auto-coder-v2

Scaffold for a local-first autonomous coding agent powered by Ollama and MCP tooling.

Getting Started

1. Install dependencies (the CLI uses npm workbench tooling):

   npm install

   Optional but recommended: Install Ollama and pull the model referenced by coder.config.json (defaults to qwen2.5-coder:14b).
2. Run the CLI in dev mode:

   npm run dev -- --help

3. Build the distributable bundle when ready:

   npm run build

4. Execute the unit test suite:

   npm test

5. Run the end-to-end demo task (plans, applies diff, runs tests):

   ./scripts/demo.sh

   Requires a clean git working tree because the agent creates a dedicated branch.
   • Pass --use-llm to skip the prebuilt diff and ask the configured Ollama model to generate one on the fly.

The current CLI wires up plan, show, run, and resume with lightweight helpers (task generation, persistence, and state tracking). MCP-backed execution still needs to be filled in per the spec.md checklist.

Current Limitation

• LLM-generated diffs for documentation tasks (README.md, AGENTS.md) are not yet reliable; the agent rejects malformed patches until we integrate a more instruction-following model.

Project Structure

• src/cli – Command-line entry point (Commander-based surface).
• src/core – Task planning, persistence, run-state helpers, and git orchestration.
• src/daemon – Placeholder for the MCP JSON-RPC server.
• src/utils – Shared filesystem/id utilities.
• coder.config.json – Default configuration values referenced by future features.
• .coder/ – Runtime state, logs (.jsonl), and task/spec artifacts generated by the CLI.
• optional .coder/patches/<taskId>/<stepId>.diff files can be dropped in to feed edit/doc steps. The demo script copies examples/sample-task/patches/s2.diff into place by default; use --use-llm to rely on the Ollama model instead.

Refer to AGENTS.md and spec.md for contributor guidance and the detailed implementation roadmap.