Orchestrate multi-LLM parallel debate and synthesis. Dispatches prompts to available omega CLI wrappers in parallel, collects independent responses, runs anonymized peer review ranking, and synthesizes via a chairman model. Supports collaboration templates, idle watchdog, inter-agent messaging, worktree isolation, and multi-turn sessions. No server required.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Orchestrate multi-LLM parallel debate and synthesis. Dispatches prompts to available omega CLI wrappers in parallel, collects independent responses, runs anonymized peer review ranking, and synthesizes via a chairman model. Supports collaboration templates, idle watchdog, inter-agent messaging, worktree isolation, and multi-turn sessions. No server required.
["Always check available CLIs before dispatching (run verify-setup.mjs for each)","Use at least 3 models for meaningful peer review rankings","Set --timeout-ms per model to prevent one slow model from blocking the council","Anonymize responses before peer review to prevent model identity bias","Store council results in .claude/context/tmp/ for downstream consumption"]
error_handling
graceful
streaming
not_supported
verified
true
lastVerifiedAt
"2026-02-22T00:00:00.000Z"
source
builtin
trust_score
100
provenance_sha
5760eb2ced34db2b
LLM Council Skill
Multi-LLM deliberation orchestrator. Implements a 3-stage council process:
Stage 1 (parallel independent responses) -> Stage 2 (anonymized peer review with ranking) ->
Stage 3 (chairman synthesis). Uses omega CLI wrapper skills for model invocation.
No server required -- pure Bash-level orchestration.
- Parallel dispatch to multiple LLM CLIs (Gemini, Codex, Cursor, Claude)
- Dynamic availability detection (only dispatches to available CLIs)
- 3-stage council protocol:
- Stage 1: Independent parallel responses
- Stage 2: Anonymized peer review with FINAL RANKING
- Stage 3: Chairman synthesis of all inputs
- Anonymized labeling (Response A, B, C...) to prevent model identity bias
- Configurable chairman model (default: current Claude session)
- Per-model timeout enforcement
- Graceful degradation (skips unavailable models)
- Structured output with all 3 stages preserved
Council Protocol
Stage 1: Independent Responses
Dispatch the user's prompt to all available omega CLIs in parallel:
Send all anonymized responses back to each available model with:
You are reviewing responses to this question: "$PROMPT"
Response A:
[content]
Response B:
[content]
Response C:
[content]
Evaluate each response for accuracy, completeness, and reasoning quality.
Then provide your FINAL RANKING (best to worst):
1. Response [X]
2. Response [Y]
3. Response [Z]
Parse FINAL RANKING from each reviewer's output using regex: \d+\.\s*Response [A-Z]
Compute aggregate ranking (average position across all reviewers)
Stage 3: Chairman Synthesis
Send to chairman model (default: current Claude session or ask-claude.mjs):
You are the Chairman synthesizing a multi-model council discussion.
Original question: "$PROMPT"
Stage 1 Responses (with de-anonymized model names):
[Model]: [Response]
...
Stage 2 Peer Review Rankings:
Aggregate ranking: [best to worst with scores]
Synthesize the best insights from all responses into a single comprehensive answer.
Highlight areas of consensus and dissent. Provide the strongest possible answer.
Usage
Full council (all available models)
Skill({ skill: 'llm-council' })
# Then in agent: run full council protocol above
Quick consultation (skip peer review)
# Stage 1 only -- parallel dispatch, collect responses, skip ranking# Use when speed matters more than rigorous evaluation
Code review where diverse model viewpoints reduce blind spots
Plan critique and validation
Resolving disagreements about implementation approach
Cross-validation of security analysis
When the user explicitly requests "council", "multiple perspectives", or "cross-validate"
Iron Laws
ALWAYS check CLI availability before dispatch -- never assume a model is present
ALWAYS anonymize responses before peer review -- model identity bias is real
NEVER skip Stage 2 for high-stakes decisions -- the peer review is the core innovation
ALWAYS preserve all 3 stages in output for transparency
ALWAYS set per-model timeout to prevent one slow model from blocking the council
Anti-Patterns
Anti-Pattern
Why Bad
Correct Approach
Dispatching to unavailable CLIs
Silent failure, missing responses
Run verify-setup.mjs first
Showing model names during peer review
Introduces identity bias
Use anonymous labels (Response A, B, C)
Using only 2 models for council
Peer review meaningless with 2
Require minimum 3 for ranking value
Ignoring failed model responses
May miss degradation
Log failures, include in metadata
Running council for simple questions
Massive overhead for trivial queries
Reserve for high-stakes decisions
Exit Codes
Code
Meaning
0
Council completed successfully (at least 2 models responded)
1
Council failed (fewer than 2 models available or all timed out)
Configuration
Env Var
Default
Purpose
LLM_COUNCIL_TIMEOUT_MS
120000
Per-model timeout for Stage 1 and Stage 2
LLM_COUNCIL_MIN_MODELS
2
Minimum models required for council to proceed
LLM_COUNCIL_CHAIRMAN
(current session)
Chairman model for Stage 3 synthesis
Integration Notes
This skill does NOT require the llm-council-master FastAPI server
Uses omega wrapper scripts directly via Bash backgrounding for parallelism
Anonymized peer review is preserved from llm-council-master's design (its core innovation)
Chairman synthesis can use the current Claude session (no additional CLI call needed)
Temporary files stored in system temp dir, cleaned up after council completes
Minimum 2 available models required for council to proceed
Collaboration Templates
Templates constrain each model's focus during council sessions, producing higher-quality synthesis than sending the same generic prompt to all models. Without a --template flag, the council operates in its default mode (all models receive the same prompt).
Available Templates
Review Template
Field
Value
Agents
2-3 (minimum 2)
Focus
Each agent reviews from their model's strengths (correctness, security, performance)
roles:-name:correctness-reviewerfocus:'Review for logic errors, edge cases, off-by-one bugs, and correctness'-name:performance-reviewerfocus:'Review for performance bottlenecks, algorithmic complexity, and scalability'-name:security-reviewerfocus:'Review for security vulnerabilities, injection vectors, and data exposure'synthesis_strategy:'merge-by-category'
When to use: Code reviews, PR reviews, audit passes where multiple review dimensions matter.
Implementation Template
Field
Value
Agents
2-4 (minimum 2)
Focus
Plan, execute, verify staged workflow with sequential handoff
Synthesis
Sequential -- architect output feeds implementer, verifier checks result
Roles:
roles:-name:architectfocus:'Design the approach, define interfaces, data flow, and module boundaries'-name:implementerfocus:"Write the implementation following the architect's design exactly"-name:verifierfocus:'Verify the implementation matches the design and passes acceptance criteria'synthesis_strategy:'sequential'
When to use: Feature implementation where design and coding benefit from separation of concerns.
Research Template
Field
Value
Agents
2-3 (minimum 2)
Focus
Independent investigation of the same topic, then cross-compare findings
Synthesis
Side-by-side comparison matrix highlighting contradictions and agreements
Roles:
roles:-name:researcher-afocus:'Research the topic independently, cite sources, provide evidence-backed findings'-name:researcher-bfocus:'Research the same topic independently from a different angle, cite sources'synthesis_strategy:'compare-and-converge'
When to use: Technology evaluation, best-practice research, exploring solution spaces.
Debug Template
Field
Value
Agents
2-3 (minimum 2)
Focus
Each agent independently diagnoses the same bug and proposes a fix
Synthesis
Convergence analysis -- if 2+ agents agree on root cause, high confidence
Roles:
roles:-name:diagnostician-afocus:'Independently reproduce and diagnose the bug, propose root cause and fix'-name:diagnostician-bfocus:'Independently reproduce and diagnose the bug, propose root cause and fix'synthesis_strategy:'convergence'
When to use: Hard-to-diagnose bugs where independent analysis reduces bias.
convergence: Report agreement/disagreement on root cause, confidence score
Council Watchdog
Per-model idle monitoring with two-tier thresholds to detect hung or stalled models during council sessions.
Two-Tier Monitoring
Tier
Threshold
Action
Idle Warning
90 seconds of no output
Log warning, optionally send nudge prompt
Stall Timeout
180 seconds of no output
Terminate model process, exclude from results
Configuration
watchdog:idle_warning_seconds:90stall_timeout_seconds:180nudge_prompt:'Please continue with your analysis.'action_on_stall:'exclude'# exclude | retry | fail
Env Var
Default
Purpose
LLM_COUNCIL_IDLE_WARNING_S
90
Seconds before idle warning
LLM_COUNCIL_STALL_TIMEOUT_S
180
Seconds before stall auto-exclude
Integration with Stage 1
During parallel dispatch, each backgrounded model process is monitored independently:
Start a per-model timer when the process launches
If no output file growth after idle_warning_seconds, log a warning
If no output file growth after stall_timeout_seconds, kill the process and set stall_timeout: true in council metadata
Continue council with remaining models (graceful degradation)
The watchdog is additive to the existing --timeout-ms global timeout. --timeout-ms remains the hard ceiling for the entire council session. The watchdog provides per-model granularity within that ceiling.
Inter-Agent Message Bus
Optional JSONL message file protocol that enables council members to share partial outputs during deliberation.
Overview
By default, council models work in complete isolation during Stage 1. The message bus enables optional partial-output sharing, useful for long deliberations where early findings from one model can inform others.
Opt-In
Enable with --enable-messaging flag. Without this flag, no message directory is created and the council operates in default isolated mode.
{"id":"msg-001","session_id":"council-2026-03-21-abc123","from":"gemini","to":"team","type":"partial_output","content":"Partial finding: JWT implementation has XSS risk in refresh flow","timestamp":"2026-03-21T10:30:00Z"}
Message Types:
Type
Purpose
When Sent
partial_output
Share an intermediate finding or observation
During Stage 1, when a model has a partial result
question
Ask the team a clarifying question
During Stage 1, when a model needs input
agreement
Signal agreement with another model's finding
During Stage 2, after reading peer outputs
disagreement
Signal disagreement with reasoning
During Stage 2, after reading peer outputs
Read/Write Protocol
Write: Each model appends messages to messages.jsonl during Stage 1 execution
Read (optional): Models can read messages.jsonl between Stage 1 and Stage 2 for informed review
Chairman read: Chairman reads all messages during Stage 3 synthesis for additional context
Cleanup
The session directory (.claude/context/tmp/council-<session-id>/) is deleted after the council completes, including the messages file.
Worktree Isolation
Optional per-agent git worktree isolation for council sessions where agents modify code.
When to Use
Worktree isolation is relevant only for templates that modify code:
Template
Worktree Applicable?
Reason
Review
No
Read-only analysis
Implementation
Yes
Agents write code that may conflict
Research
No
Read-only research
Debug
Yes
Agents may apply experimental fixes
Opt-In
Enable with --use-worktrees flag. Only takes effect when combined with --template implementation or --template debug.
Worktree Lifecycle
Create: For each participating agent, create a worktree:
When worktrees have overlapping changes to the same files:
Chairman identifies conflicting hunks from each agent's diff
Chairman selects the best version based on Stage 2 peer review rankings
If ranking is inconclusive, chairman presents both versions with rationale for human decision
Platform Note
Worktree isolation follows existing agent-studio worktree safety rules. On Windows, ensure paths stay under the OS path length limit. Reference .claude/rules/ for worktree safety patterns.
Multi-Turn Council Sessions
Persistent session state that allows council deliberations to span multiple turns with user feedback between rounds.
Overview
By default, each council invocation is one-shot. Multi-turn sessions enable iterative refinement: the user reviews the synthesis, provides feedback, and the council continues deliberation with the additional context.
Opt-In
--multi-turn: Create a resumable session (generates a session ID)
--resume SESSION_ID: Continue a previous session with new user input
{"session_id":"council-2026-03-21-abc123","status":"active","turn_count":2,"max_turns":5,"original_prompt":"Review this auth implementation","template":"review","models":["gemini","codex","claude"],"turns":[{"turn":1,"stage_1_responses":{"gemini":"...","codex":"...","claude":"..."},"stage_2_rankings":{"aggregate":["gemini","claude","codex"]},"stage_3_synthesis":"...","user_feedback":"Focus more on the JWT refresh flow"}],"created_at":"2026-03-21T10:00:00Z","last_active":"2026-03-21T10:15:00Z"}
Continuation Protocol
When --resume SESSION_ID is used:
Load session.json from the session directory
Verify session is not expired (24-hour TTL from last_active)
Include previous turns' synthesis and user feedback in the next Stage 1 prompt:
Previous council synthesis: [Stage 3 output from last turn]
User feedback: [feedback text]
Continue the analysis with this additional context.
Models receive full conversation history for context continuity
Increment turn_count and update last_active
Configuration
Env Var
Default
Purpose
LLM_COUNCIL_MAX_TURNS
5
Maximum turns per session before forced closure
Session Expiry
Sessions older than 24 hours are considered expired. Attempting to --resume an expired session returns an error with the session's last synthesis as context.
Cleanup
Session directories are cleaned up when:
The session reaches max_turns
The user does not resume within 24 hours
The user explicitly closes the session (no --resume after final turn)
