mars-llm 0.0.2

Multi-Agent Resolution Synthesis — LLMs debate to find the best answer

MARS — Multi-Agent Resolution Synthesis

Multiple LLMs debate your question through structured rounds of argumentation, critique, and synthesis to produce a single, well-reasoned answer.

Installation

Requires Python 3.11+.

git clone https://github.com/jpollock/multi-agent-resolution-synthesis.git
cd multi-agent-resolution-synthesis
pip install -e .

Copy .env.example to .env and add your API keys:

cp .env.example .env

MARS_OPENAI_API_KEY=sk-...
MARS_ANTHROPIC_API_KEY=sk-ant-...
MARS_GOOGLE_API_KEY=AIza...
MARS_OLLAMA_BASE_URL=http://localhost:11434

Quick Start

# Two providers debate (default: openai + anthropic)
mars debate "What is the best sorting algorithm for nearly-sorted data?"

# Pick specific providers
mars debate "Explain CAP theorem" -p openai -p google

# Include context from files
mars debate @prompt.md -c @context.txt

# Check which providers are configured
mars providers

CLI Reference

mars debate PROMPT [OPTIONS]

Run a multi-LLM debate on PROMPT. PROMPT can be plain text or @file to read from a file.

Option	Default	Description
PROMPT	(required)	Question or @file path
-c, --context		Context text or @file (repeatable)
-p, --provider	openai anthropic	Provider name or provider:model (repeatable)
-m, --mode	round-robin	Debate mode: round-robin or judge
-r, --rounds	3	Maximum debate rounds
-j, --judge-provider		Provider to act as judge (judge mode)
-s, --synthesis-provider		Provider for final synthesis (auto if omitted)
--model		provider:model override (repeatable)
--threshold	0.85	Convergence similarity threshold (0.0-1.0)
--max-tokens	8192	Max output tokens per LLM call
-t, --temperature	(provider default)	Temperature (0.0-2.0)
-v, --verbose	off	Stream responses in real-time
-o, --output-dir	./mars-output	Output directory

mars configure

Set up MARS integration with Claude Code. Installs the /mars:debate slash command so you can run debates from any Claude Code session.

mars configure

Then in Claude Code:

/mars:debate Should we use Redis or Postgres for caching?

mars providers

List configured providers with their default models and configuration status.

Providers

Provider	Env Var	Default Model
openai	MARS_OPENAI_API_KEY	gpt-4o
anthropic	MARS_ANTHROPIC_API_KEY	claude-sonnet-4-20250514
google	MARS_GOOGLE_API_KEY	gemini-2.0-flash
ollama	MARS_OLLAMA_BASE_URL	llama3.2

Override models per-run with -p provider:model or --model provider:model.

Debate Modes

Round-Robin (default)

All providers answer the prompt independently. Each provider then critiques the others' answers and produces an improved response. This repeats until answers converge (similarity exceeds --threshold) or max rounds are reached. A final synthesis step merges the best points into one answer.

Judge

All providers answer independently. A designated judge provider (-j) evaluates every response and produces a final ruling with resolution reasoning.

mars debate "Is Rust better than Go for CLI tools?" \
  -p openai -p anthropic -p google \
  -m judge -j anthropic

Examples

Basic two-provider debate:

mars debate "What are the trade-offs between microservices and monoliths?"

Three providers with model overrides:

mars debate "Design a rate limiter" \
  -p openai -p anthropic -p google \
  --model openai:gpt-4.1 --model anthropic:claude-opus-4-20250514

Using context files:

mars debate @question.md -c @codebase-summary.txt -c @requirements.txt

Tuning convergence and temperature:

mars debate "Optimal database indexing strategy" \
  -p openai -p anthropic \
  --threshold 0.70 -t 0.3 -r 5

Output Structure

Each debate produces a timestamped directory:

mars-output/<timestamp>_<slug>/
├── final-answer.md
└── audit/
    ├── 00-prompt-and-context.md
    ├── 01-round-1-responses.md
    ├── 02-round-2-critiques.md
    ├── 03-round-3-critiques.md
    ├── attribution.md
    ├── costs.md
    ├── round-diffs.md
    ├── convergence.md
    └── resolution.md

File	Contents
final-answer.md	The synthesized final answer
00-prompt-and-context.md	Original prompt and all context
NN-round-N-responses.md	Each provider's response for that round
NN-round-N-critiques.md	Cross-critiques and improved answers
attribution.md	Per-provider contribution, survival, and influence metrics
costs.md	Token counts and estimated cost per provider
round-diffs.md	How each provider's answer changed between rounds
convergence.md	Why the debate stopped (converged or max rounds)
resolution.md	Synthesis reasoning: which points were accepted/rejected

Analysis Output

Attribution

Three metrics per provider, computed via sentence-level similarity:

• Contribution — percentage of final answer sentences traced to this provider (best-match attribution above threshold).
• Survival — percentage of this provider's round-1 sentences that appear in the final answer.
• Influence — rate at which other providers adopted this provider's sentences in subsequent rounds.

Cost Tracking

Token counts (input + output) and estimated USD cost per provider. Pricing uses prefix-matched model lookup (e.g., claude-sonnet-4-20250514 matches claude-sonnet-4 pricing). Ollama models show zero cost.

Claude Code Integration

MARS can be used as a slash command inside Claude Code.

# One-time setup after installing MARS
mars configure

This installs /mars:debate into ~/.claude/commands/, making it available in every Claude Code session. Usage:

/mars:debate What is the best approach to database sharding?
/mars:debate Compare Kubernetes vs Docker Swarm for container orchestration

Claude Code will check your configured providers, run the debate with streaming output, and summarize the result.

Configuration Tips

Temperature: 0.0 for deterministic/factual answers, 0.7 for creative tasks, 1.0+ for experimental diversity. Each provider uses its own default when -t is omitted.

Convergence threshold: Lower values (e.g., 0.70) stop debate sooner when answers are roughly similar. Higher values (e.g., 0.95) force more rounds of refinement. Default 0.85 balances quality and cost.

Synthesis provider: By default, MARS prefers Anthropic then OpenAI for synthesis. Use -s to override.

Retries: All provider calls retry up to 3 times with exponential backoff on transient errors (timeouts, rate limits, 503s).

License

MIT