conclave v3

Multi-LLM deliberation protocol engine — like a conclave of AI cardinals.

V3 is a complete rewrite. It's not just "ask two models and compare" — it's a multi-round deliberation protocol where models critique each other, revise their answers, and a judge synthesizes a consensus report.

What's new in v3

• 🔄 Multi-round deliberation — Initial → Critique → Revise → Synthesize
• 🔌 Plugin provider system — CLI (Claude Code/Codex) + API (OpenAI/Anthropic/DeepSeek)
• 🧠 Semantic synthesizer — LLM-powered judge generates consensus/divergence/insight report
• 📡 Streaming — Real-time output for each provider
• ⚙️ Config system — ~/.conclave/config.yaml
• 🧪 Test suite — 22 tests, CI via GitHub Actions

Install

pip install conclave-llm

Prerequisites

conclave doesn't manage API keys. It connects through:

Quick Start

# Initialize config
conclave config init

# Ask 2 models the same question
conclave "Should we use Redis or Postgres for this caching layer?"

# Multi-round deliberation with synthesis
conclave "Design a rate-limiting strategy for our API" --rounds 3

# Use specific providers
conclave "Review this architecture" --providers claude,deepseek

# JSON output
conclave "Compare Kubernetes vs Nomad" --format json --rounds 2

# Prompt from a file (long / multiline specs — no shell pipes needed)
conclave run --prompt-file ./spec.md --providers claude,codex --rounds 1

# Persist result to disk (JSON is un-truncated; markdown is ANSI-free)
conclave run --prompt-file ./spec.md \
    --providers claude,codex \
    --format json --output ./deliberation.json

--prompt-file and --output

• --prompt-file PATH reads the prompt from a UTF-8 file. Mutually exclusive with the positional PROMPT. Empty/missing/non-UTF-8/directory inputs raise a clear error.
• --output PATH writes the full result to disk.
  • --format json writes the complete DeliberationResult as JSON (response text is not truncated, unlike the terminal preview).
  • --format markdown writes plain, ANSI-free markdown.
  • If writing fails (permissions, missing parent dir, …) conclave prints a warning but keeps the deliberation exit code.
• Passing the same path to --prompt-file and --output is rejected so the input can never be clobbered.

Migrating from hermes_fusion.py

The old fusion runner used echo "$SPEC" | python3 hermes_fusion.py …, which Hermes Tirith blocks as pipe_to_interpreter. Replace it with a --prompt-file call — no stdin pipe, no interactive TUI:

# ❌ blocked by Hermes and hangs on multiline prompts
echo "$SPEC" | python3 ~/.hermes/scripts/hermes_fusion.py --providers claude,codex

# ✅ conclave equivalent
conclave run --prompt-file ./spec.md \
    --providers claude,codex --rounds 1 \
    --format json --output ./out.json

The default Claude preset now runs claude -p "{prompt}" --model {model}, so long or multiline prompts no longer drop into the interactive TUI and hang, and the model: you set in ~/.conclave/config.yaml is passed through to the Claude CLI via --model instead of being a metadata-only label.

Deliberation Protocol

Round 1 (INITIAL):  All models answer independently
Round 2 (CRITIQUE):  Each model critiques others' answers
Round 3 (REVISE):    Models revise based on feedback
       ↓
    SYNTHESIS:       Judge generates consensus report:
                     ✅ Consensus points
                     🔀 Divergences
                     💡 Unique insights

Provider System

conclave v3 supports any backend through a provider plugin system:

# ~/.conclave/config.yaml
providers:
  claude:
    provider_type: cli
    model: claude-fable-5
    executable: claude
  openai:
    provider_type: openai
    model: gpt-5.6-sol
    api_key: ${OPENAI_API_KEY}

Fable 5 必须写全名 claude-fable-5；别名 fable 可能不可用。

Architecture

Layer 4: CLI (click + rich)
Layer 3: Protocol Engine (DeliberationEngine)
Layer 2: Scheduler (parallel ThreadPoolExecutor)
Layer 1: Providers (CLI / OpenAI / Anthropic / DeepSeek)
Layer 0: Config + Cache + Formatters

Development

git clone https://github.com/doctorzero666/conclave
cd conclave
pip install -e ".[dev]"
pytest       # 22 tests
ruff check   # lint

License

MIT — Zhichao Jiang