oma-coordination Skill
手动协调分配的专项任务和交接。用于逐步监督多智能体项目。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
Multi-Agent Coordination Rules
Scheduling
Goal
Supply the coordination rules every multi-agent run shares (plan reuse, contract alignment, workspace ownership, QA handoff), and guide manual step-by-step coordination when the user drives each dispatch.
Intent signature
- A workflow (
/work,/orchestrate,/ultrawork) loads this skill and applies its Guardrails. - User wants step-by-step coordination, manual agent spawning, or multi-domain work planning without full automation.
- Task spans multiple specialist agents and requires contract alignment.
When to use
- Any workflow that dispatches more than one specialist agent
- Complex feature spanning multiple domains where the user wants to drive each step
- Coordination needed between agents that share an API or data contract
When NOT to use
- Simple single-domain task -> use the specific agent directly
- Automated dispatch, monitoring, and verification ->
oma-orchestrationowns execution; this skill supplies only the rules it applies - Quick bug fixes or minor changes
Expected inputs
- Complex feature or project goal
- Current plan: task ownership, acceptance checks, dependencies, and priority tiers
- Workspace/session constraints and API/data contract needs
Expected outputs
- Coordination sequence: plan reuse or PM decomposition, spawn order, monitoring guidance, and QA review
- API/data contract alignment checkpoints
Dependencies
oma-pmfor decomposition,oma-qafor review, andoma-orchestrationfor dispatch resolution, the parallelism limit, the recovery budget, and the review loop- CLI
oma agent spawnand the progress/result conventions in../_shared/runtime/memory-protocol.md
Structural Flow
Entry
- Confirm the task needs more than one specialist agent.
- Identify priority tiers and shared contracts from the current plan.
- Apply the shared collaboration policy before assigning parallel work or coordinating ownership handoffs and result collection.
Transitions
- If task is simple, route to one specialist.
- If the user wants automated execution, hand execution to
oma-orchestration; the Guardrails below still apply. - If QA reports a blocking defect or required verification gap, route the affected work to the responsible agent within the existing recovery budget.
Failure and recovery
- If contracts diverge, pause dependent work until the owning task's API/data contract is reconciled.
- If agent workspaces conflict, split ownership boundaries.
- If progress stalls, inspect progress files and reissue focused instructions.
Exit
- Success: specialist outputs are coordinated and QA-reviewed.
- Partial success: blocked agents, contract conflicts, or QA failures are explicit.
Logical Operations
Actions
| Action | SSL primitive | Evidence |
|--------|---------------|----------|
| Read request and domains | READ | User prompt and project context |
| Select agent plan | SELECT | PM decomposition and priority tiers |
| Spawn agents | CALL_TOOL | oma agent spawn |
| Monitor progress | READ | injected task/run-scoped progress, claims, and receipts |
| Validate contracts | VALIDATE | API/data model alignment |
| Notify coordination status | NOTIFY | Final coordination summary |
Canonical command path
# One spawn per ready task; agent, task ID, and workspace come from the current plan.
oma agent spawn <agent-id> <prompt-file> <session-id> --task-id <task.id> -w <workspace>
Resolve the dispatch path per agent as oma-orchestration defines: the runtime's native subagent path when the agent's target vendor equals the current runtime vendor, otherwise oma agent spawn.
Useful agent spawn options: --vendor (CLI vendor override), --isolation worktree (git worktree per spawn, prevents file conflicts), --read-only (non-destructive tools only, e.g. for review/QA passes).
While agents run, apply ../_shared/core/execution-policy.md for the execution deadline, resource ownership, and cancellation. Read injected run artifacts and receipts; progress and result file names are defined in ../_shared/runtime/memory-protocol.md.
Resource scope
| Scope | Resource target |
|-------|-----------------|
| LOCAL_FS | Progress/result files and workspaces |
| PROCESS | Agent spawn commands |
| MEMORY | Session state and task board |
| CODEBASE | Shared contracts and implementation areas |
Preconditions
- Task requires multiple domains.
- A current plan identifies task ownership, acceptance checks, dependencies, and priority tiers.
Effects and side effects
- Spawns or guides multiple agents.
- Coordinates workspace ownership and QA feedback.
Guardrails
- Reuse a valid plan; call PM only when decomposition or contract changes are required. Preserve injected session/task/run IDs and do not regenerate a plan already frozen by dispatch.
- Run independent tasks of the same priority tier in parallel, within the parallelism limit
oma-orchestrationresolves from project configuration. - Reuse or update contracts for changed API boundaries before dependent work starts.
- Accept an implementation only after QA review. Review follows the iterative loop
oma-orchestrationdefines (verify, QA cross-review, fix, review again within the recovery budget), not a single final pass. - Assign separate workspaces to avoid file conflicts (or use
--isolation worktreefor a git worktree per spawn). - Resume from the current handoff state and execute applicable steps in dependency order; record why a branch is not applicable.
References
- Parallel worker collaboration:
../_shared/core/collaboration-policy.md(before assigning parallel work or coordinating peer issues, ownership handoffs, and result collection) - Deadline, resource ownership, and cancellation:
../_shared/core/execution-policy.md(while agents run) - Progress and result file naming:
../_shared/runtime/memory-protocol.md(when reading run artifacts) - Dispatch resolution, parallelism limit, recovery budget, and review loop:
../oma-orchestration/SKILL.md(when dispatching or reviewing)