workerflow-internal-workflow

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.