Fable5 Fusion (Claude Code + Codex)

---

A Claude Code plugin that turns one question into a multi‑model panel. Three models answer the same task in parallel — none seeing the others' work — then GPT‑5.5 judges all of them and a Claude Opus 4.8 synthesizer writes the final answer grounded in that analysis.

Independent agreement is your highest‑confidence signal. Independent disagreement is exactly what a single model's self‑review hides.

---

🧭 The pipeline

                      YOUR TASK
                          │
                          │   verbatim · no lenses · blind
        ┌─────────────────┼─────────────────┐
        ▼                 ▼                 ▼
    Opus 4.8 ①        Opus 4.8 ②        GPT‑5.5          ╮  PANEL · parallel
   (claude -p)        (claude -p)        (codex)         │  & independent
   max · locked       max · locked      xhigh · locked   ╯  web+bash · blind
        └─────────────────┼─────────────────┘
                          ▼
                  JUDGE · GPT‑5.5             ◄─ xhigh · fast
                  structured analysis            no vote · no average
                          │
                          ▼
                  SYNTH · Opus 4.8           ◄─ max
                  writes the final answer        (runs & merges code)
                          │
                          ▼
                     FINAL ANSWER
              (answer first, then audit trail)

Even two cold runs of the same model diverge enough — different reasoning paths, tool calls, and sources — that synthesizing them beats running it once. The diversity is harvested from independence, not manufactured with personas.

---

🎛️ The panel

Seat	Model	Runs via	Effort	Job
🟠 Panelist 1	Claude Opus 4.8	claude -p (run_claude.sh)	max (locked)	independent answer
🟠 Panelist 2	Claude Opus 4.8	claude -p (run_claude.sh)	max (locked)	independent answer (2nd cold run)
🟣 Panelist 3	GPT‑5.5	codex (run_codex.sh)	xhigh (locked)	independent answer
⚖️ Judge	GPT‑5.5	codex (run_codex.sh)	xhigh (locked)	structured analysis — not the final answer
✍️ Synthesizer	Claude Opus 4.8	claude -p (run_claude.sh)	max (locked)	writes the one final answer

All reasoning seats are locked subprocesses; the orchestrator only coordinates. Every seat runs as a wrapped runner at a fixed effort, so the Fusion session's own effort no longer affects output. The two Opus panelists and the Opus synthesizer are set to max explicitly — run_claude.sh exports CLAUDE_CODE_EFFORT_LEVEL=max (the highest‑precedence effort knob, above the --effort flag and settings.json) and also passes --effort max, so they hit max regardless of the session's inherited effort (only a settings.json env pin of a different value could override it). The GPT‑5.5 panelist and the GPT‑5.5 judge are locked at xhigh in run_codex.sh.

Why split judge and synthesizer? GPT‑5.5 analyzes the panel; a separate Claude Opus 4.8 writes the answer. Splitting "analyze" from "write" — and crossing model families between them — keeps the final answer honest instead of one seat defending its own draft.

---

🚀 Quickstart

# 1. add the marketplace + install
claude plugin marketplace add https://github.com/Rylaa/fable5-fusion
claude plugin install fable5-fusion

# 2. run it
/fable5-fusion:fusion <your question or task>

That's it. The skill auto‑detects the panel, fans out, judges, synthesizes, and hands you the final answer followed by a per‑seat audit trail.

---

🧠 How it works