LLM Council Skill

Multi-model council: parallel drafts → adversarial critique → validated synthesis.

Prerequisite: This skill requires the the-llm-council Python package to be installed. The skill provides IDE integration but the actual council runs via the installed CLI. If you see command not found: council, run pip install the-llm-council first.

Setup

1. Install

pip install the-llm-council>=0.5.0

# With specific provider SDKs
pip install the-llm-council[anthropic,openai,google]

2. Configure API Keys

Provider	Environment Variable	Notes
OpenRouter	OPENROUTER_API_KEY	Recommended - single key for all models
OpenAI	OPENAI_API_KEY	Direct GPT access
Anthropic	ANTHROPIC_API_KEY	Direct Claude access
Google	GOOGLE_API_KEY or GEMINI_API_KEY	Direct Gemini access

# Minimum setup (OpenRouter)
export OPENROUTER_API_KEY="your-key"

3. Verify

council doctor

Usage

council run <subagent> "<task>" [options]

CLI Options

Option	Description
--mode	Agent mode (e.g., impl/arch/test for drafter)
--json	Output structured JSON
--verbose, -v	Verbose output
--models, -m	Comma-separated model IDs
--providers, -p	Comma-separated provider list
--no-artifacts	Disable artifact storage (faster)

Other Commands

council doctor    # Check provider health
council config    # Show current configuration

Subagents (v0.5.0)

Core Agents

Subagent	Modes	Use For	Details
drafter	impl, arch, test	Code, architecture, tests	See below
critic	review, security	Code review, security audit	See below
synthesizer	-	Merge and finalize outputs	See subagents/synthesizer.md
researcher	-	Technical research	See subagents/researcher.md
planner	plan, assess	Roadmaps, decisions	See subagents/planner.md
router	-	Task classification	See subagents/router.md

Agent Modes

drafter modes:

• --mode impl - Feature implementation, bug fixes (default)
• --mode arch - System design, API schemas
• --mode test - Test suite design

critic modes:

• --mode review - Code review with CWE IDs (default)
• --mode security - Security threat analysis

planner modes:

• --mode plan - Execution roadmaps (default)
• --mode assess - Build vs buy decisions

Deprecated Aliases (Backwards Compatible)

The following legacy agent names still work but will be removed in v1.0:

Old Name	Use Instead	Removed In
implementer	drafter --mode impl	v1.0
architect	drafter --mode arch	v1.0
test-designer	drafter --mode test	v1.0
reviewer	critic --mode review	v1.0
red-team	critic --mode security	v1.0
assessor	planner --mode assess	v1.0
shipper	synthesizer	v1.0

Multi-Model Configuration

Run multiple models in parallel for adversarial debate:

# Via CLI flag
council run drafter --mode arch "Design caching layer" \
  --models "anthropic/claude-3.5-sonnet,openai/gpt-4o,google/gemini-pro"

# Via environment variable
export COUNCIL_MODELS="anthropic/claude-3.5-sonnet,openai/gpt-4o,google/gemini-pro"

Model Pack Overrides

Fine-tune which models handle specific task types:

export COUNCIL_MODEL_FAST="anthropic/claude-3-haiku"      # Quick tasks
export COUNCIL_MODEL_REASONING="anthropic/claude-3-opus"  # Deep analysis
export COUNCIL_MODEL_CODE="openai/gpt-4o"                 # Code generation
export COUNCIL_MODEL_CRITIC="anthropic/claude-3.5-sonnet" # Adversarial critique

Config File

Optional YAML configuration:

# ~/.config/llm-council/config.yaml
providers:
  - name: openrouter
    api_key: ${OPENROUTER_API_KEY}
    default_model: anthropic/claude-3-opus

defaults:
  providers:
    - openrouter
  timeout: 120
  max_retries: 3
  summary_tier: actions

Python API

from llm_council import Council
from llm_council.protocol.types import CouncilConfig

config = CouncilConfig(
    providers=["openrouter"],
    mode="impl"  # Optional: set agent mode
)
council = Council(config=config)
result = await council.run(
    task="Build a login page with OAuth",
    subagent="drafter"
)
print(result.output)

When to Use

Use council for:

• Feature implementation requiring production quality
• Code review with security analysis (CWE IDs)
• Architecture design decisions
• Technical research informing decisions
• Build vs buy assessments
• Security threat modeling

Skip council for:

• Quick file lookups
• Single-line fixes
• Simple questions

Examples

# Feature implementation (new v0.5.0 syntax)
council run drafter --mode impl "Add pagination to users API" --json

# Code review
council run critic --mode review "Review the authentication changes" --json

# Multi-model architecture design
council run drafter --mode arch "Design caching layer" \
  --models "anthropic/claude-3.5-sonnet,openai/gpt-4o" --json

# Security threat model
council run critic --mode security "Analyze auth system vulnerabilities" --json

# Build vs buy decision
council run planner --mode assess "Should we build or buy a payment system?" --json

# Legacy syntax (still works, shows deprecation warning)
council run implementer "Add pagination" --json
council run reviewer "Review changes" --json

Security Notes

• API Keys: Never embed secrets in task descriptions or skill files. Use environment variables.
• Data Sensitivity: Avoid sending files containing secrets (.env, credentials) to the council. Context is sent to external LLM providers.
• Skill Integrity: Treat SKILL.md and subagents/*.md as configuration code. Keep under version control.

Troubleshooting

# Check all providers
council doctor

# Verbose output for debugging
council run drafter --mode impl "task" --verbose

# Faster runs (skip artifact storage)
council run drafter "task" --no-artifacts