Worker Internal Workflow

Use this skill before splitting work inside a single Worker. WorkerFlow is not TeamHarness delegation.

Decision

Need	Path
Current Worker can do it directly	No subagent
Same Worker needs short parallel inspection	QwenPaw native subagent
Same Worker needs visible fan-out to custom roles, workspaces, AGENTS.md, or skills	Dynamic workflow via worker_agentflow workflow_run
Same Worker needs one manually controlled temporary subagent	create_temp_agent directly
Work needs a persistent team member or Leader acceptance	Do not use WorkerFlow; use the team workflow

Native Subagent

Use QwenPaw native subagents for short internal parallelism. The current Worker still owns the final result.

Do not create a temporary agent when no custom prompt, workspace, or skill set is required.

Temporary QwenPaw Agent

Use worker_agentflow with workflow_run when a custom AgentSpec-style template has already been placed in the default QwenPaw workspace:

<QWENPAW_WORKING_DIR>/workspaces/default/subagents/<role>/
  AGENTS.md
  PROFILE.md        # optional
  SOUL.md           # optional
  skills/
    <skill-id>/
      SKILL.md

Use list_subagents first when you need to discover the templates available in the default workspace. Use create_temp_agent directly only when you need manual control outside a dynamic workflow.

Subagent template rules:

• AGENTS.md should tell the temporary agent to begin analysis immediately.
• AGENTS.md should forbid greetings, self-introductions, and capability summaries.
• AGENTS.md should define the subagent's focus area so the Worker can send the complete source input instead of manually slicing the input.
• When multiple subagent results will be merged, each AGENTS.md should require the same fixed output shape.

Dynamic workflow lifecycle:

1. Call workflow_run with the explicit current DM/conversation roomId, title, source input, optional merge.instruction, and either subagents or DAG nodes.
2. Each subagent or node must name a default-workspace template with subagent and a bounded task.
3. workflow_run starts the Matrix card, creates tmp-... subagents, creates the run-level shared directory, stores workflow.json, and returns submitInstructions for nodes that are ready now.
4. Send each returned submitPrompt to its agentId using the available QwenPaw agent communication tools.
5. When a subagent finishes, call workflow_update with a steps row whose id matches the node id, status is done, and summary contains the short result.
6. For DAG nodes, inspect every workflow_update response. If it returns readyInstructions, immediately send each returned submitPrompt to its agentId.
7. Use workflow_update at meaningful phase changes: submitted, running, retrying, merging, cleanup, done, or failed.
8. Retry a timed-out subagent once when the task is still useful.
9. Mark subagents that still fail as missing or failed in the merged result.
10. Merge subagent results into the current Worker's own output.
11. Delete every temporary agent with delete_temp_agent.
12. Call workflow_finish or workflow_fail.
13. After merging, either keep the shared run directory as evidence or remove it with cleanup_shared.

Example workflow_run input:

{
  "action": "workflow_run",
  "roomId": "!current-dm:example.test",
  "runId": "run-review-auth",
  "title": "Review auth module",
  "input": "Review auth.py for this task...",
  "subagents": [
    {
      "id": "security",
      "subagent": "security-reviewer",
      "task": "Check authentication and authorization risks."
    },
    {
      "id": "tests",
      "subagent": "test-reviewer",
      "task": "Check test coverage gaps."
    }
  ],
  "merge": {
    "instruction": "Return one severity-ordered finding list."
  }
}

DAG nodes example:

{
  "action": "workflow_run",
  "roomId": "!current-dm:example.test",
  "runId": "pig-diagnosis-b001",
  "title": "猪病诊断报告生成",
  "input": "批次号 B001 的猪病诊断请求...",
  "nodes": [
    {
      "id": "base_info",
      "name": "基础信息梳理",
      "role": "基础信息梳理",
      "subagent": "base-info-agent",
      "task": "处理基础信息。"
    },
    {
      "id": "feed_cough",
      "name": "采食与咳喘指标",
      "role": "采食与咳喘指标",
      "subagent": "feed-cough-agent",
      "task": "分析采食及咳嗽波动。"
    },
    {
      "id": "pathogen",
      "name": "病原线索分析",
      "role": "病原线索分析",
      "subagent": "pathogen-agent",
      "task": "分析病原及亚型。"
    },
    {
      "id": "disease_analysis",
      "name": "疾病综合研判",
      "role": "疾病综合研判",
      "subagent": "disease-analysis-agent",
      "task": "综合基础信息、采食咳嗽和病原结果，判断猪病。",
      "dependsOn": ["base_info", "feed_cough", "pathogen"]
    },
    {
      "id": "medication",
      "name": "用药与处置措施",
      "role": "用药与处置措施",
      "subagent": "medication-measures-agent",
      "task": "基于猪病分析生成药物及措施方案。",
      "dependsOn": ["disease_analysis"]
    },
    {
      "id": "diagnosis_advice",
      "name": "诊断建议",
      "role": "诊断建议",
      "subagent": "diagnosis-advice-agent",
      "task": "基于猪病分析生成诊断建议。",
      "dependsOn": ["disease_analysis"]
    },
    {
      "id": "report",
      "name": "最终报告汇总",
      "role": "最终报告汇总",
      "subagent": "report-agent",
      "task": "汇总上游结果生成最终诊断报告。",
      "dependsOn": ["medication", "diagnosis_advice"]
    }
  ]
}

For DAG workflows, workflow_run creates every node's temporary agent. Only dependency-free nodes appear in submitInstructions. Dependent nodes appear in waitingInstructions. The current Worker is the scheduler: submit ready nodes, wait for their replies, call workflow_update with done steps, then submit any downstream readyInstructions returned by that update. WorkerFlow uses the done step ids as the dependency-completion signal and appends upstream summaries to newly unblocked submit prompts.

File sharing:

• Native QwenPaw subagents share the current Worker's workspace.
• Temporary QwenPaw agents created through /api/agents have separate workspaces. Do not point their workspaceDir at the default workspace.
• workflow_run and create_temp_agent expose a run-level shared directory: <default-workspace>/shared/workerflow/<sharedRunId>/.
• Put input files under shared/inputs/.
• Each subagent writes only under shared/outputs/<agent-id>/.
• The subagent workspace contains .workerflow/shared.json and a shared symlink for discovery, but prompts should still include the absolute shared.path, shared.inputs, and shared.output returned by the tool.
• Do not place the shared directory inside a temporary agent workspace, because cleanupWorkspace may delete that workspace.

Workflow visibility:

• The current Worker is the Matrix sender for WorkerFlow workflow cards.
• Temporary agents do not send Matrix status directly.
• Prefer workflow_run for visible multi-subagent work; it creates the first Matrix m.notice card and records the returned eventId.
• workflow_run and workflow_start require an explicit current DM/conversation roomId. WorkerFlow does not fallback to Team Room or personal room.
• Use workflow_update, workflow_finish, and workflow_fail to edit the same card through Matrix m.replace.
• The card state is also stored in <default-workspace>/shared/workerflow/<run-id>/workflow.json.
• Show phase-level changes only: spawning subagents, running, retrying, merging, cleanup, done, or failed.
• Do not expose full prompts, secrets, raw large inputs, or tool traces in the card. Use shared paths and short summaries.

Fan-out rules:

• Send the full source input to each subagent unless there is a clear token, privacy, or cost reason to narrow it.
• Let each subagent extract the subset relevant to its AGENTS.md.
• Request a fixed output shape for all subagents when the results need programmatic or low-friction merging.
• Keep a small ledger of created temporary agent ids so cleanup can run even when one subagent fails.
• Keep the shared run id in the same ledger when file sharing is used.

Rules:

• Temporary agents are implementation details of the current Worker.
• Always use tmp- ids for temporary agents.
• Do not store temporary agent ids as TeamHarness workers.
• Delete temporary agents after the bounded task completes or fails; treat this as a finally step.
• If the work must survive the current Worker run, use a durable team workflow instead.