Skip to content

AGENTS.md

Latest commit

 

History

History
92 lines (75 loc) · 5.93 KB

AGENTS.md

File metadata and controls

92 lines (75 loc) · 5.93 KB

Agents

Overview

Swarm Command ships one Copilot CLI skill and one standalone agent. The skill (swarm-command) is the primary entry point — it orchestrates the full multi-model consensus swarm. The agent (swarm-command.agent.md) is a standalone agent version of the same capability for use when you prefer the agent selector over skill triggers.

Available Agents

swarm-command (Skill)

  • Purpose: Multi-model swarm orchestrator with consensus scoring and Shadow Score Spec L2 conformance. Launches 50–250+ AI agents across 15 models in a 5-layer hierarchy (Nexus → Commander → Squad Lead → Worker → Reviewer), performs cross-family review, validates outputs against sealed acceptance criteria, and synthesizes the final output through a 4-stage consensus algorithm.
  • Trigger: swarm command
  • Usage: Install via Copilot CLI, then invoke with natural language: 
    swarm command                           # Default SS-100 deployment
swarm command ss-250                    # Full 250+ agent deployment
swarm command ss-50                     # Quick 50-agent deployment
swarm command ss-250 "refactor auth"    # With specific task
  • Model: Dispatches to 15 models across Claude and GPT families
  • Install: 
    mkdir -p ~/.copilot/skills/swarm-command && \
  curl -sL https://raw.githubusercontent.com/DUBSOpenHub/swarm-command/main/skills/swarm-command/SKILL.md \
    -o ~/.copilot/skills/swarm-command/SKILL.md
  • Location: skills/swarm-command/SKILL.md

swarm-command (Agent)

  • Purpose: Standalone agent version of the swarm orchestrator. Same orchestration logic as the skill but framed as a named agent for use in the Copilot CLI agent selector. Use when you want to invoke swarm orchestration via the @swarm-command mention or agent picker rather than the swarm command trigger phrase.
  • Usage: After installing ~/.copilot/agents/swarm-command.agent.md, invoke via the agent selector in Copilot CLI.
  • Install: 
    mkdir -p ~/.copilot/agents && \
  curl -sL https://raw.githubusercontent.com/DUBSOpenHub/swarm-command/main/agents/swarm-command.agent.md \
    -o ~/.copilot/agents/swarm-command.agent.md
  • Location: agents/swarm-command.agent.md

File Ownership

File/Directory Owner Purpose
skills/swarm-command/SKILL.md Skill Core orchestration logic
agents/swarm-command.agent.md Agent Standalone agent wrapper
templates/commander.md Skill (reference) Commander prompt template
templates/worker.md Skill (reference) Worker prompt template
templates/reviewer.md Skill (reference) Cross-Reviewer prompt template
templates/squad-lead.md Skill (reference) Squad Lead prompt template
protocols/depth-guard.md Skill (reference) Depth Guard protocol
protocols/circuit-breaker.md Skill (reference) Circuit Breaker protocol
protocols/context-capsule.md Skill (reference) Context Capsule schemas
protocols/meta-reviewer.md Skill (reference) Reviewer quality gate protocol
.github/workflows/ci.yml CI CI validation (YAML + SKILL.md parity)
config.yml Both Shared configuration
catalog.yml Meta Skill metadata for discovery
docs/example-output.md Documentation Representative completed run output
docs/*.md Documentation Architecture, consensus, scaling

Configuration

Reference vs Operational Files

The templates/ and protocols/ directories contain reference documentation — canonical prompt templates and protocol specifications. The SKILL.md embeds this logic inline for self-contained execution. These files exist for:

  • Contributors who want to understand or modify the prompt templates

  • Documentation of the protocols that SKILL.md implements

  • Copy-paste source material when building custom deployments

  • No API keys required — all models are accessed through your active Copilot subscription

  • No servers or infrastructure — swarm state lives in memory during the session

  • Scaling modes: SS-50 (starter), SS-100 (default), SS-250 (full consensus swarm)

  • Consensus threshold: 70% for auto-accept, 50% for majority, below 50% for Nexus arbitration

  • Shadow scoring: Shadow Score Spec L2 — sealed acceptance criteria generated before execution, validated after, hardened on failure

  • Cost ceiling: $20 hard cap with automatic kill-switch

  • Timeout cascade: SS-50: 60/40/30s · SS-100: 75/50/35/25s · SS-250: 90/60/40/30s (children always finish before parents)

Agent Prompt Rules

Terminology Note

  • "task tool": The Copilot CLI tool that launches sub-agents (task() function). Used by Commanders to spawn workers.
  • "task" agent type: A lightweight agent type (Haiku model, all tools, no sub-agent spawning). Used for workers that need bash/command execution.
  • Depth Guard rule: "DO NOT use the task tool" means "do not spawn sub-agents" — it does NOT mean "do not be a task-type agent."
  1. Depth Guard is non-negotiable — Workers MUST be told "DO NOT use the task tool. You are a leaf node." Workers are ALWAYS explore or task agent type, never general-purpose.
  2. Parent controls spawning — The parent computes can_launch for every child. The child never decides for itself.
  3. Context compresses at every layer — 128K → 2K → 512 → 128 tokens. Strip rationale, keep only the task.
  4. All outputs are strict JSON — No prose, no markdown, just schema-validated JSON at every layer boundary.
  5. Pipeline overlap — Reviewers start as soon as the first 2 commander bundles arrive, don't wait for all 5.
  6. Canary before swarm — Deploy 1 canary worker per pod before full deployment to verify task feasibility.
  7. Fail parsably — Every failure is a structured JSON object with status, reason, and diagnostics.