Coding Assistant

Hybrid coding assistant with a LangGraph orchestration loop, structured tool execution, and FastAPI control-plane API.

Prerequisites

• Python 3.12+
• Docker (for PostgreSQL and Redis)

Quick start

docker compose up -d
docker compose build sandbox-image
cp .env.example .env
# Set OPENAI_API_KEY, ANTHROPIC_API_KEY, and/or GEMINI_API_KEY in .env

pip install -e ".[dev]"
coding-assistant serve

API: http://127.0.0.1:8000/docs

Terminal chat

In one terminal, start the server. In another, run the interactive client:

coding-assistant chat
# optional: --workspace /path/to/repo --api-url http://127.0.0.1:8000

Slash commands: /help, /workspace <path>, /status, /exit. Each line you type starts an agent run on the server.

Create a run (HTTP)

curl -s -X POST http://127.0.0.1:8000/v1/runs \
  -H 'Content-Type: application/json' \
  -d '{"workspace_root": "/path/to/repo", "message": "List Python files in src"}'

Architecture

LLM → StructuredAction → Host Validation → Tool Execution

See docs/phase-1-core-runtime.md through docs/phase-10-mcp-ecosystem.md, and procedure.md.

Metrics: GET http://127.0.0.1:8000/metrics

Model routing

Planner calls use capability-based routing (routing.config.json). Preview:

Gemini (Google)

1. Set GEMINI_API_KEY in .env (from Google AI Studio).

2. Either use the Gemini-first routing file:

   ROUTING_CONFIG_PATH=routing.config.gemini.json coding-assistant serve

   or override every tier with one model:

   ROUTING_DEFAULT_MODEL=gemini/gemini-2.5-flash coding-assistant serve

   The default routing.config.json also lists gemini/gemini-2.0-flash in fallback_chain after OpenAI/Anthropic.

   For Gemini-only setups, use routing.config.gemini.json (Gemini-only fallback chain; no OpenAI attempts).

Supported LiteLLM ids include gemini/gemini-2.5-flash, gemini/gemini-2.0-flash, and gemini/text-embedding-004 (for EMBEDDING_MODEL).

curl -s 'http://127.0.0.1:8000/v1/routing/resolve?capability=planning&complexity=high'