name: headroom

services:
    # ===========================================================================
    # CLI — interactive session for ad-hoc headroom commands
    # Not needed for production proxy operation.
    # Only start when you need to run headroom commands manually.
    # ===========================================================================
    cli:
        image: ${HEADROOM_IMAGE:-ghcr.io/chopratejas/headroom:latest}
        entrypoint: ["headroom"]
        working_dir: /workspace
        stdin_open: true
        tty: true
        profiles:
            - cli
        env_file:
            - .env
        environment:
            HOME: /tmp/headroom-home
            HEADROOM_WORKSPACE_DIR: /tmp/headroom-home/.headroom
            HEADROOM_CONFIG_DIR: /tmp/headroom-home/.headroom/config
        volumes:
            - ${HEADROOM_WORKSPACE:-.}:/workspace
            - ${HEADROOM_HOST_HOME:?set HEADROOM_HOST_HOME}/.headroom:/tmp/headroom-home/.headroom
            - ${HEADROOM_HOST_HOME:?set HEADROOM_HOST_HOME}/.claude:/tmp/headroom-home/.claude
            - ${HEADROOM_HOST_HOME:?set HEADROOM_HOST_HOME}/.codex:/tmp/headroom-home/.codex
            - ${HEADROOM_HOST_HOME:?set HEADROOM_HOST_HOME}/.gemini:/tmp/headroom-home/.gemini
        command: ["--help"]
        networks:
            - headroom

    # ===========================================================================
    # Proxy — drop-in LLM proxy with compression, caching, and persistent memory
    #
    # Agent (Zed) sends requests → Headroom compresses, caches, adds context → LLM
    # Persistent memory (Tier 1): local SQLite + HNSW for session-level context
    # Shared memory (Tier 2): Agent queries Chroma directly before sending to proxy
    #
    # Docs: https://headroom-docs.vercel.app/docs/proxy
    # ===========================================================================
    proxy:
        image: ${HEADROOM_IMAGE:-ghcr.io/chopratejas/headroom:latest}
        # Install rtk-ai at startup (auto-detected by Headroom for CLI output filtering)
        # Pip sees it's already installed on subsequent restarts — no delay.
        entrypoint:
            [
                "sh",
                "-c",
                'pip install rtk-ai -q 2>/dev/null && exec headroom proxy "$0" "$@"',
            ]
        restart: unless-stopped
        env_file:
            - .env
        environment:
            HOME: /tmp/headroom-home
            HEADROOM_HOST: 0.0.0.0
            HEADROOM_WORKSPACE_DIR: /tmp/headroom-home/.headroom
            HEADROOM_CONFIG_DIR: /tmp/headroom-home/.headroom/config
            HEADROOM_TELEMETRY: ${HEADROOM_TELEMETRY:-off}
            HEADROOM_LOG_LEVEL: ${HEADROOM_LOG_LEVEL:-INFO}
            # Langfuse / OpenTelemetry tracing
            # Keys (LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY) come from .env via env_file
            LANGFUSE_HOST: http://langfuse-web:3000
            LANGFUSE_PUBLIC_KEY: ${LANGFUSE_PUBLIC_KEY}
            LANGFUSE_SECRET_KEY: ${LANGFUSE_SECRET_KEY}
            OTEL_SERVICE_NAME: headroom-proxy
            OTEL_EXPORTER_OTLP_ENDPOINT: http://langfuse-web:3000/api/public/otel/v1
        ports:
            - "${HEADROOM_PORT:-8787}:${HEADROOM_PORT:-8787}"
        volumes:
            # Headroom persistent data (memories, config, cache, learn artifacts)
            - ${HEADROOM_HOST_HOME:?set HEADROOM_HOST_HOME}/.headroom:/tmp/headroom-home/.headroom
            # Agent-specific directories are NOT mounted here — they're for the
            # CLI / wrap mode. The proxy only needs the shared .headroom dir.
        command:
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "${HEADROOM_PORT:-8787}"
            - "--openai-api-url"
            - "${OPENAI_TARGET_API_URL:-https://api.venice.ai/v1}"
            - "--memory"
            - "--memory-db-path"
            - "/tmp/headroom-home/.headroom/memory.db"
            # Traffic learning — writes AGENTS.md / MEMORY.md with failure
            # patterns mined from agent sessions
            - "--learn"
        # healthcheck:
        #     test: ["CMD", "headroom", "--help"]
        #     interval: 30s
        #     timeout: 10s
        #     retries: 3
        #     start_period: 15s
        deploy:
            resources:
                limits:
                    memory: 8G
                    cpus: "8"
                reservations:
                    memory: 4G
                    cpus: "2"
        networks:
            headroom: {}
            langfuse:
                aliases:
                    - headroom-proxy
            pipeline:
                aliases:
                    - headroom-proxy

networks:
    headroom:
        name: headroom
        driver: bridge
    langfuse:
        name: langfuse_langfuse
        external: true
    pipeline:
        name: pipeline
        external: true