{"id":50849541,"url":"https://github.com/frahlg/fusion","last_synced_at":"2026-06-14T12:34:50.312Z","repository":{"id":364764992,"uuid":"1269038658","full_name":"frahlg/fusion","owner":"frahlg","description":"Fusion — a Claude Code skill: convene a panel of frontier models (Opus 4.8 + GPT-5.5 xhigh), judged into one grounded answer. Deep Thought, with receipts.","archived":false,"fork":false,"pushed_at":"2026-06-14T11:47:01.000Z","size":199,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-06-14T12:24:00.749Z","etag":null,"topics":["agents","ai","ai-agents","anthropic","claude","claude-code","ensemble","gpt-5","llm","llm-tools","multi-agent","opus"],"latest_commit_sha":null,"homepage":null,"language":"Shell","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/frahlg.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-06-14T08:27:12.000Z","updated_at":"2026-06-14T11:47:05.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/frahlg/fusion","commit_stats":null,"previous_names":["frahlg/fusion"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/frahlg/fusion","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/frahlg%2Ffusion","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/frahlg%2Ffusion/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/frahlg%2Ffusion/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/frahlg%2Ffusion/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/frahlg","download_url":"https://codeload.github.com/frahlg/fusion/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/frahlg%2Ffusion/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34322074,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-14T02:00:07.365Z","response_time":62,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["agents","ai","ai-agents","anthropic","claude","claude-code","ensemble","gpt-5","llm","llm-tools","multi-agent","opus"],"created_at":"2026-06-14T12:34:49.761Z","updated_at":"2026-06-14T12:34:50.303Z","avatar_url":"https://github.com/frahlg.png","language":"Shell","funding_links":[],"categories":[],"sub_categories":[],"readme":"# Fusion\n\n[![License: MIT](https://img.shields.io/badge/License-MIT-black.svg)](LICENSE)\n[![Claude Code skill](https://img.shields.io/badge/Claude%20Code-skill-d97757)](https://claude.com/claude-code)\n[![Panel](https://img.shields.io/badge/panel-Opus%204.8%20%2B%20GPT--5.5-blue)](#the-panels)\n[![Answer](https://img.shields.io/badge/answer-42-9cf)](#the-panels)\n\n**Stop trusting one model. Convene a panel.**\n\nFusion is a [Claude Code](https://claude.com/claude-code) skill for questions where a single model answer is\ntoo cheap to trust. It runs your hardest prompt through a **panel → judge** pipeline: several frontier models\nanswer *in parallel and blind*, then Opus 4.8 judges every answer and forges one you can actually trust — with\nthe full audit trail underneath.\n\n\u003e *Deep Thought ran for 7½ million years and returned a single number. Fusion runs a panel of frontier models\n\u003e and returns one grounded answer — plus the receipts.*\n\n\u003cp align=\"center\"\u003e\n  \u003cimg src=\"assets/fusion-demo.gif\" alt=\"An illustrative /fusion run: an Opus 4.8 + GPT-5.5 (xhigh) panel answers blind, then Opus judges into one verdict with a consensus / contradictions / blind-spots audit trail\" width=\"860\"\u003e\n\u003c/p\u003e\n\n\u003cp align=\"center\"\u003e\u003csub\u003eAn illustrative \u003ccode\u003e/fusion\u003c/code\u003e run — verdict first, audit trail underneath.\u003c/sub\u003e\u003c/p\u003e\n\n**The counterintuitive part:** you don't need *different* models to beat one model. Even two cold runs of the\n*same* model diverge — different reasoning paths, different searches, different mistakes — and synthesizing\nthat divergence beats running it once. Fusion harvests that diversity instead of faking it with personas or\n\"lenses\".\n\n```\n                      ┌──────────────┐\n                 ┌──▶ │  panelist 1  │ ─┐   (web + bash, independent)\n                 │    │  Opus 4.8    │  │\n prompt ──▶ fan ─┤    └──────────────┘  ├─▶ ┌──────────────┐\n            out  │    ┌──────────────┐  │   │   Opus 4.8   │ ──▶ Fusion's\n                 └──▶ │  panelist 2  │ ─┘   │  (judge +    │     final answer\n                      │  GPT-5.5     │      │  synthesize) │   (grounded in\n                      │  (xhigh)     │      └──────────────┘    the analysis)\n                      └──────────────┘\n              each answers blind          consensus · contradictions ·\n                                          partial · unique · blind spots\n```\n\nMany great minds computing toward one answer — except this answer ships with receipts, disagreement, and its\nown confidence boundaries. Opus 4.8 **always** judges and writes the final answer; the pipeline can't be\nreversed, because the panelist models can't call back out to spawn Opus.\n\n## What an answer looks like\n\nFusion always leads with the verdict, then shows its work (illustrative):\n\n```text\n\u003e /fusion Single MQTT broker or a quorum for our edge control loop?\n\nRun the quorum. Every independent line of reasoning converged here: a single\nbroker is a single point of failure your control loop can't survive, and the\nlatency cost of consensus stays inside your 200 ms budget. One real caveat — no\npanelist could verify failover time under network partition, so prove that on\nyour hardware before you trust it in production.\n\n──────────────────────────────────────────────────────────────\nPanel: opus4.8-gpt5.5 — Opus 4.8 ✓  ·  GPT-5.5 (xhigh) ✓\nConsensus      · quorum for availability; latency fits the 200 ms budget\nContradictions · broker count (3 vs 5) — adjudicated to 3 on the cited benchmark\nUnique insight · GPT-5.5 flagged split-brain on even-sized clusters\nBlind spots    · failover-under-partition timing unverified by either panelist\n```\n\nThe verdict you can read in ten seconds. The audit trail is there for when being wrong is expensive.\n\n## The panels\n\n| Slug | Panel | Requires |\n| --- | --- | --- |\n| `opus4.8-gpt5.5` | Opus 4.8 + **GPT-5.5 (xhigh)** in parallel → Opus judges | the `codex` CLI |\n| `opus4.8-4.8` | the **same prompt run twice** as 2 independent Opus 4.8 panelists → Opus judges | nothing — works everywhere (fallback) |\n\nThe skill auto-detects whether the `codex` CLI is installed and usable, and falls back gracefully to the\npure-Opus panel when it isn't. The default is **Opus 4.8 + GPT-5.5 (xhigh)**.\n\n*Two independent minds in; one answer out. Deep Thought only had the one.*\n\n## When to use it\n\nReach for Fusion when one model being **confidently wrong** would cost you more than an extra model pass:\n\n- architecture and design calls\n- high-stakes research\n- gnarly / incident debugging\n- vendor or framework decisions\n- claims that need sources, commands, or cross-checking\n\nSkip it for the easy stuff — Fusion is deliberately slower and more expensive than one model.\n\n## Install\n\n```bash\ngit clone https://github.com/frahlg/fusion.git\ncd fusion\n./install.sh\n```\n\nThis copies the skill to `~/.claude/skills/fusion` and prints which panels your machine can run. The skill is\nitself invocable as `/fusion`, so no separate command file is needed. Restart Claude Code (or run\n`/reload-skills`) afterward.\n\n\u003e Override the target with `CLAUDE_CONFIG_DIR=/path/to/.claude ./install.sh`.\n\n## Use\n\nRun one hard prompt:\n\n```\n/fusion Should we run our edge control loop on a single MQTT broker or a quorum?\n```\n\n…or just ask in prose to \"run it through Fusion\" / \"ask the panel\". Fusion returns the final answer first,\nthen the audit trail (Consensus / Contradictions / Partial coverage / Unique insights / Blind spots) so you\ncan trace every claim back to a panelist.\n\nDon't panic about the cost: a panel runs roughly N× the tokens of a single answer and is as slow as its\nslowest panelist. That's the deliberate trade for an answer worth trusting — not a default.\n\n## Requirements\n\n- Claude Code with Opus 4.8 (judge + an always-available panelist via subagents).\n- Optional: the [`codex`](https://github.com/openai/codex) CLI for the GPT-5.5 (xhigh) panelist. Without it,\n  Fusion runs the pure-Opus `opus4.8-4.8` panel.\n\n## License\n\nMIT — see [LICENSE](LICENSE).\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ffrahlg%2Ffusion","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Ffrahlg%2Ffusion","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ffrahlg%2Ffusion/lists"}