nibzard/agentlab

최근 기록된 소스 활동: 2026년 3월 19일 17:15

SkillsMP 카탈로그 업데이트: 2026년 10월 29일 18:30

수집된 skills: 1

GitHub 스타: 4
GitHub 포크: 2

AI 어시스턴트로 설치

이 프롬프트를 복사해 사용 중인 AI 어시스턴트에 입력하세요.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/nibzard/agentlab.

저장소 README

README 원본 파일: README.md

README 원본 보기

README 전체 펼치기

# AgentLab: Proxmox Sandbox & Job Orchestration AgentLab is a production-ready sandbox orchestration system for Proxmox VE, enabling automated VM provisioning, job execution, workspace management, and artifact collection. Built with Go, it provides a daemon (`agentlabd`) and CLI (`agentlab`) for seamless sandbox lifecycle management. ## Overview AgentLab provisions unattended, network-isolated VM sandboxes on Proxmox VE for running AI coding agents in "dangerous"/YOLO mode with full API-based infrastructure integration. **Key components:** - `agentlabd` daemon on Proxmox host (owns Proxmox access, enforces policy) - `agentlab` CLI for local control via a Unix socket - `agentlab-dashboard` optional web UI for sandbox/job/workspace visibility - Proxmox API backend (recommended) or shell backend (fallback) - Guest `agent-runner` service inside VM template for bootstrap + execution **Security posture by default:** - Full outbound Internet access with RFC1918/ULA egress blocks - No host bind mounts; optional persistent workspaces via separate disks - One-time secrets delivery into tmpfs only - Host-owned exposures (Tailscale Serve) with audit events for every expose/unexpose - API token authentication instead of shell access ### Network Modes Profiles can set `network.mode` to one of `off`, `nat` (default), or `allowlist`. AgentLab maps these modes to Proxmox firewall groups and enables the firewall automatically: - `off` -> `agent_nat_off` (no network) - `nat` -> `agent_nat_default` (NAT + RFC1918/ULA blocks) - `allowlist` -> `agent_nat_allowlist` (egress allowlist rules) Ensure the firewall groups exist in Proxmox before using non-default modes. ## Prerequisites - Proxmox VE 9.x+ host (API backend) or 8.x+ (shell backend) - Storage pool suitable for templates/clones (ZFS or LVM-thin recommended) - `vmbr0` for LAN/WAN and ability to create `vmbr1` for agent subnet (defaults to `10.77.0.0/16`) - Go 1.24.0 or higher to build binaries (Go toolchain will auto-download if needed) - Tailscale on the host for remote SSH access (recommended) - Optional: Tailscale Admin API credentials if you want `agentlab bootstrap` to auto-approve subnet routes - Template VM with qemu-guest-agent enabled ## Quickstart (host setup) 1) Build binaries: ```bash make build ``` 2) Install binaries + systemd unit, plus AgentLab Claude skill bundle: ```bash sudo scripts/install_host.sh ``` 3) Configure networking (agent bridge + NAT/egress blocks): ```bash sudo scripts/net/setup_vmbr1.sh --apply sudo scripts/net/apply.sh --apply ``` 4) (Recommended) Enable Tailscale subnet routing: ```bash sudo scripts/net/setup_tailscale_router.sh --apply ``` 5) Create Proxmox API token and configure: ```bash # Create API token pveum user token add root@pam agentlab-api --privsep=0 # Configure /etc/agentlab/config.yaml with: # proxmox_backend: api # proxmox_api_url: https://localhost:8006 # proxmox_api_token: root@pam!agentlab-api=<token-uuid> ``` 6) Create secrets and minimal config/profile, then build the template: ```bash # Register the LLM keys and git credentials every VM needs. These go through # the daemon, so the bundle stays age-encrypted on disk (no local age key needed # on a laptop driving the CLI remotely over --endpoint). agentlab secrets set-env --name ANTHROPIC_API_KEY --value sk-ant-... agentlab secrets set-env --name OPENAI_API_KEY --value sk-... agentlab secrets set-git --git-token ghp_... --username x-access-token # Authorize SSH and enroll each VM onto the tailnet as agentlab-{vmid} agentlab secrets add-ssh-key --name laptop --key-file ~/.ssh/id_ed25519.pub agentlab secrets set-tailscale \ --admin-api-key tskey-api-XXXXXXXX \ --tailnet example.com \ --hostname-template 'agentlab-{vmid}' \ --tag tag:agentlab \ --extra-arg --ssh # Inspect or validate the bundle (values redacted by default) agentlab secrets show agentlab --json secrets validate # Build the template (installs Node 20 + the tailscale CLI), then restart the daemon sudo scripts/create_template.sh sudo systemctl restart agentlabd.service ``` There are two ways to enroll VMs onto the tailnet: - **Per-VM minting (recommended).** Pass `--admin-api-key tskey-api-...` (a Tailscale tailnet API access key). At each VM's bootstrap the daemon calls the Tailscale Admin API to mint a fresh single-use, ephemeral, preauthorized auth key for that VM alone. The minted key is transient — it is never stored, never written to a cloud-init snippet, and is delivered to the guest only over the subnet-gated, single-use bootstrap channel. `--tailnet` is optional and defaults to the key's own tailnet (`-`). - **Shared reusable key (fallback).** Pass `--authkey tskey-auth-...` instead. One reusable auth key is reused across all VMs (the hostname is still templated per-vmid). If both are configured, the daemon mints per VM and falls back to the shared key only if the Admin API call fails — so a transient Tailscale outage does not strand a VM. The Admin API key itself is treated as a secret: it is age-encrypted at rest, scrubbed from logs and every response, and used only to mint keys server-side. After upgrading Node, Tailscale, or any pinned agent CLI in the template, re-run `scripts/create_template.sh` so new VMs pick up the changes. 7) Check control-plane status: ```bash agentlab status ``` `agentlab status` includes installed skill bundle metadata under `Skill Bundle`. 8) Run a job from the host: ```bash agentlab job run --repo <git-url> --task "<task>" --profile yolo-ephemeral ``` ## Web Dashboard (Optional) AgentLab includes an optional `agentlab-dashboard` binary that serves a browser UI and proxies requests to `agentlabd` over its Unix socket. Start the dashboard: ```bash agentlab-dashboard --listen :8080 --socket /run/agentlab/agentlabd.sock ``` If daemon auth is enabled, pass a bearer token: ```bash agentlab-dashboard --listen :8080 --token <token> ``` Open `http://localhost:8080` in your browser. For full operator setup, see the runbook below. ## Documentation The documentation is organized with the [Diátaxis](https://diataxis.fr/) framework and published as a MkDocs site. Browse it under [`docs/`](docs/): - **Architecture**: [`docs/explanation/architecture.md`](docs/explanation/architecture.md) - system architecture, components, and data flow - **State model**: [`docs/reference/state-machine.md`](docs/reference/state-machine.md) - sandbox, workspace, session, and job states - **Remote CLI**: [`docs/how-to/connect-remote-daemon-over-tailnet.md`](docs/how-to/connect-remote-daemon-over-tailnet.md) - **Runbook and troubleshooting**: [`docs/how-to/debug-a-stuck-sandbox.md`](docs/how-to/debug-a-stuck-sandbox.md) - **Secrets**: [`docs/reference/secrets.md`](docs/reference/secrets.md) - **Local control API**: [`docs/reference/http-api.md`](docs/reference/http-api.md) - **Releases**: [`docs/reference/releases.md`](docs/reference/releases.md) ## Architecture Overview ```mermaid graph TB CLI[agentlab CLI] -->|Unix Socket| Daemon[agentlabd Daemon] Daemon --> Manager[Sandbox/Workspace/Job Managers] Manager --> Backend[Proxmox Backend] Backend --> Proxmox[Proxmox VE] Daemon --> DB[(SQLite Database)] Manager --> Bootstrap[Bootstrap API] Manager --> Artifacts[Artifact API] VM[Sandbox VMs] -->|HTTP| Bootstrap VM -->|HTTP| Artifacts classDef client fill:#e1f5fe,stroke:#01579b,stroke-width:2px classDef daemon fill:#f3e5f5,stroke:#4a148c,stroke-width:2px classDef storage fill:#fff3e0,stroke:#e65100,stroke-width:2px classDef proxmox fill:#e8f5e9,stroke:#1b5e20,stroke-width:2px class CLI client class Daemon,Manager,Backend,Bootstrap,Artifacts daemon class DB storage class Proxmox,VM proxmox ``` **For detailed architecture diagrams including:** - Complete system architecture with all components - Sandbox state machine with all transitions - Network topology and security rules - Job execution data flow - Database schema and relationships - Request lifecycle and error handling See [docs/explanation/architecture.md](docs/explanation/architecture.md) ## Sandbox States ```mermaid stateDiagram-v2 [*] --> REQUESTED REQUESTED --> PROVISIONING PROVISIONING --> BOOTING BOOTING --> READY READY --> RUNNING RUNNING --> COMPLETED RUNNING --> FAILED RUNNING --> TIMEOUT RUNNING --> STOPPED COMPLETED --> STOPPED FAILED --> STOPPED TIMEOUT --> STOPPED STOPPED --> DESTROYED REQUESTED --> DESTROYED PROVISIONING --> DESTROYED BOOTING --> DESTROYED READY --> DESTROYED RUNNING --> DESTROYED DESTROYED --> [*] ``` | State | Description | Allowed Operations | |--------|-------------|-------------------| | REQUESTED | VM creation requested | show, logs | | PROVISIONING | VM being created | show, logs | | BOOTING | VM is booting | show, logs | | READY | VM ready but not running | show, logs, destroy | | RUNNING | VM actively running | show, logs, destroy, lease renew | | STOPPED | VM stopped | show, logs, start, destroy | | TIMEOUT | Lease expired, VM may be gone | show, logs, destroy (--force) | | DESTROYED | VM destroyed | show | | FAILED | VM provisioning failed | show, logs | **Notes:** - Use `--force` with `sandbox destroy` to bypass state restrictions - Use `sandbox prune` to remove orphaned TIMEOUT sandboxes - Lease renewal only allowed in RUNNING state ## Proxmox Backend Configuration AgentLab supports two Proxmox backends: ### API Backend (Recommended) **Benefits:** - More reliable than shell commands - Better error handling and debugging - No Proxmox IPC layer issues - Standard HTTP authentication - Future-proof (Proxmox recommended approach) **Configuration:** ```yaml proxmox_backend: api # Use API backend (recommended) proxmox_api_url: https://localhost:8006 proxmox_api_token: root@pam!token-id=token-uuid proxmox_node: "" # Auto-detected if empty proxmox_tls_insecure: false # Default: verify TLS certificates # proxmox_tls_ca_path: /etc/agentlab/certs/proxmox-ca.pem # proxmox_tls_insecure: true # INSECURE: disable verification for self-signed certs ``` **Creating API Token:** ```bash # Create API token for AgentLab pveum user token add root@pam agentlab-api --privsep=0 # Output will show: # full-tokenid: root@pam!agentlab-api # value: 0d08f8bc-0a12-4072-8f18-264a33793bb9 # Add to config proxmox_api_token: root@pam!agentlab-api=0d08f8bc-0a12-4072-8f18-264a33793bb9 ``` **Testing API Backend:** ```bash # Test API connectivity (use -k only when proxmox_tls_insecure is true) curl -k -H "Authorization: PVEAPIToken=root@pam!token=uuid" \ https://localhost:8006/api2/json/nodes # Check template VM curl -k -H "Authorization: PVEAPIToken=root@pam!token=uuid" \ https://localhost:8006/api2/json/nodes/<node>/qemu/9000/config ``` ### Shell Backend (Fallback) **Use cases:** - Proxmox API not available - Legacy Proxmox versions - Troubleshooting **Configuration:** ```yaml proxmox_backend: shell # Use shell commands (qm, pvesh) ``` **Note:** Shell backend may encounter Proxmox IPC issues (e.g., `ipcc_send_rec[1] failed`). ## CLI Usage ### Global Flags ```text agentlab [global flags] <command> [command flags] Global flags: --endpoint URL Control plane endpoint (http(s)://host:port) --token TOKEN Control plane auth token (Authorization: Bearer) --socket PATH Path to agentlabd socket (default: /run/agentlab/agentlabd.sock) --json Output JSON format --timeout Request timeout (e.g., 30s, 2m) ``` ### Remote CLI When the daemon is configured with a remote control listener, you can connect from another machine: ```bash agentlab connect --endpoint http://host.tailnet.ts.net:8845 --token <token> ``` The remote listener serves **plaintext HTTP** behind the Tailscale TCP tunnel, so use `http://` (not `https://`) — the TLS-like termination happens at the Tailscale layer, not on the daemon port. This writes a client config file at `$XDG_CONFIG_HOME/agentlab/client.json` (or `~/.config/agentlab/client.json`). Subsequent commands use the saved endpoint and token automatically. Precedence (highest to lowest): CLI flags → environment variables (`AGENTLAB_ENDPOINT`, `AGENTLAB_TOKEN`) → config file → defaults. To remove the saved config: ```bash agentlab disconnect ``` ### Secrets Management AgentLab stores host-side credentials in encrypted bundles under `secrets_dir` (default: `/etc/agentlab/secrets`). Use the CLI to inspect and update them instead of editing bundle files manually: ```bash # Register LLM keys / arbitrary env (merge into the bundle) agentlab secrets set-env --name ANTHROPIC_API_KEY --value sk-ant-... agentlab secrets set-env --from-file env.txt # bulk: KEY=VAL lines or JSON # Register git credentials for private-repo cloning agentlab secrets set-git --git-token ghp_... --username x-access-token agentlab secrets set-git --ssh-key-file ~/.ssh/id_ed25519 --known-hosts-file ~/.ssh/known_hosts # Show the default bundle with sensitive values redacted agentlab secrets show # Show raw values when you explicitly need them agentlab secrets show --reveal --format json # Validate the configured bundle agentlab --json secrets validate # Add or remove guest SSH keys agentlab secrets add-ssh-key --name laptop --key-file ~/.ssh/id_ed25519.pub agentlab secrets remove-ssh-key --name laptop # Configure or clear guest Tailscale bootstrap settings. Per-VM minting is the # default (--admin-api-key); a shared --authkey is the fallback. agentlab secrets set-tailscale \ --admin-api-key tskey-api-XXXXXXXX \ --tailnet example.com \ --hostname-template 'agentlab-{vmid}' \ --tag tag:agentlab \ --extra-arg --ssh agentlab secrets clear-tailscale ``` `set-env`, `set-git`, and the other mutations go through the daemon: they work identically over the local socket (host) and a remote `--endpoint` (a laptop driving AgentLab from inside a coding agent), so a laptop agent can stage every credential a VM needs without a local age key. Mutations never echo raw secret values back — `show` and every mutation response redact them. Notes: - The CLI resolves `secrets_dir`, `secrets_bundle`, and `secrets_age_key_path` from `/etc/agentlab/config.yaml` by default. - Pass `--bundle`, `--dir`, `--age-key`, or `--config` to target a different bundle. - Plaintext bundle reads and writes require `--allow-plaintext`. - Existing `.sops.*` bundles can be read and validated, but in-place writes are not supported yet. ### Sandbox Management #### List Sandboxes ```bash agentlab sandbox list # Include destroyed sandboxes agentlab ls --all # JSON output agentlab sandbox list --json ``` Destroyed sandboxes are hidden by default in both table and JSON output. Use `--all` (or `-a`) to include them. Output: ``` VMID NAME PROFILE STATE IP LEASE 1020 test-api yolo-ephemeral RUNNING 10.77.0.130 2026-02-02T07:34:40Z ``` #### Create Sandbox ```bash agentlab sandbox new \ --profile yolo-ephemeral \ --name my-sandbox \ --ttl 10m # With modifiers (resolved to a profile name like secure-small) agentlab sandbox new \ --name my-sandbox \ --ttl 10m \ +small +secure # With workspace agentlab sandbox new \ --profile yolo-ephemeral \ --name my-sandbox \ --ttl 10m \ --workspace my-workspace # With specific VMID (advanced) agentlab sandbox new \ --profile yolo-ephemeral \ --vmid 1000 # With existing job (rare) agentlab sandbox new \ --profile yolo-ephemeral \ --job <job-id> ``` **Flags:** - `--profile <name>` (required unless using `+modifier`s): Profile name from `/etc/agentlab/profiles/` - `--name <name>` (optional): Sandbox name - `--ttl <duration>` (optional): Time-to-live (default from profile) - `--workspace <id>` (optional): Attach existing workspace - `--vmid <id>` (optional): Force specific VMID (advanced) - `--job <id>` (optional): Create sandbox for existing job - `--keepalive` (optional): Don't auto-destroy on TTL expiration Note: Modifiers are derived from profile names split on `-`, and combined modifiers are resolved by sorting and joining with `-`. #### Validate Sandbox Plan ```bash agentlab sandbox validate \ --profile yolo-ephemeral \ --name my-sandbox \ --ttl 10m # JSON output for automation agentlab sandbox validate \ --name my-sandbox \ --vmid 1001 \ --profile yolo-ephemeral \ --json ``` Use `agentlab sandbox validate` before `sandbox new` to confirm modifiers, workspace/VMID, and profile compatibility before creating a sandbox. #### Show Sandbox Details ```bash agentlab sandbox show 1020 # JSON output agentlab sandbox show 1020 --json ``` When the control plane can read live VM config from Proxmox, `sandbox show` includes current CPU core and memory settings under `resources` / `CPU Cores` and `Memory`. #### Update Sandbox Resources ```bash # Set a sandbox to 4 vCPU cores and 8 GiB RAM agentlab sandbox update --cores 4 --memory 8GiB 1020 # Update only memory agentlab sandbox update --memory 12288 1020 # JSON output for automation agentlab sandbox update --memory 12GiB 1020 --json ``` **Flags:** - `--cores <n>` (optional): Set the VM CPU core count - `--memory <size>` (optional): Set VM memory in MiB or GiB, for example `8192`, `8192MiB`, or `8GiB` - VMID (required): Target sandbox ID At least one of `--cores` or `--memory` is required. This updates the existing sandbox in place and records an audited sandbox configuration event in the control plane. #### Destroy Sandbox ```bash agentlab sandbox destroy 1020 # Force destroy (bypasses state checks) agentlab sandbox destroy --force 1020 ``` **Allowed transitions:** - Any state → DESTROYED (with `--force`) - TIMEOUT/DESTROYED → DESTROYED (normal) #### Revert Sandbox (Rollback to Clean Snapshot) ```bash # Revert to the canonical "clean" snapshot agentlab sandbox revert 1020 # Revert without restarting agentlab sandbox revert --no-restart 1020 # Force revert even if a job is running agentlab sandbox revert --force 1020 ``` #### Expose Sandbox Port (Tailnet) ```bash # Expose a sandbox port over the tailnet (host-owned + audited) agentlab sandbox expose 1020 :8080 # For non-interactive scripts, add --force to skip the confirmation prompt agentlab sandbox expose --force 1020 :8080 # List active exposures agentlab sandbox exposed # Remove an exposure by name agentlab sandbox unexpose sbx-1020-8080 ``` Exposure names default to `sbx-<vmid>-<port>`. Example list output: ``` NAME VMID PORT TARGET URL STATE UPDATED sbx-1020-8080 1020 8080 10.77.0.130 tcp://host.tailnet.ts.net:8080 serving 2026-02-08T20:30:00Z ``` #### Renew Lease ```bash agentlab sandbox lease renew --ttl 30m 1020 ``` **Flags:** - `--ttl <duration>` (required): New lease duration - VMID (required): Target sandbox ID #### Prune Orphaned Sandboxes ```bash agentlab sandbox prune ``` Removes sandboxes in TIMEOUT state that no longer exist in Proxmox. ### Workspace Management #### Create Workspace ```bash agentlab workspace create --name my-workspace --size 10 # With specific storage agentlab workspace create --name my-workspace --size 10 --storage local-zfs ``` **Flags:** - `--name <name>` (required): Workspace name - `--size <gb>` (required): Size in GB - `--storage <name>` (optional): Proxmox storage (default: local-zfs) #### List Workspaces ```bash agentlab workspace list ``` #### Attach Workspace to Sandbox ```bash agentlab workspace attach my-workspace 1020 ``` **Flags:** - Workspace name (required): Target workspace - VMID (required): Target sandbox #### Detach Workspace ```bash agentlab workspace detach my-workspace ``` **Note:** Workspace remains available after detaching. #### Rebind Workspace to New Sandbox ```bash agentlab workspace rebind my-workspace \ --profile yolo-ephemeral \ --ttl 30m ``` Creates a new sandbox and automatically attaches the workspace. **Flags:** - Workspace name (required): Target workspace - `--profile <name>` (required): Profile for new sandbox - `--ttl <duration>` (optional): TTL for new sandbox - `--keep-old` (optional): Keep old sandbox (default: destroys old) ### Profile Management #### List Profiles ```bash agentlab profile list # JSON output agentlab profile list --json ``` Outputs profile names and template VMIDs loaded by `agentlabd`. ### Job Execution #### Run Job ```bash agentlab job run \ --repo https://github.com/user/repo \ --task "run tests" \ --profile yolo-ephemeral \ --ttl 30m # With specific branch agentlab job run \ --repo https://github.com/user/repo \ --ref main \ --task "npm test" \ --profile yolo-ephemeral # With keepalive (no auto-termination) agentlab job run \ --repo https://github.com/user/repo \ --task "debug issue" \ --profile yolo-ephemeral \ --keepalive # With workspace agentlab job run \ --repo https://github.com/user/repo \ --task "data analysis" \ --profile yolo-ephemeral \ --workspace my-workspace # With workspace wait (wait for detach) agentlab job run \ --repo https://github.com/user/repo \ --task "data analysis" \ --profile yolo-ephemeral \ --workspace my-workspace \ --workspace-wait 2m # Create workspace on demand agentlab job run \ --repo https://github.com/user/repo \ --task "data analysis" \ --profile yolo-ephemeral \ --workspace new:ws-data \ --workspace-size 80 \ --workspace-storage local-zfs # Stateful shorthand (creates default workspace) agentlab job run \ --repo https://github.com/user/repo \ --task "data analysis" \ --profile yolo-ephemeral \ --stateful ``` #### Validate Job Plan ```bash agentlab job validate \ --repo https://github.com/user/repo \ --task "run tests" \ --profile yolo-ephemeral \ --ttl 30m # JSON output for automation agentlab job validate \ --repo https://github.com/user/repo \ --task "run tests" \ --profile yolo-ephemeral \ --ttl 30m \ --json ``` `agentlab job validate` validates the exact job preflight request without creating a job. The `job validate` command accepts the same flags as `job run` (including `--branch` and workspace flags) and returns non-zero on validation failure. **Flags:** - `--repo <url>` (required): Git repository URL - `--task <task>` (required): Task description or command - `--profile <name>` (required): Profile for sandbox - `--ref <branch>` (optional): Git branch/commit (default: default branch) - `--ttl <duration>` (optional): Sandbox TTL (default from profile) - `--keepalive` (optional): Don't auto-destroy on completion - `--workspace <id|name|new:name>` (optional): Attach existing workspace or create `new:<name>` - `--workspace-create <name>` (optional): Create workspace with name (requires `--workspace-size` unless `--stateful`) - `--workspace-size <size>` (optional): Workspace size for creation (e.g. 80G) - `--workspace-storage <name>` (optional): Workspace storage (default `local-zfs` for `--stateful`) - `--workspace-wait <duration>` (optional): Wait for workspace detach (e.g. 2m, 30s) - `--stateful` (optional): Create a default workspace named `stateful-<repo-slug>` with size 80G on local-zfs `--stateful` naming rule: `repo-slug` is the repo basename (strip `.git`), lowercased with non-alphanumeric characters replaced by `-`. API example for a workspace-backed job: ```json { "repo_url": "https://github.com/user/repo", "ref": "main", "profile": "yolo-ephemeral", "task": "data analysis", "mode": "dangerous", "workspace_id": "workspace-001", "workspace_wait_seconds": 60 } ``` API example creating a workspace during job create: ```json { "repo_url": "https://github.com/user/repo", "ref": "main", "profile": "yolo-ephemeral", "task": "data analysis", "mode": "dangerous", "workspace_create": { "name": "ws-data", "size_gb": 80, "storage": "local-zfs" } } ``` #### Show Job Status ```bash agentlab job show <job-id> # With event tailing agentlab job show <job-id> --events-tail 10 # JSON output agentlab job show <job-id> --json ``` #### List Job Artifacts ```bash agentlab job artifacts <job-id> ``` #### Download Artifacts ```bash # Download latest bundle agentlab job artifacts download <job-id> # Download to specific directory agentlab job artifacts download <job-id> --out ./artifacts # Download specific file agentlab job artifacts download <job-id> --name test-results.xml # Download all artifacts as bundle agentlab job artifacts download <job-id> --bundle # Download latest file agentlab job artifacts download <job-id> --latest ``` ### SSH Access #### SSH to Sandbox ```bash agentlab ssh 1020 # With custom user agentlab ssh 1020 --user ubuntu # With custom SSH key agentlab ssh 1020 --identity ~/.ssh/my_key # Execute command without interactive shell agentlab ssh 1020 --exec -- "uname -a" # Execute command and return output (non-interactive) agentlab ssh 1020 --exec "cat /etc/os-release" ``` **Flags:** - VMID (required): Target sandbox - `--user <name>` (optional): SSH user (default: root) - `--port <port>` (optional): SSH port (default: 22) - `--identity <path>` (optional): SSH private key path - `--exec <command>` (optional): Execute command instead of interactive shell ### Logs #### View Sandbox Logs ```bash # View all logs agentlab logs 1020 # Follow logs (tail -f) agentlab logs 1020 --follow # Tail last N lines agentlab logs 1020 --tail 50 # JSON output (one JSON object per line) agentlab logs 1020 --json ``` ## Architecture & Internals ### Daemon Architecture ``` agentlabd (Daemon) ├── HTTP API Server (Unix socket) │ ├── /v1/sandboxes │ ├── /v1/jobs │ ├── /v1/workspaces │ ├── /v1/artifacts │ └── /metrics ├── Proxmox Backend Layer │ ├── API Backend (HTTP) │ └── Shell Backend (CLI) ├── Sandbox Manager │ ├── Lifecycle management │ ├── Lease tracking │ └── State reconciliation ├── Workspace Manager │ ├── Volume management │ └── Attachment logic ├── Job Orchestrator │ ├── Provisioning │ ├── Execution │ └── Artifact collection └── SQLite Database ├── Sandboxes ├── Jobs ├── Workspaces └── Events ``` ### Proxmox Backend Comparison | Feature | API Backend | Shell Backend | |---------|-------------|---------------| | Reliability | High (HTTP) | Medium (IPC) | | Error Handling | Detailed | Basic | | Performance | Good | Good | | Debugging | HTTP logs | Command output | | Proxmox Issues | None (avoids IPC) | Possible | | Recommended | ✅ Yes | Fallback | ### State Reconciliation AgentLab runs a state reconciler every 5 minutes that: 1. Checks Proxmox for VMs not in database (orphaned VMs) 2. Checks database for VMs not in Proxmox (stale entries) 3. Auto-cleanup of zombie sandboxes 4. State synchronization between Proxmox and database **Example log output:** ``` reconcile: VM 1016 not found in Proxmox, marking as destroyed reconcile: found orphaned VM 1020 in Proxmox, importing to database ``` ## Monitoring & Observability ### Metrics Endpoint Configure in `/etc/agentlab/config.yaml`: ```yaml metrics_listen: 127.0.0.1:9090 ``` Access metrics: ```bash curl http://127.0.0.1:9090/metrics ``` ### Available Metrics ```prometheus # Sandbox counts by state agentlab_sandboxes_total{state="REQUESTED"} 5 agentlab_sandboxes_total{state="RUNNING"} 10 agentlab_sandboxes_total{state="DESTROYED"} 100 # Sandbox lifecycle counters agentlab_sandboxes_created_total 150 agentlab_sandboxes_destroyed_total 150 # Job metrics agentlab_jobs_total{status="success"} 120 agentlab_jobs_total{status="failed"} 5 agentlab_jobs_duration_seconds{quantile="0.95"} 45.2 # Workspace metrics agentlab_workspaces_total 15 agentlab_workspaces_attached_total 42 ``` ### Log Locations - **Daemon logs**: `/var/log/agentlab/agentlabd.log` - **Systemd journal**: `journalctl -u agentlabd.service -f` - **Database**: `/var/lib/agentlab/agentlab.db` - **Unix socket**: `/run/agentlab/agentlabd.sock` ### Debug Mode Enable verbose logging: ```bash # Add to systemd service [Service] Environment=AGENTLAB_DEBUG=1 # Restart systemctl restart agentlabd ``` Or temporarily: ```bash AGENTLAB_DEBUG=1 /usr/local/bin/agentlabd --config /etc/agentlab/config.yaml ``` ## Troubleshooting ### Common Issues Start with a quick status snapshot: ```bash agentlab status ``` #### Daemon Not Starting ```bash # Check logs journalctl -u agentlabd.service -n 50 --no-pager # Check configuration /usr/local/bin/agentlabd --config /etc/agentlab/config.yaml --help 2>&1 | head -20 # Verify socket directory ls -la /run/agentlab/ ``` #### Proxmox API Connection Failed ```bash # Test API manually curl -k -H "Authorization: PVEAPIToken=root@pam!token=uuid" \ https://localhost:8006/api2/json/nodes # Check token is valid pveum user token list root@pam # Verify API is accessible pvesh get /nodes ``` **Error:** `dial tcp: lookup localhost: no such host` **Fix:** Use `127.0.0.1` instead of `localhost` in config. #### Sandbox Stuck in REQUESTED ```bash # Check daemon logs tail -f /var/log/agentlab/agentlabd.log | grep "provision" # Check Proxmox VM status qm list | grep <vmid> # Check template VM qm config 9000 # Manual intervention agentlab sandbox destroy --force <vmid> ``` #### VM Not Getting IP Address ```bash # Check guest agent status on Proxmox qm guest cmd <vmid> ping # Check DHCP leases cat /var/lib/dnsmasq/dnsmasq.leases | grep <vmid-mac> # Restart guest agent in VM (if accessible via console) systemctl restart qemu-guest-agent # Verify network bridge ip link show vmbr1 ip addr show vmbr1 ``` #### Zombie Sandboxes (VM exists but not in database) ```bash # Automatic cleanup (via reconciler) agentlab sandbox prune # Manual cleanup for vmid in $(seq 1000 1020); do qm stop $vmid 2>/dev/null || true qm destroy $vmid --purge 1 2>/dev/null || true done # Verify cleanup qm list ``` #### CLI "socket not found" Error ```bash # Check socket exists ls -la /run/agentlab/agentlabd.sock # Check daemon is running systemctl status agentlabd # Use correct socket path agentlab --socket /run/agentlab/agentlabd.sock status ``` ## Releases agentlab ships as prebuilt binaries. Push a `v*` tag and the [release workflow](.github/workflows/release.yml) builds and publishes a GitHub Release with [GoReleaser](.goreleaser.yaml). No manual packaging is needed. Install the CLI from the latest release: ```bash curl -fsSL https://raw.githubusercontent.com/nibzard/agentlab/main/scripts/install.sh | bash ``` Each release contains: - `agentlab_<tag>_<os>-<arch>.tar.gz` - the CLI, for `linux` and `darwin` on `amd64` and `arm64`. - `agentlabd_<tag>_<os>-<arch>.tar.gz` - the daemon, for `linux` on `amd64` and `arm64`. - `checksums.txt` - SHA-256 sums for every archive. agentlab follows semantic versioning. Pre-release tags such as `v1.2.3-rc.1` publish as prereleases. See `docs/reference/releases.md` for the full procedure and `docs/how-to/upgrade-and-migrate.md` for upgrade steps. ## Development ### Building from Source ```bash # Using Go 1.24.0+ go build -v -o agentlabd ./cmd/agentlabd go build -v -o agentlab ./cmd/agentlab # Using Make make build ``` ### Running Tests ```bash # Run all tests go test ./... # Run with coverage go test -cover ./... # Run specific package go test ./internal/proxmox go test ./internal/daemon # Run verbose go test -v ./... ``` ### Code Structure ``` agentlab/ ├── cmd/ │ ├── agentlab/ # CLI application │ │ ├── main.go # Entry point │ │ ├── commands.go # Command handlers │ │ └── api.go # API client │ └── agentlabd/ # Daemon application │ └── main.go # Entry point ├── internal/ │ ├── config/ # Configuration management │ │ └── config.go # Config loading/validation │ ├── daemon/ # Daemon logic & HTTP API │ │ ├── daemon.go # Service startup │ │ ├── api.go # HTTP handlers │ │ ├── sandbox_manager.go │ │ └── workspace_manager.go │ ├── db/ # SQLite database │ │ └── store.go # Database operations │ ├── models/ # Data models │ │ ├── sandbox.go │ │ └── job.go │ ├── proxmox/ # Proxmox backend │ │ ├── proxmox.go # Backend interface │ │ ├── api_backend.go # API implementation │ │ ├── shell_backend.go # Shell implementation │ │ └── errors.go # Error types │ └── secrets/ # Secret management ├── docs/ # Documentation │ ├── runbook.md │ ├── api.md │ └── troubleshooting.md ├── scripts/ # Setup/maintenance scripts ├── go.mod # Go module definition └── README.md # This file ``` ### Contributing 1. Fork the repository 2. Create a feature branch (`git checkout -b feature/amazing-feature`) 3. Make your changes 4. Run tests (`go test ./...`) 5. Commit your changes (`git commit -m 'Add amazing feature'`) 6. Push to the branch (`git push origin feature/amazing-feature`) 7. Open a Pull Request **Coding standards:** - Follow Go conventions (gofmt, go vet) - Write tests for new features - Update documentation - Add metrics for new operations ## Performance Tuning ### Proxmox Backend Performance ```yaml # Reduce API timeout for faster failures proxmox_command_timeout: 30s # Default: 2m ``` ### Sandbox Provisioning Performance ```yaml # Increase parallelism (in future versions) provisioning_parallelism: 5 # Default: 1 ``` ### Memory Usage ```yaml # Reduce daemon memory footprint # Use smaller SQLite cache # Limit concurrent operations ``` ## Security Best Practices ### API Token Management - Use dedicated API token (not root password) - Set minimal required permissions (`--privsep=0` for no privilege separation) - Consider token expiration for production - Store token in `/etc/agentlab/config.yaml` with `chmod 600` ### Network Security - Use agent subnet (`vmbr1`) for isolation - Apply egress filtering (RFC1918 blocks) - Enable Tailscale for secure remote access - Restrict access to Proxmox API ### Sandbox Isolation - No host bind mounts by default - Workspaces use separate disks (not shared) - One-time secrets delivery (tmpfs only) - VM-level isolation via Proxmox ## License MIT License - See LICENSE file for details ## Acknowledgments - Proxmox VE team for excellent virtualization platform - Go community for fantastic tooling - Open-source contributors to related projects --- **AgentLab**: Orchestrate sandboxes with confidence. 🚀

저장소 README를 불러오는 중…

이 저장소의 skills

네트워크·컴퓨터 시스템 관리자

직업 카테고리 1개 · 100% 분류됨

skill

직업 분류

설명

업데이트

agentlab

네트워크·컴퓨터 시스템 관리자

Task-focused command bundles for AgentLab operations in Claude workflows.

원문 언어: 영어

2026-03-19

수집된 skill 1개 중 1개를 표시합니다.