fusion

name	fusion
description	Use to raise answer quality on a hard prompt by running it through a blind multi-model panel, then having Opus judge and synthesize one grounded answer. Panelists: 2x Opus subagents, GPT-5.5 via codex, and a local LMStudio model via pi — all over subscription, no API keys. Triggers on: fusion, run through fusion, fuse models, multi-model panel, panel of models, second opinion from multiple models.
user-invocable	true

Fusion — multi-model panel with a local LMStudio seat

Ported from fusion-fable (MIT). This variant replaces the Gemini/agy panelist with the local LMStudio model via pi. Opus and GPT-5.5 panelists run over subscription; no API keys.

Pipeline: detect → preflight → blind parallel fan-out → Opus judge → grounded final → provenance.

Read references/panel.md (fan-out contract) and references/judge_rubric.md (judging) before fanning out. Scripts live in ${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/.

Argument

/fusion [panel] — optional panel override:

• (none) → auto-detect the richest available panel
• opus → force opus-opus (offline, no external CLI)
• codex → force opus-codex
• local → force opus-codex-local

The user's actual question is the rest of the conversation / prompt — pass it to panelists verbatim.

Step 0 — Detect

Run scripts/detect_panel.sh and grep the SLUG= line for the auto-detected panel.

SLUG="$("${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/detect_panel.sh" | grep -oE 'SLUG=(opus-opus|opus-codex|opus-codex-local)$' | cut -d= -f2)"

If the user forced a panel, honor it only if available. If a forced panel needs a seat that detection shows is absent (e.g. local but no model loaded, or codex but no codex CLI), tell the user the seat is unavailable and fall back to the detected $SLUG — never claim a panelist ran when it did not.

Step 1 — Preflight

Write the user's verbatim task to a temp prompt file, then:

"${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/preflight.sh" "$SLUG" "$PROMPT_FILE"

Show the output. It is informational only — never block on it.

Step 2 — Blind fan-out (concurrent)

Launch every seat for $SLUG at once, each blind to the others. Use a per-run id (e.g. a timestamp) so temp files never collide.

• Opus seats (always 1, or 2 in opus-opus): dispatch each as an Agent subagent, handing it the user's task verbatim with instructions to use web search + bash and return only its answer. Do not let subagents see each other's work.
• GPT-5.5 seat (opus-codex, opus-codex-local): run scripts/run_codex.sh "$PROMPT_FILE" /tmp/fusion-$RUN-codex.md in the background.
• Local seat (opus-codex-local): run scripts/run_local.sh "$PROMPT_FILE" /tmp/fusion-$RUN-local.md in the background.

Wait for all to finish. Any seat whose runner exited non-zero is dropped — note it and continue with whatever answers landed.

Step 3 — Judge

Following references/judge_rubric.md: read every landed answer fresh, classify the task (artifact vs research), weight web-grounded panelists above the no-web local seat, and produce one grounded result. For artifact tasks, run/verify the merged candidate before delivering.

Step 4 — Deliver

Present the working artifact or grounded analysis. Never average or smooth.

Step 5 — Provenance

Archive every panelist answer plus the final to ~/.claude/fusion-runs/<timestamp>/ (outside the repo; never committed). Include the panel slug and which seats ran vs were dropped.

Step 6 — Present audit trail

Tell the user the panel composition, which seats contributed, which were dropped and why, and the provenance directory path.

Prerequisites

• opus-opus: always works, fully offline.
• opus-codex: authenticated codex CLI on PATH.
• opus-codex-local: codex + pi on PATH + LMStudio running at http://127.0.0.1:1234 (override with FUSION_LMSTUDIO_URL) with a model loaded. jq and curl are required for model detection.