recursive-swarm（递归群）

Recursive Swarm

Use this skill to turn one large task into a small, bounded task tree with explicit node state, artifacts, merge points, and audit events.

Use this workflow

Use recursive-swarm only when all of these are true:

1. The task is too large or messy for one agent turn.
2. The task has at least 2 meaningful subproblems.
3. The outputs can be merged back into one answer, report, or code result.

Do not recurse for simple reads, one-shot edits, destructive workflows, or open-ended exploration with no clear output contract.

Defaults

• Default max depth: 2
• Hard cap: 3
• Preferred fan-out per node: 2-4 children
• Hard cap per node: 5 children
• Default concurrency: 2-3 active leaves
• Hard cap concurrency: 5
• Default max nodes per run: 9
• Hard cap nodes per run: 15
• Default workspace mode: artifact folders, not git worktrees
• Append-only audit log: events.jsonl
• Use git worktrees only for coding leaves inside a git repo when sibling edits benefit from isolation

Node types

Assign each node one primary type:

• research — extraction, investigation, analysis
• coding — implementation, refactor, test work
• ops — shell, system, or environment workflows
• browser — web/UI automation
• synthesis — combine child outputs into one merged result
• review — challenge weak claims, reconcile conflicts, prune bad findings

Execution routing

Route nodes like this:

• research, synthesis, review → subagents
• coding → ACP sessions by default; use worktrees only when useful
• ops, browser → direct tools or a narrowly scoped subagent if needed
• destructive, external, or approval-sensitive nodes → pause and ask before execution

Quiet child execution (default)

For routine child runs, prefer silent child completion + parent-owned final delivery.

That means:

• child sessions do the work
• child results are harvested from session history and/or artifact files
• routine child announce steps should reply exactly ANNOUNCE_SKIP
• the parent/orchestrator sends the one final merged answer to the user

Only allow a child announce message to reach the user when:

• the child is blocked
• the child needs approval to continue
• the child hit a major error the parent must surface immediately
• the child itself is the intended final user-facing delivery step

Do not let every child produce routine completion chatter in the user channel.

Atomic vs composite rule

Treat a node as atomic when one agent run can finish it without meaningful internal planning.

Examples of atomic nodes:

• extract top travel dates from the indexed threads
• summarize one PDF
• implement one endpoint
• validate one config file

Treat a node as composite when it clearly breaks into independent workstreams.

Examples of composite nodes:

• reconstruct a work year from exported messages
• build a feature spanning backend, frontend, and tests
• audit a system and produce remediation steps

If a child split feels artificial or the merge plan is vague, stop decomposing.

Workflow

1) Initialize the run

Create a run folder with scripts/init_run.py.

Recommended output layout:

runs/<run-id>/
  tree.json
  events.jsonl
  summary.md
  nodes/
    1/
      spec.json
      notes.md
      result.md

2) Create the root node

Use scripts/upsert_node.py to record the root task and defaults.

Record at minimum:

• id
• parentId
• goal
• type
• depth
• executor
• status
• workspaceMode
• approvalRequired

3) Decompose only when worth it

For each composite node:

• split into 2-5 children
• assign node type and executor
• keep child goals concrete
• define how the parent will merge the results
• stop at depth 2 unless there is a strong reason to go to 3

Never recurse just to make the tree look smart.

4) Persist state and audit events

Before executing a node:

• mark it running
• create or update nodes/<id>/spec.json
• note the intended output in nodes/<id>/notes.md
• append an event to events.jsonl

After executing a node:

• save the result in nodes/<id>/result.md
• mark it completed, failed, or waiting_for_approval
• record artifacts and confidence
• append an event to events.jsonl

Use these scripts:

• scripts/upsert_node.py
• scripts/mark_node.py
• scripts/list_ready_nodes.py
• scripts/list_events.py
• scripts/merge_results.py
• scripts/render_tree.py

5) Execute leaves

Use scripts/list_ready_nodes.py to identify executable leaves.

Execution guidance:

• prefer subagents for analysis/synthesis leaves
• prefer ACP for coding leaves
• prefer direct tools for small ops/browser leaves
• keep concurrency modest; avoid flooding the system with low-value leaves
• when spawning routine child runs, instruct them to keep their work in session/artifacts and reply ANNOUNCE_SKIP during the announce step unless they are blocked or explicitly responsible for final delivery
• do not clean up or delete child sessions until the parent has harvested the needed results

6) Merge upward

When all child nodes under a parent are complete:

• harvest child outputs from child result.md, session history, or other saved artifacts
• use scripts/merge_results.py to bundle child file results when applicable
• create a parent synthesis or review node if needed
• write the merged result to the parent result.md

If a child was run in quiet mode with ANNOUNCE_SKIP, treat that as normal. Silence is not failure; it just means the parent owns user-facing delivery.

Use type-aware merge behavior:

• research → combine evidence, themes, and caveats
• coding → combine patches, test notes, and integration risks
• ops → combine findings, risks, and recommended actions
• review → identify weak claims, contradictions, and missing coverage

Do not silently average conflicting child outputs. Bubble disagreements up.

7) Finish with a skeptical pass

For important runs, add one final review node that:

• challenges overconfident claims
• checks whether decomposition went too far or not far enough
• identifies ambiguous results
• tightens the final summary

Approval gates

Pause and ask before executing any node that would:

• write, edit, move, or delete important files
• install or update software
• restart services
• change config
• send external messages
• perform destructive system actions

Mark these nodes waiting_for_approval instead of forcing them through.

Worktrees

Do not use git worktrees by default.

Use a worktree only when all of these are true:

• the node type is coding
• the target is a git repo
• sibling coding tasks benefit from isolation
• there is a plausible merge path afterward

Otherwise keep the node in normal artifact-folder mode.

Stop conditions

Stop decomposing when any of these is true:

• the node is already atomic
• the merge plan is unclear
• decomposition is becoming repetitive
• depth 2 already produced workable leaves
• time/cost budget would be wasted by further splitting
• adding another node would exceed the run budget

References

Read these only if needed:

• references/tree-schema.json — canonical run and node schema
• references/example-run.md — example run layout and sample tree
• references/quiet-mode.md — quiet child execution pattern using ANNOUNCE_SKIP