Skip to content

About

Deterministic multi-model planning & review councils for Claude Code (Claude role panel + GPT-5.5). Installable plugin + marketplace.

Resources

Stars

4 stars

Watchers

0 watching

Forks

Latest commit

 

History

35 Commits

Folders and files

Repository files navigation

Fusion Council

Fusion Council hero illustration showing multiple read-only review perspectives converging into a synthesized report

Two deterministic, read-only councils for Claude Code that vet your work with more than one model family before you commit to it:

  • /fusion-council:fusion-plan <task> — before a non-trivial change. A Claude role panel (architect · skeptic · test-strategist · maintainer) plus GPT Sol deliberate in parallel; one merged judge+synthesize call compares them on merit and writes the plan from its own verdict. Since v0.4.0 this is the same one-call shape fusion-review has run since v0.3.0, ported on that arm's bench result — the plan council itself is unbenched (EVAL.md §B0); args: { splitSynth: true } restores the two-call pipeline.
  • /fusion-council:fusion-review — after the change. One generalist Claude seat plus GPT Sol review your current git diff, judged and synthesized → one findings report. This duo is the bench-validated default (matched recall on the settled case set — n=1 hard case, so recall deltas are anecdote — at ~half the clean-diff noise and −21% cost; EVAL.md §B5′); pass args: { roles: true } to restore the role panel (skeptic · test-strategist · maintainer + GPT Sol). Since v0.3.0 the judge and the report writer are one call: the same seat that grounds and gates the panel's claims writes the final report from its own verdict, which removes the judge→synthesizer handoff both measured information-loss bugs lived in. A same-era bench pass tied recall (5/5 bug cases incl. the hard one), lowered clean-diff noise on every clean case, and cost −7% (EVAL.md §B0); args: { splitSynth: true } restores the two-call pipeline.

Both are read-only: panelists may read the repo but never edit it. Only you (the main session) implement, and only after you approve.

Install

/plugin marketplace add halfmoon-mind/fusion-council   # or: /plugin marketplace add /path/to/this/repo
/plugin install fusion-council@fusion-council

Local testing without GitHub: point the marketplace at this directory directly (/plugin marketplace add . from the repo root).

Use

/fusion-council:fusion-plan add rate limiting to the upload endpoint
# ...review the plan, approve, implement...
/fusion-council:fusion-review

The GPT Sol seat (optional)

The non-Claude panel seat shells out to the local Codex CLI (codex exec, ChatGPT subscription — no metered API). If codex isn't installed — or a run fails or times out — that seat drops out and the council runs Claude-only; every report's Council Coverage line says so, so reduced diversity is never silent. Install Codex CLI to enable it. It runs the newest gpt-*-sol in Codex's model list (~/.codex/models_cache.json, resolved per run, falling back to gpt-6.1-sol; args.codexModel pins one). The seat hands its prompt to codex exec on stdin (the - / piped form, so large tasks/diffs aren't capped by ARG_MAX), which needs a Codex CLI recent enough to support it (verified on v0.142.0); an older CLI just shows up as UNAVAILABLE in Council Coverage.

Telemetry & self-evaluation

The council writes nothing. Claude Code journals every Workflow run at ~/.claude/projects/<project>/<session>/workflows/wf_*.json — the returned judge, panel, coverage and stats (arm, per-seat claim counts, GPT ran/retried) plus what the plugin itself could never record: wall time, total tokens, per-seat model and tokens, and killed/failed status. eval/analyze.sh reads those journals and prints the scoreboard (--bench includes scratch-repo bench runs; eval/journal.sh looks a run up by cwd or runId). Until 0.4.0 a sonnet seat appended a subset of this to ~/.fusion-council/telemetry.jsonl through a heredoc relay; that seat cost ~33k tokens and ~40s per run and was removed in 0.5.0 (args.noTelemetry is now a no-op). The full validation methodology (verified literature, seeded-bug benchmark design, ablation arms noGpt/roles, kill metrics, and the §B5′ re-validation that made duo the review default) lives in EVAL.md.

Layout

.claude-plugin/plugin.json        plugin manifest
.claude-plugin/marketplace.json   one-plugin marketplace (this repo)
agents/                           5 read-only role agents (Read/Glob/Grep only)
skills/fusion-plan, fusion-review entry points (call the workflows by path)
workflows/*.js                    the orchestration (bundled; called via ${CLAUDE_PLUGIN_ROOT})
EVAL.md, eval/                    validation methodology + journal scoreboard scripts

Current Claude Code registers plugin workflows by name: real sessions launch Workflow({ name: "fusion-council:fusion-review" }) and the skill text shown to the model is the workflow's own meta, not skills/*/SKILL.md. The skills/ entries stay as the documented contract and as the scriptPath fallback for older harnesses.

About

Deterministic multi-model planning & review councils for Claude Code (Claude role panel + GPT-5.5). Installable plugin + marketplace.

Resources

Stars

4 stars

Watchers

0 watching

Forks

Releases

Contributors

Languages