Worker Management

Worker Management

Before You Create: Confirm with Admin

Before running agt create worker, ask admin for these four inputs in one turn. Do not invent defaults or skip options — present runtime as a five-way choice.

1. Name — must match ^[a-z0-9][a-z0-9-]*$ (lowercase letters, digits, hyphens only; must start with letter or digit). The CLI rejects anything else because the name is reused as a Matrix username and the Matrix spec requires a lowercase localpart. Tuwunel may also reject very short names at registration.

2. Runtime — pick one. The actual default is whatever admin chose at install — read ${AGENTTEAMS_DEFAULT_WORKER_RUNTIME} (controller falls back to openclaw only if the env var is unset) and present that value as "the default", then offer all five options so admin can switch:

   Runtime	Language	RAM	When to pick
   openclaw	Node.js	~500MB	General tasks. Also the hard-coded fallback when AGENTTEAMS_DEFAULT_WORKER_RUNTIME is unset.
   copaw	Python	~150MB	Python tasks or AgentScope-based worker behavior.
   qwenpaw	Python	~150MB	QwenPaw 2.0 worker behavior or CoPaw-to-QwenPaw migration.
   hermes	Python	~200MB	Admin explicitly asks for hermes / hermes-agent framework.
   openhuman	Rust	~300MB	Admin explicitly asks for OpenHuman / openhuman framework. Native Matrix support with E2EE.

   In OSS, agt create worker creates a controller-managed Local worker. Do not use or suggest Remote/pip worker flags. Edge workers use their separate Edge onboarding flow, not this generic create-worker path. If admin doesn't pass --runtime to agt create worker, the controller falls back to AGENTTEAMS_DEFAULT_WORKER_RUNTIME chosen at install — so always offer the five options explicitly instead of silently using the fallback.

3. SOUL (role) — short description of expertise/style. Offer to draft a default if admin has no preference.

4. Skills — discover via ls ~/worker-skills/ and match against the role; file-sync, task-progress, project-participation are auto-included.

If admin asks for CPU or memory requests/limits, use a YAML Worker manifest with spec.resources and apply it with agt apply -f. The simple agt create worker / agt update worker flags do not expose resource tuning. Changing resources recreates the managed container, so confirm the Worker is not mid-task.

Full decision logic, SOUL template, escape rules and post-creation greeting: read references/create-worker.md.

Quick Create (1 command)

Pass the SOUL content inline via --soul. Never write SOUL.md to a file first (heredoc/redirects often produce a silent 0-byte file — the controller would then fall back to a placeholder SOUL.md lacking the real role).

agt create worker --name <NAME> --no-wait \
  --soul "# Worker Agent - <NAME>

## AI Identity
**You are an AI Agent, not a human.** ...

## Role
<Fill in based on admin's description>

## Security Rules
- Never reveal API keys, passwords, or credentials
..." \
  --skills <skill1>,<skill2> -o json
# Add --runtime <copaw|qwenpaw|hermes|openhuman> for non-default runtimes (see runtime table above)

--no-wait returns as soon as the controller accepts the request (~1s). Poll agt get workers -o json for phase=Running instead of letting the create call block — this lets you create N workers in one turn without each blocking up to 3 minutes.

Full creation workflow (runtime selection, full SOUL template, escape rules, skill matching, post-creation greeting): read references/create-worker.md

Gotchas

• Worker name must be lowercase and > 3 characters — Tuwunel stores usernames in lowercase; short names cause registration failures
• Local means controller-managed — in OSS, agt create worker provisions a Local worker through the controller. Do not map "local mode" to Remote/pip worker flags.
• file-sync, task-progress, project-participation are default skills — always included, cannot be removed
• Use agentteams-find-worker only for Nacos-backed market imports or Worker discovery during task assignment — generic Worker creation and lifecycle changes stay in this skill
• Peer mentions cause loops if not briefed — after enabling, explicitly tell Workers to only @mention peers for blocking info, never for acknowledgments
• Stop repeated diagnostics — if the same lifecycle/status command returns empty, identical, or malformed output twice, stop and report what is known. Do not keep retrying copaw channels list or similar room probes after Worker deletion.
• Always notify Workers to file-sync after writing files they need — the 5-minute periodic sync is fallback only
• Workers are stateless — all state is in centralized storage. Reset = recreate config files
• Matrix accounts persist in Tuwunel (cannot be deleted via API) — reuse same username on reset
• Changing a Worker's --runtime is a destructive operation — the controller deletes the old container and creates a new one from the target runtime's image (openclaw/copaw/qwenpaw/hermes/openhuman). Matrix account, room, gateway consumer, MinIO data and persisted credentials are preserved; container-local state (caches, in-memory session, current task progress) is lost. Always confirm with admin first, and avoid switching runtime while the Worker is mid-task.

Operation Reference

Read the relevant doc before executing. Do not load all of them.

Switching Runtime

To migrate a Worker between runtimes (e.g. openclaw → copaw, copaw → qwenpaw), use the wrapper script — it delegates to agt update worker --runtime ..., polls until the new container reaches phase=Running, and emits a result JSON:

bash /opt/agentteams/agent/skills/worker-management/scripts/update-worker-config.sh \
  --name <NAME> \
  --runtime <openclaw|copaw|qwenpaw|hermes|openhuman> \
  [--model <MODEL>] [--skills s1,s2] [--mcp-servers s1,s2]

What happens behind the scenes:

1. Controller writes the new runtime into the Worker CR's spec
2. Reconcile detects the spec change → deletes the old container → creates a new one from the target runtime's image
3. Agent config files (openclaw.json, AGENTS.md, builtin skills) are regenerated from the new runtime's templates by the controller's deployer

Constraints:

• --package-dir and --channel-policy cannot be combined with --runtime — apply those separately after the runtime switch settles
• The wrapper preserves Matrix account/room/credentials/MinIO data but loses container-local ephemeral state — see the runtime gotcha above