| name | swarm-intelligence |
| description | Swarminator orchestration: Mode Single-Node (one bounded external node, immutable artifact) or Mode Full Swarm (multi-phase multi-persona multi-model). Use for isolated research/second opinions/patch suggestions, or for diverse-perspective cross-validation when risk is too high for one pass. Activate on "swarm", "multi-agent", "external delegated node", "single-node", "swarminator", "second opinion", "cross-validate", "parallel agents". Commands: /swarm → Full Swarm, /external-subagent → Single-Node. Skip for routine single-agent implementation. |
Swarm Intelligence (Swarminator)
One skill, two modes. Choose the mode first — compliance and preflight differ. Commands pin the mode: /external-subagent → Single-Node, /swarm → Full Swarm.
Loading this skill binds you to the chosen mode’s full procedure. Do not run a “lite swarm” that is neither a locked single node nor a full three-phase swarm. Real swarminator invocations only — no prose simulations. See rules/skill-compliance.md for Full Swarm hard gates.
Compose with subagent-dispatch for prompt contracts before any node launch.
Mode router (mandatory)
| Signal | Mode | Cost |
|---|
/external-subagent command, or one concrete deliverable / one node / research / second opinion / review / patch suggestion | Single-Node | low |
| User asks for one isolated delegated node / worker to offload a small task | Single-Node | low |
/swarm command, or multi-phase / ambiguous / high-stakes / quorum / challenge cycles | Full Swarm | high |
| User says swarm, multi-agent, diverse perspectives, cross-validate | Full Swarm | high |
| Unsure | Prefer Single-Node; escalate to Full Swarm only if the artifact is still insufficient after one tight retry | — |
Is the deliverable one bounded artifact from one model context?
YES → Mode Single-Node (references/single-node.md)
NO → Mode Full Swarm (references/preflight-and-phases.md)
Do not slowly grow Single-Node into an ad-hoc multi-node mini-swarm. Escalate modes explicitly.
Shared runtime
Discover CLI at runtime — do not hard-code stale flags:
$SHELL -l -c 'command -v swarminator'
$SHELL -l -c 'swarminator --help'
$SHELL -l -c 'swarminator --list-agents'
$SHELL -l -c 'swarminator --list-models --agent NAME'
Shared assets (both modes):
- Personas:
references/personas/ via references/discover-personas.sh
- Models:
references/models/{free,premium}.json
- Default node preference:
command-code + deepseek-v4-pro while budget allows
Shared hard rules (both modes):
- Delegated nodes are read-only on the workspace: no file writes, patch apply, stage, or destructive commands.
- Host defines contract (scope, output, allowed files, stop conditions) before invoke.
- Host validates and applies any changes.
- Prefer
subagent-dispatch pillars in every node prompt.
Mode Single-Node
Goal: one task → one swarminator node → one immutable artifact.
Procedure: references/single-node.md
Artifact modes: analysis | review | patch | transform.
Minimum flow: lock scope + artifact mode + allowed files → one node → validate artifact → host applies if needed.
Retry: fresh bounded invocation with a tighter contract — not automatic multi-node expansion. If scope is unbounded, switch to Full Swarm.
Mode Full Swarm
Goal: multi-persona, multi-model, three-phase orchestration with synthesis between phases.
Preflight: all 10 steps required (user confirm, domain, deliverable, read-only OK, swarminator checks, catalogs, personas).
Detail: references/preflight-and-phases.md
Models: 2–3 models per persona each phase — single-model shortcuts forbidden.
| Phase | Purpose | Required sections (summary) |
|---|
| 1 Gather | Inputs + findings | Goals, Key Inputs, Extracted Findings, Open Questions, Synthesis Log |
| 2 Synthesize | Spec + challenge | Spec Summary, Design Decisions, Acceptance Criteria, Task Hints, Synthesis Log |
| 3 Produce | Decompose + deliver | Executive Summary, Task Results, Blocked Tasks, Open Items, Synthesis Log |
Forbidden: describing what the swarm would produce; skipping preflight; “simplified” single-model swarm.
Composition
| With | How |
|---|
subagent-dispatch | Build every node prompt (scope, output contract, obstacles, allowed actions) before swarminator |
bounded-iteration | Full Swarm (or design) first → locked TASK.md → iteration loop (bounded-iteration/references/swarm-integration.md) |
reviewer | Single-Node review artifact mode, or Full Swarm review personas |
model-benchmarking | Choose catalogs/tiers → then this skill for execution |
| Environment-native task workers | Prefer native workers when available; use this skill when you need swarminator isolation or multi-model quorum |
Deliverable checklist
Either mode
Single-Node
Full Swarm
