swarm git:20261014.6c60297A

Immutable. This exact content never changes and is served at /api/v1/blob/c262c46f7f6043f0.

---
name: swarm
description: Delegate an approved finite batch of independent engineering or research responsibilities using the operator's existing worker channel. Use for broad repository discovery, multiple search directions, repeated checks and substantial dirty work; keep coupled single tasks with one worker and explicit solo work in the main session.
---

# Bounded swarm

Use worker-routing for the coordinator/worker contract and channel authority. This method
expands already-authorized independent responsibilities; it grants no new provider, data,
workspace, account or cost authority. One configured route may serve multiple sessions,
but its shared state must actually support that arrangement. Workers remain peers and
cannot recursively route or create a reviewer hierarchy.

## Decide the batch before launch

State the finite outcome, criteria, candidate/input coverage, independent responsibility
list, exact read/write roots and shared resources. Small repeated operations can form one
complete script-based responsibility. Do not first investigate every candidate yourself.
Keep an internally coupled implementation with one worker; split only where execution and
acceptance are meaningfully independent. Batch size is an execution choice, not a reason
to expand the goal or require the user to configure each worker manually.

Use the operator's current authorized channel. ACP follows acp-worker with explicit current
permissions; native uses the live host tools and their existing boundaries. Use the actual
finite-batch interface the installed integration exposes, not a fabricated slash command.
One authored order spec mechanically becomes the order files and manifest; nothing needs
hand-assembling per item. Read [batch execution and collection](references/batch.md) before
launching a batch.

## Keep useful work on the worker side

Each worker owns a complete local outcome: investigate, implement if authorized, check,
repair ordinary defects and deliver evidence. A finished slot may receive the next item
already in the approved list; do not wait for the slowest unrelated item before proceeding.
Throttle starts when collection or acceptance backs up. A blocked item does not reclaim all
other execution. Missing required items mean partial, even when most receipts completed.

If a worker needs a decision, supply the smallest useful increment and return execution.
If only its report is deficient, consolidate existing findings rather than repeat research.
Correct evidenced caller/permission/environment mistakes under existing authority. Preserve
supported artifacts, original identity and stop/cleanup facts before rework or replacement.

## Accept a batch without repeating all discovery

Collect bounded summaries and expandable artifacts. Mechanical collection can check identity,
duplicates, coverage, missing/stale evidence and source locations; it cannot prove semantics.
Settled results are readable while the rest of the batch still runs; each carries a bounded
worker-reported excerpt and candidate next actions that stay hints, never auto-executed.
Keep ready/partial/needs_input declarations, runtime status and coordinator acceptance
separate. A result with unknown evidence stays unknown; an access failure is not no-support.

Read key conclusions with enough supporting context, coverage, counterevidence and unresolved
questions. Expand or delegate only doubts that can change the decision. If cross-result
reasoning warrants it, give one peer merger a bounded responsibility over collected packages
and necessary evidence; no default review tier. It resolves identity/version distinctions,
preserves source lineage and sends exact missing questions back to their original workers.

For parallel code edits, verify the combined final state and shared interfaces. Reuse current
valid checks, and identify the narrow old evidence invalidated by a change. Tests do not cover
unstated semantics. A high-risk permission/data/lifecycle boundary may need directed main
review; neither worker confidence nor reviewer count substitutes for evidence.

Report covered/partial/unprocessed inputs, useful decisions and exact unresolved gaps. Public
sharing uses only its existing allowlist; batch reports, orders, paths and detailed artifacts
remain private local state. No mandatory review stamps or new task database are required.