fusion-3
Fusion full panel — Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro in parallel, judged by Opus 4.8…
OMC interview (interactive) → 3-round Opus 4.8 + GPT-5.5 fusion deepening → concise .omc/plans/ output → review/execute handoff
$ npx -y skills add duolahypercho/fusion-fable --agent claude-codeHow it fires
How this command gets triggered: by you, by Claude, or both.
/fusion-planContext preview
What this command does when you run it.
OMC interview (interactive) → 3-round Opus 4.8 + GPT-5.5 fusion deepening → concise .omc/plans/ output → review/execute handoff
description: OMC interview (interactive) → 3-round Opus 4.8 + GPT-5.5 fusion deepening → concise .omc/plans/ output → review/execute handoff argument-hint: "[--no-interview] <what to plan / decide>"
Invoke the **fusion-plan** skill on the task below. Run the full OMC-integrated flow:
1. **Requirements — pick the mode FIRST (Step 0 of SKILL.md):**
interview, do NOT call AskUserQuestion, do NOT auto-chain omc-plan.** Derive requirements from existing context (story doc `docs/stories/STORY-*.md`, the flagged section of the larger `.omc/plans/*.md`, and the task args) as the seed. If context is insufficient, stop and report what's missing — never block on an unanswerable question.
`Skill("oh-my-claudecode:omc-plan")` on the task verbatim (no `--consensus`; auto-detect interview vs direct) → initial plan in `.omc/plans/<slug>.md`. Capture that path as the seed; do not let OMC auto-execute. 2. **Deepen with the panel (3 rounds, seeded):** each round, ONE Opus 4.8 panelist (Agent subagent, `model: opus`) and ONE GPT-5.5 panelist (`codex exec`, **high** effort) independently critique-and-improve the current plan IN PARALLEL and BLIND. Round 1 seeds from the OMC plan (so every interview requirement is preserved); rounds 2–3 seed from the previous synthesis. After each round YOU (Opus 4.8) judge both (Consensus / Contradictions / Partial / Unique / Blind spots) and synthesize one tighter plan. Stop early on `NO_MATERIAL_CHANGE` (round ≥ 2). If `codex` is missing, fall back to two Opus 4.8 panelists per round. 3. **Write back:** the synthesized plan must be **concise but content-dense** (no padding) and is written back to the SAME `.omc/plans/<slug>.md` in OMC standard format, preserving all gathered requirements. 4. **Handoff (offer, don't auto-run execution):** quality gate via `/omc-plan --review .omc/plans/<slug>.md`, then `/team` or `/ralph` on approval.
Follow `skills/fusion-plan/SKILL.md` exactly. Do not edit OMC's own skills — integrate only through the `.omc/plans/*.md` artifact and by invoking `omc-plan`.
Task: $ARGUMENTS
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.
Repo: duolahypercho/fusion-fable
Fusion full panel — Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro in parallel, judged by Opus 4.8…
Fusion panel of Opus 4.8 + Gemini 3.1 Pro in parallel, judged by Opus 4.8…
Fusion panel of Opus 4.8 + GPT-5.5 in parallel, judged by Opus 4.8 (opus4.8-gpt5.5)
Fusion panel of two independent Opus 4.8 runs, judged by Opus 4.8 (opus4.8-4.8)