whalecouncil 0.1.0

Independent second opinions for AI agent workflows.

WhaleCouncil

Independent second opinions for AI agent workflows.

When you use a single AI model, you get a single perspective — one that can anchor on wrong assumptions, miss edge cases, or confidently hallucinate. WhaleCouncil sends your task to multiple models in parallel, collects independent responses, surfaces disagreements, and synthesizes a final report.

$ council review "Should I use Redis or Postgres for this session store?"

WhaleCouncil — reviewing with: claude, codex, gemini
────────────────────────────────────────────────────
claude   ✓  Redis for session data. Low latency reads, built-in TTL...
codex    ✓  Postgres if you already have it. Avoid infra complexity...
gemini   ✓  Depends on scale. Redis at >10k concurrent sessions...
────────────────────────────────────────────────────
⚡ Disagreement: Redis vs Postgres — codex flags infra cost, others favor Redis.
→ Key question to resolve: Do you already run Postgres in this stack?

---

Why

Heavy AI users constantly switch between Claude, GPT, Gemini — manually copying context, pasting outputs back, comparing results by eye. This is slow, error-prone, and easy to skip under pressure.

WhaleCouncil automates the loop:

• One task in → multiple independent opinions out
• No anchoring — models don't see each other's responses in the first round
• Structured diff — disagreements are extracted, not buried in prose
• Actionable output — synthesis points to the next decision, not a summary of summaries

The core insight: one AI versus two AI is a qualitative difference, not a quantitative one.

---

Install

pip install whalecouncil

Or from source:

git clone https://github.com/openwhale-labs/whalecouncil
cd whalecouncil
pip install -e ".[dev]"

Set your API keys in .env (copy from .env.example):

cp .env.example .env
# edit .env with your keys

---

Usage

Review a task

# inline
council review "Is this approach correct: use a global dict as a cache in a FastAPI app?"

# from file
council review --file mycode.py

# choose models
council review --models claude,codex "Review this SQL query for performance issues"

# output to markdown file
council review --output markdown --file plan.md > report.md

Pipe input

cat diff.patch | council review --models claude,gemini
git diff HEAD~1 | council review "Any bugs introduced in this diff?"

List available models

council models

Configuration file

Persistent defaults can be set in ~/.council.toml (copy from council.toml.example):

[defaults]
models = ['claude-cli']
output = 'terminal'

[models.claude]
model = 'claude-sonnet-4-6'

CLI flags always override the config file, which overrides the built-in defaults. Per-provider [models.*] entries set the model each adapter uses.

---

Supported Models

Key	Provider	Notes
claude	Anthropic	Requires ANTHROPIC_API_KEY
claude-cli	Local Claude subscription (Claude Code CLI)	No API key needed. Requires claude CLI installed and logged in.
codex	OpenAI	Requires OPENAI_API_KEY
codex-cli	Local Codex CLI (codex exec)	No API key needed. Requires codex CLI installed and logged in.
gemini	Google	Requires GEMINI_API_KEY

---

How It Works

Input task
    │
    ├──► Claude  (independent, no other model's output)
    ├──► Codex   (independent)
    └──► Gemini  (independent)
         │
         ▼
   Collect opinions
         │
         ▼
   Extract disagreements
         │
         ▼
   Synthesize + surface next question
         │
         ▼
   Markdown or terminal report

Round 1 — Independent: Each model receives only the task and a neutral system prompt. No model sees another's output.

Round 2 — Synthesis (optional): A judge model receives all opinions and produces a structured diff: what they agree on, where they diverge, what the divergence reveals, and what question to resolve next.

---

Roadmap

• CLI skeleton
• Parallel model dispatch (asyncio)
• Model adapters: Claude, OpenAI, Gemini
• Disagreement extraction
• Synthesis / judge layer
• Markdown report output
• Pipe / stdin support
• Config file (~/.council.toml)
• WhaleTrace — save and replay council runs
• WhaleBench — structured benchmark tasks for agent evaluation

---

Design Philosophy

• CLI-first — works in any terminal, composable with pipes
• Independent opinions — first round is always blind; no anchoring
• Disagreement over consensus — the value is in the diff, not the summary
• Minimal infra — no server, no database, no account required
• Extensible — add any OpenAI-compatible model endpoint

---

Contributing

See CONTRIBUTING.md.

---

License

MIT © noetherly