fusion-fable

Fusion-Fable

Fuse a panel of frontier models into one Fable-tier answer.

Fusion-Fable is a Claude Code skill that runs a hard question through a panel → judge pipeline. The same prompt is dispatched to several models in parallel — each answering independently with web search and bash, none seeing the others' work — and then Opus 4.8 judges every answer into a structured analysis (consensus, contradictions, partial coverage, unique insights, blind spots) and writes a final answer grounded in it.

The mechanism is independence, then synthesis. The diversity that makes a panel beat a single model is harvested, not manufactured: running the same prompt independently yields different reasoning paths, tool calls, and sources — even two cold runs of the same model diverge enough that synthesizing them beats running it once. So there are no contrived "lenses" or personas; every panelist gets the task verbatim and answers it straight. Fuse Opus 4.8 + Opus 4.8, or Opus 4.8 + GPT-5.5 (via the codex CLI), into a result better than either alone — a Fable-tier fusion.

                      ┌──────────────┐
                 ┌──▶ │  panelist 1  │ ─┐   (web + bash, independent)
                 │    └──────────────┘  │
                 │    ┌──────────────┐  │   ┌──────────────┐
 prompt ──▶ fan ─┼──▶ │  panelist 2  │ ─┼─▶ │   Opus 4.8   │ ──▶ final answer
            out  │    └──────────────┘  │   │   (judge +   │     (grounded in
                 │    ┌──────────────┐  │   │  synthesize) │      the analysis)
                 └──▶ │  panelist 3  │ ─┘   └──────────────┘
                      └──────────────┘
              Opus 4.8 / GPT-5.5 / Gemini      consensus · contradictions ·
              (each answers blind)             partial · unique · blind spots

Opus 4.8 always judges and writes the final answer — the pipeline can't be reversed, because the panelist models can't call back out to spawn Opus.

The panels

The skill auto-detects which panelist CLIs are installed and uses the richest panel available, falling back gracefully when one is missing.

Install

git clone https://github.com/duolahypercho/fusion-fable.git
cd fusion-fable
./install.sh

This copies the skill to ~/.claude/skills/fusion and the slash commands to ~/.claude/commands, then prints which panels your machine can run. Restart Claude Code (or run /reload-skills) afterward.

Override the target with CLAUDE_CONFIG_DIR=/path/to/.claude ./install.sh.

Use it

Three ways, all equivalent under the hood:

• Natural language — just ask. The skill auto-triggers and picks the richest panel:

  "Run this through Fusion: is it safe to ALTER TABLE … ADD COLUMN on a 200M-row Postgres table in prod?"

• Pinned slash commands:

  /fusion-opus4.8  does my JWT refresh-rotation design have a replay hole?
  /fusion-gpt5.5   is git push --force-with-lease actually safe on a shared branch?
  /fusion-3        full 3-family panel (Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro)
  

• Force a panel in prose — "run the opus4.8-gpt5.5 Fusion on …".

Every run returns the same structure: a Final answer up top, then the audit trail — Consensus / Contradictions / Partial coverage / Unique insights / Blind spots — with each point attributed to the panelist that raised it, so you can see how the answer was assembled. Every run is also written to a timestamped provenance file under ~/.claude/fusion-runs/ (raw panelist answers + analysis + final answer) for auditing.

Requirements

• Claude Code, with the session running Opus 4.8 (panelist subagents and the judge inherit the session model — on another model the slug is nominal, not literal).
• For opus4.8-gpt5.5: the codex CLI installed and logged in to an account with GPT-5.5 access. The runner uses codex exec (tested against codex-cli 0.139). It runs against a throwaway copy of the current repo/workdir with trusted local access so tools such as gh, local test runners, Docker, and SDK-managed toolchains behave like they do in your terminal without writing back to the live checkout. Each Fusion invocation uses its own temporary prompt/output directory, so concurrent runs in different Claude Code sessions do not read each other's GPT panelist artifacts.
• For the 3-model panel: the agy (Antigravity) CLI installed and its keyring seeded — run agy once interactively to complete the Google OAuth, after which headless runs reuse that login. run_gemini.sh works around agy's print-mode bug (empty stdout under no TTY) by running it under a pseudo-TTY with a transcript-JSONL fallback and a hard anti-empty guard, so it never returns a silently empty answer. The pseudo-TTY is allocated by a pty.fork() Python helper (_pty_run.py) so it keeps working when the orchestrator itself runs in a socket (cmux / headless), where script aborts on tcgetattr.

Only the opus4.8-4.8 panel is truly zero-setup; the GPT-5.5 and Gemini panels light up once their CLIs are installed and authenticated. Note: there is no timeout/gtimeout on stock macOS, so the runners use a self-contained perl timeout helper (FUSION_TIMEOUT, default 300s per panelist).

What's in here

skills/fusion/
  SKILL.md                  detect → preflight → blind fan-out → judge → grounded final → save
  scripts/
    _fusion_lib.sh          shared helpers: perl-based per-panelist timeout, have()
    _pty_run.py             pty.fork() runner so agy gets a TTY even under socket stdio (cmux/headless)
    detect_panel.sh         picks the richest available panel
    preflight.sh            non-blocking token/call estimate + Codex cap reminder
    run_codex.sh            runs the GPT-5.5 panelist (web + bash) with a timeout, captures its answer
    run_gemini.sh           runs the Gemini 3.1 Pro panelist via agy (pseudo-TTY + transcript fallback)
    save_run.sh             writes the timestamped provenance .md to ~/.claude/fusion-runs/
  references/
    panel.md                why independent parallel runs (no lenses) — the panel mechanism
    judge_rubric.md         the structured analysis + grounded final answer
commands/
  fusion-opus4.8.md         /fusion-opus4.8  (pinned opus4.8-4.8 panel)
  fusion-gpt5.5.md          /fusion-gpt5.5   (pinned opus4.8-gpt5.5 panel)
  fusion-3.md               /fusion-3        (pinned full opus4.8-gpt5.5-gemini3.1pro panel)
install.sh                  copies the above into ~/.claude

Why a panel beats one model

On the DRACO deep-research benchmark, OpenRouter found that fusing model answers consistently beats the individual models — and that a meaningful chunk of the lift comes from the synthesis step itself, not just from mixing architectures: two independent runs of one model, synthesized, beat that model run once. Fusion-Fable implements that same independence-then-judge pipeline locally in Claude Code.

Cost & latency

A panel costs roughly N× a single answer in tokens and runs as slow as its slowest panelist. That's the deliberate trade: spend more to stop being confidently wrong where that's expensive. For quick or low-stakes questions, a single direct answer is the right call.

License

MIT — see LICENSE.