diff --git a/.env.example b/.env.example index d5d801e..fc3c96a 100644 --- a/.env.example +++ b/.env.example @@ -94,6 +94,13 @@ FORWARD_AUTH_SECRET= # [SECRET] Shared secret for Caddy ↔ # Store all project secrets here so formulas reference env vars, never hardcode. BASE_RPC_URL= # [SECRET] on-chain RPC endpoint +# ── Local Qwen dev agent (optional) ────────────────────────────────────── +# Set ENABLE_LLAMA_AGENT=1 to emit agents-llama in docker-compose.yml. +# Requires a running llama-server reachable at ANTHROPIC_BASE_URL. +# See docs/agents-llama.md for details. +ENABLE_LLAMA_AGENT=0 # [CONFIG] 1 = enable agents-llama service +ANTHROPIC_BASE_URL= # [CONFIG] e.g. http://host.docker.internal:8081 + # ── Tuning ──────────────────────────────────────────────────────────────── CLAUDE_TIMEOUT=7200 # [CONFIG] max seconds per Claude invocation diff --git a/AGENTS.md b/AGENTS.md index e647d24..d768f20 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -118,6 +118,7 @@ bash dev/phase-test.sh | Reproduce | `docker/reproduce/` | Bug reproduction using Playwright MCP | `formulas/reproduce.toml` | | Triage | `docker/reproduce/` | Deep root cause analysis | `formulas/triage.toml` | | Edge dispatcher | `docker/edge/` | Polls ops repo for vault actions, executes via Claude sessions | `docker/edge/dispatcher.sh` | +| agents-llama | `docker/agents/` (same image) | Local-Qwen dev agent (`AGENT_ROLES=dev`), gated on `ENABLE_LLAMA_AGENT=1` | [docs/agents-llama.md](docs/agents-llama.md) | > **Vault:** Being redesigned as a PR-based approval workflow (issues #73-#77). > See [docs/VAULT.md](docs/VAULT.md) for the vault PR workflow details. diff --git a/bin/disinto b/bin/disinto index bbb11ec..84200c9 100755 --- a/bin/disinto +++ b/bin/disinto @@ -890,6 +890,19 @@ p.write_text(text) echo "Config: CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 saved to .env" fi + # Write local-Qwen dev agent env keys with safe defaults (#769) + if ! grep -q '^ENABLE_LLAMA_AGENT=' "$env_file" 2>/dev/null; then + cat >> "$env_file" <<'LLAMAENVEOF' + +# Local Qwen dev agent (optional) — set to 1 to enable +ENABLE_LLAMA_AGENT=0 +FORGE_TOKEN_LLAMA= +FORGE_PASS_LLAMA= +ANTHROPIC_BASE_URL= +LLAMAENVEOF + echo "Config: ENABLE_LLAMA_AGENT keys written to .env (disabled by default)" + fi + # Create labels on remote create_labels "$forge_repo" "$forge_url" diff --git a/docs/agents-llama.md b/docs/agents-llama.md new file mode 100644 index 0000000..6764360 --- /dev/null +++ b/docs/agents-llama.md @@ -0,0 +1,42 @@ +# agents-llama — Local-Qwen Dev Agent + +The `agents-llama` service is an optional compose service that runs a dev agent +backed by a local llama-server instance (e.g. Qwen) instead of the Anthropic +API. It uses the same Docker image as the main `agents` service but connects to +a local inference endpoint via `ANTHROPIC_BASE_URL`. + +## Enabling + +Set `ENABLE_LLAMA_AGENT=1` in `.env` (or `.env.enc`) and provide the required +credentials: + +```env +ENABLE_LLAMA_AGENT=1 +FORGE_TOKEN_LLAMA= +FORGE_PASS_LLAMA= +ANTHROPIC_BASE_URL=http://host.docker.internal:8081 # llama-server endpoint +``` + +Then regenerate the compose file (`disinto init ...`) and bring the stack up. + +## Prerequisites + +- **llama-server** (or compatible OpenAI-API endpoint) running on the host, + reachable from inside Docker at the URL set in `ANTHROPIC_BASE_URL`. +- A Forgejo bot user (e.g. `dev-qwen`) with its own API token and password, + stored as `FORGE_TOKEN_LLAMA` / `FORGE_PASS_LLAMA`. + +## Behaviour + +- `AGENT_ROLES=dev` — the llama agent only picks up dev work. +- `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=60` — more aggressive compaction for smaller + context windows. +- `depends_on: forgejo (service_healthy)` — does **not** depend on Woodpecker + (the llama agent doesn't need CI). +- Serialises on the llama-server's single KV cache (AD-002). + +## Disabling + +Set `ENABLE_LLAMA_AGENT=0` (or leave it unset) and regenerate. The service +block is omitted entirely from `docker-compose.yml`; the stack starts cleanly +without it. diff --git a/lib/generators.sh b/lib/generators.sh index 3b42b5d..6157710 100644 --- a/lib/generators.sh +++ b/lib/generators.sh @@ -381,6 +381,63 @@ services: networks: - disinto-net +COMPOSEEOF + + # ── Conditional agents-llama block (ENABLE_LLAMA_AGENT=1) ────────────── + # Local-Qwen dev agent — gated on ENABLE_LLAMA_AGENT so factories without + # a local llama endpoint don't try to start it. See docs/agents-llama.md. + if [ "${ENABLE_LLAMA_AGENT:-0}" = "1" ]; then + cat >> "$compose_file" <<'LLAMAEOF' + + agents-llama: + build: + context: . + dockerfile: docker/agents/Dockerfile + container_name: disinto-agents-llama + restart: unless-stopped + security_opt: + - apparmor=unconfined + volumes: + - agent-data:/home/agent/data + - project-repos:/home/agent/repos + - ${CLAUDE_SHARED_DIR:-/var/lib/disinto/claude-shared}:${CLAUDE_SHARED_DIR:-/var/lib/disinto/claude-shared} + - ${HOME}/.claude.json:/home/agent/.claude.json:ro + - CLAUDE_BIN_PLACEHOLDER:/usr/local/bin/claude:ro + - ${HOME}/.ssh:/home/agent/.ssh:ro + - ${HOME}/.config/sops/age:/home/agent/.config/sops/age:ro + - woodpecker-data:/woodpecker-data:ro + environment: + FORGE_URL: http://forgejo:3000 + FORGE_REPO: ${FORGE_REPO:-disinto-admin/disinto} + FORGE_TOKEN: ${FORGE_TOKEN_LLAMA:-} + FORGE_PASS: ${FORGE_PASS_LLAMA:-} + FORGE_BOT_USERNAMES: ${FORGE_BOT_USERNAMES:-} + WOODPECKER_TOKEN: ${WOODPECKER_TOKEN:-} + CLAUDE_TIMEOUT: ${CLAUDE_TIMEOUT:-7200} + CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC: ${CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC:-1} + CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: "60" + ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-} + ANTHROPIC_BASE_URL: ${ANTHROPIC_BASE_URL:-} + FORGE_ADMIN_PASS: ${FORGE_ADMIN_PASS:-} + DISINTO_CONTAINER: "1" + PROJECT_NAME: ${PROJECT_NAME:-project} + PROJECT_REPO_ROOT: /home/agent/repos/${PROJECT_NAME:-project} + WOODPECKER_DATA_DIR: /woodpecker-data + WOODPECKER_REPO_ID: "PLACEHOLDER_WP_REPO_ID" + CLAUDE_CONFIG_DIR: ${CLAUDE_CONFIG_DIR:-/var/lib/disinto/claude-shared/config} + POLL_INTERVAL: ${POLL_INTERVAL:-300} + AGENT_ROLES: dev + depends_on: + forgejo: + condition: service_healthy + networks: + - disinto-net +LLAMAEOF + fi + + # Resume the rest of the compose file (runner onward) + cat >> "$compose_file" <<'COMPOSEEOF' + runner: build: context: .