fusion

name	fusion
description	Run a local OpenRouter Fusion style multi-model deliberation using available CLI agents such as claude -p and codex exec. Use for deep research, high-stakes synthesis, architecture decisions, code review second opinions, product comparisons, strategy memos, contradiction finding, or any task where independent model perspectives and a judged synthesis are worth extra cost and latency.

Fusion Panel

Overview

Use this skill to replicate the Fusion pattern locally: run independent panel agents in parallel, preserve their raw answers, extract consensus and contradictions, then synthesize a final answer grounded in the panel.

Default to this workflow for expensive-to-be-wrong tasks. Do not use it for simple tactical questions where one direct answer is enough.

Workflow

1. Clarify the evaluation target: what would make the answer useful, correct, and complete.
2. Choose a small panel. Start with claude:sonnet + codex; add claude:opus or duplicate runs only when the result justifies the cost.
3. Run scripts/fusion.py from this skill folder, passing the user's task through stdin or --prompt.
4. Read final.md first, then inspect judge.md and raw panel outputs when disagreement matters.
5. Verify the final answer with the real task surface: source checks for research, tests for code, browser/app checks for UI, or a rubric for decisions.

Helper Script

Run from any workspace:

python3 ~/.codex/skills/fusion/scripts/fusion.py \
  --panel claude:sonnet \
  --panel codex \
  --synthesizer codex \
  --prompt "Research the tradeoffs of X and recommend a path."

Pipe longer prompts:

cat task.md | python3 ~/.codex/skills/fusion/scripts/fusion.py \
  --panel claude:sonnet \
  --panel codex \
  --panel claude:sonnet \
  --synthesizer claude:opus

Use claude:* specs only on machines where claude -p is authenticated. If Claude is unavailable, the script keeps going with successful panel members by default and records the failure in run.json.

The script writes a run directory under /tmp/fusion-runs/ by default:

• prompt.md - original task
• panel/*.md - raw panel responses
• judge.md - consensus, contradictions, missing coverage, unsupported claims
• final.md - synthesized answer
• run.json - command metadata and exit statuses

Panel specs:

• claude or claude:sonnet
• claude:opus
• codex
• codex:gpt-5 or any installed Codex model alias

Useful flags:

• --out-dir PATH - choose where artifacts are written
• --timeout SECONDS - per-agent timeout
• --codex-search - enable Codex web search when the installed CLI supports it
• --fail-fast - stop instead of synthesizing when a panel member fails

Prompting Rules

Ask each panel member to answer independently. Require:

• direct answer or recommendation
• key claims
• evidence and sources when available
• uncertainties
• missing information
• things the synthesizer should verify

Ask the judge to produce:

• consensus
• contradictions
• unique useful insights
• unsupported claims
• missing coverage
• best-supported final position

Ask the synthesizer to write only from supported claims, with uncertainty visible.

Safety

Panel calls are intended to be read-only. The helper script runs Codex with a read-only sandbox and asks Claude with editing tools disallowed. If the task requires code edits, use Fusion Panel for analysis first, then make edits in the main session after reviewing the result.

Do not outsource final responsibility to the panel. The main agent must still verify the result before reporting it as true or done.