code integration
- Merge simulation with staleness cascade
- Lazy refresh for stale relationship and repo docs
- CLAUDE.md for agent context
