teamworklisted
Install: claude install-skill sumitake/agent-collab
# Teamwork - role-based coordination
Use the active host's permitted subagent or inbox mechanisms; the skill runs
standalone from the installed package. The active primary retains objective
interpretation, architecture, integration, secrets, and every merge/deploy/
destructive decision.
## Roles
- Explorer: read-only research and option mapping.
- Worker: bounded implementation or output generation in an isolated area.
- Reviewer: independent family review of the worker artifact.
- Integrator: the active primary; verifies and combines accepted outputs.
## Workflow
1. State the decomposition economics and define non-overlapping ownership,
outputs, budgets, stop conditions, and trust posture.
2. Capture the current primary/model/session. Reviewer family is selected
dynamically; never encode a fixed model or assume the host name is a family.
3. For a managed model route, resolve the **plugin root** from this loaded file
and invoke only `python3 "<plugin-root>/coordinator.py"` using the exact
route/action contract. Claude remains asynchronous inbox-only.
4. Use isolated worktrees or paths for mutating workers. Model-generated output
is untrusted until the primary reviews and tests it.
5. Exclude the artifact author and active primary families from independent
review. Unknown-family governance fails closed; non-governance work carries
an independence warning.
6. Stop on new scope, ambiguous authority, exhausted budget, or an operator
gate. Do not let a