duolahypercho/fusion-fable

Fusion‑Fable – A Claude Code skill that fuses multiple LLMs into a single, higher‑quality answer

What it is – Fusion‑Fable is a skill for Claude Code (the “Opus 4.8” model) that lets you run a prompt through a panel of independent LLMs, have each model answer the same question (with its own web‑search and bash tool calls), and then let Opus 4.8 act as a judge that analyses the panel’s outputs and synthesises a final, grounded response.

Why it matters – Independent runs of the same model already produce diverse reasoning paths; combining several models (or two runs of the same model) and then synthesising the results has been shown to improve accuracy on hard benchmarks. Fusion‑Fable packages that idea into a ready‑to‑use Claude Code skill, so you can get “Fable‑tier” answers without writing any orchestration code yourself.

---

How it works

1. Fan‑out – The skill detects which model CLIs are installed on your machine and launches a panel:
   • opus4.8-4.8 – two independent Opus 4.8 runs (zero‑setup).
   • opus4.8-gpt5.5 – Opus 4.8 plus OpenAI’s GPT‑5.5 via the codex CLI.
   • opus4.8-gpt5.5-gemini3.1pro – adds Google Gemini 3.1 Pro via the agy CLI.
2. Blind execution – Each panelist receives the exact same prompt, runs in parallel, and may invoke web search or bash tools. They never see each other’s output.
3. Judging & synthesis – Opus 4.8 receives all raw answers, produces a structured analysis (consensus, contradictions, partial coverage, unique insights, blind spots) and writes a final answer that is explicitly grounded in that analysis.
4. Provenance – Every run is saved as a timestamped markdown file under ~/.claude/fusion‑runs/ containing the raw panelist answers, the analysis, and the final answer.

---

Installation (quick)

git clone https://github.com/duolahypercho/fusion-fable.git
cd fusion-fable
./install.sh          # copies skill to ~/.claude/skills/fusion
# optional: set CLAUDE_CONFIG_DIR if you keep Claude config elsewhere

The installer also installs slash‑commands (/fusion‑opus4.8, /fusion‑gpt5.5, /fusion‑3, /fusion‑plan) and an optional hook script.

---

Using the skill

Method	Example	What happens
Natural‑language trigger	Run this through Fusion: is it safe to ALTER TABLE … on a 200M‑row Postgres table?	Claude automatically runs the richest available panel and returns the final answer with the analysis block.
Pinned slash command	/fusion‑gpt5.5 is git push --force‑with‑lease actually safe on a shared branch?	Forces the opus4.8‑gpt5.5 panel; useful when you know which models you want to involve.
Explicit panel name in prose	run the opus4.8‑gpt5.5 Fusion on …	Same as the slash command but written in ordinary text.

All three produce identical output format:

**Final answer**
…
---
**Consensus** … (attributed)
**Contradictions** … (attributed)
**Partial coverage** …
**Unique insights** …
**Blind spots** …

---

Planning mode – /fusion‑plan

Fusion‑Fable also ships an iterative planning skill that plugs into the oh‑my‑claudecode (OMC) planning system:

1. Requirements – optional interactive interview or automatic extraction from an existing plan file.
2. Three refinement rounds – each round runs a two‑model panel (Opus 4.8 + GPT‑5.5) that critiques the current plan; Opus 4.8 judges and produces a tighter plan for the next round.
3. Write‑back – the converged plan overwrites the original .omc/plans/<slug>.md in a concise, content‑dense form.
4. Handoff – you can then run OMC’s quality‑gate (/omc‑plan --review) and proceed to execution.

A lightweight optional hook (hooks/fusion‑plan‑nudge.sh) can be enabled to remind you to run /fusion‑plan before delegating a heavy implementation task to a sub‑agent.

---

Requirements & optional dependencies

Component	Needed for	How to get
Claude Code (Opus 4.8)	Core skill & judge	Already part of Claude Code environment
codex CLI	GPT‑5.5 panelist	npm i -g @openai/codex-cli (or follow repo link) and log in with a GPT‑5.5‑enabled account
agy CLI	Gemini 3.1 Pro panelist	Install from its repo, run once to complete Google OAuth
perl timeout helper	All panels (fallback on macOS)	Included in the repo; no extra install needed
python (for _pty_run.py)	Gemini panel (needs a pseudo‑TTY)	Any recent Python 3 interpreter

If only the zero‑setup panel is needed, no extra tools are required.

---

What’s inside the repo

skills/fusion/            # core skill implementation
  SKILL.md               # skill definition
  scripts/               # helpers for timeouts, panel detection, running each CLI, saving provenance
    _fusion_lib.sh
    _pty_run.py
    detect_panel.sh
    preflight.sh
    run_codex.sh
    run_gemini.sh
    save_run.sh
  references/            # design docs (panel rationale, judge rubric)
skills/fusion‑plan/       # OMC‑integrated planning skill
commands/                # slash‑command markdown files
hooks/                   # optional pre‑tool‑use hook (disabled by default)
install.sh                # installer script

---

Strengths

• Model diversity without custom prompting – each model runs the same prompt independently, preserving its native reasoning style.
• Structured synthesis – the judge produces a clear audit trail, making it easy to see where the answer came from.
• Extensible panels – automatically picks the richest panel your system can support; you can add more panelists by writing a new run_*.sh script.
• Auditable provenance – every run is saved locally for later review or debugging.
• Planning integration – the iterative /fusion‑plan workflow can improve high‑stakes design documents.

---

Limitations

• Cost & latency – token usage and wall‑clock time scale linearly with the number of panelists; a three‑model panel can be several times slower and more expensive than a single answer.
• Dependency on external CLIs – GPT‑5.5 and Gemini panels require the codex and agy CLIs plus valid credentials; without them you fall back to the basic Opus‑only panel.
• Claude‑only environment – the skill runs inside Claude Code; it cannot be used as a standalone CLI or web service.
• No built‑in timeout on macOS – the repo ships a Perl helper; if that fails you may need to adjust FUSION_TIMEOUT.

---

License

MIT – see the LICENSE file in the repository.

---

Bottom line – Fusion‑Fable gives Claude Code users a plug‑and‑play way to harness the proven “panel‑then‑judge” technique, turning multiple LLM answers into a single, well‑justified response and even supporting iterative planning when paired with OMC.