Docker Compose: FastAPI Coding Agent + Langfuse

This project runs a FastAPI coding agent service (/health, /runs) and a local Langfuse observability stack.

1) Start from scratch

cp .env.example .env
docker compose up --build -d

Set the model API endpoint explicitly if your provider is not OpenAI:

OPENAI_BASE_URL=http://seasonsolt.myds.me:8888/v1 \
OPENAI_API_BASE=http://seasonsolt.myds.me:8888/v1 \
OPENAI_API_KEY=<your_key> \
DEEP_AGENT_MODEL=openai:gpt-4o-mini \
LANGFUSE_ENABLED=true \
LANGFUSE_PUBLIC_KEY=<lf_public_key> \
LANGFUSE_SECRET_KEY=<lf_secret_key> \
docker compose up -d --force-recreate coding-agent

2) Verify startup

docker compose ps
curl http://localhost:8000/health
curl -X POST http://localhost:8000/runs \
  -H "Content-Type: application/json" \
  -d '{"task":"List files in the workspace root."}'

Services:

• langfuse-web: http://localhost:3000
• coding-agent: http://localhost:8000
• langfuse-worker
• postgres
• clickhouse
• redis
• minio
• proxy (optional): route / to agent and /langfuse/ to Langfuse

3) Optional services to consider

• proxy (nginx): unify public entry (/ -> agent, /langfuse/ -> Langfuse) and easy TLS termination.
• monitoring: Prometheus + Grafana for host/container metrics.
• backup: scheduled snapshots of postgres / clickhouse / minio volumes.

Logs:

docker compose logs -f coding-agent

Skill Evaluation Lab

Run a sample skill evaluation:

mkdir -p /tmp/agent-lab-runs
docker compose up -d --build
python3 -m skill_lab.cli run \
  --skill ./skills/example-coding-skill \
  --task-pack ./task_packs/coding-basic \
  --api-url http://localhost:8000 \
  --network

Compare a treatment skill against a baseline skill:

python3 -m skill_lab.cli compare \
  --baseline ./skills/example-coding-skill \
  --treatment ./skills/ddia-system-design \
  --task-pack ./task_packs/ddia-coding-real \
  --api-url http://localhost:8000 \
  --runs 3 \
  --network \
  --timeout-seconds 300

The comparison output includes mean scores, score lift, pass-rate delta, error rate, timeout rate, run IDs, and a verdict.

Eval API:

• POST /eval-runs
• GET /eval-runs/{id}
• GET /eval-runs/{id}/report