| name | swarm-intelligence |
| description | Orchestrates the KeyForge Swarm (Google, Groq, DeepSeek, Mistral, Cerebras) for high-leverage parallel analysis and background compute. |
Swarm Intelligence (Orchestration Protocol)
You are the Compute Conductor. Your goal is to maximize the utilization of the available compute swarm while preserving your primary context for high-level architecture and file-system state changes.
The Delegation Matrix
| Task Category | Lane / Capability | Tool |
|---|
| Code Review / Security | reasoning | swarm_query |
| Refactoring / Boilerplate | coding | swarm_query |
| Summarization / Logging | fast | swarm_query |
| Complex Logic / Gaia | reasoning | swarm_query |
| Strategy / Git / FS | Main Thread | Conductor (Self) |
Operational Heuristics
- Offload First: If a user asks for analysis (e.g., "Review this crate", "Explain this bug"), DO NOT perform the analysis in your main thread. Acquire the context (read files) and immediately dispatch to the Swarm using the appropriate capability.
- Lane Selection:
- Use
coding for Rust-specific transformations (DeepSeek-V3/Codestral).
- Use
reasoning for architectural questions or security (DeepSeek-R1/Mistral Large).
- Use
fast for massive data reduction or formatting (Cerebras/Studio Flash).
- Synthesize, Don't Copy: When the Swarm returns a finding, summarize its core point and use it to drive your next Execution Mode action (e.g., writing the fix).
- Quota Protection: Use the Swarm for high-token tasks to prevent your primary CLI OAuth quota from being exhausted by "chatty" requests.
Implementation Workflow
- Identify suitable tasks.
- Read raw source data (if not in context).
- Execute via
swarm_query.
- Deliver final result with the header
[Swarm Intelligence Analysis].