fusion

Fusion

Fusion turns one prompt into a panel. The task goes to the two strongest frontier models at the same time, each answering independently — with web search and bash, and with no knowledge of the other. Then GPT-5.6 Sol (via codex) reads every answer and extracts the structure of the panel's reasoning (what they agree on, where they conflict, what only one saw, what they all missed), and a Claude Fable 5 synthesizer writes the final answer grounded in that analysis.

The mechanism is independence, then synthesis. The diversity that makes a panel beat a single model is harvested, not manufactured: running the same task independently yields different reasoning paths, tool calls, and sources — two different frontier families (Fable 5 and GPT-5.6 Sol) diverge far more than two runs of one model. So there are no assigned "lenses" or personas; every panelist gets the user's task verbatim and answers it straight. (See references/panel.md.)

Fixed pipeline (this build)

No Opus seat, no model fallback. This build removed Opus 4.8 entirely: every Claude seat runs claude-fable-5 and a seat that fails permanently is absent (never silently swapped to another model). run_claude.sh still accepts an optional 5th fallback_model argument as a general mechanism, but the panel does not use it. The judge's CLI-level cross-family fallback (codex fails → retry the judge on claude) runs on Fable 5.

Every reasoning seat is a locked subprocess; the orchestrator only coordinates. All four seats run as wrapped runners at a fixed effort, so the Fusion session's own effort no longer affects output quality — the orchestrator just writes prompt files, launches the runners, and presents results; it does no judging or synthesis reasoning itself. The Fable 5 panelist and the Fable 5 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 passes --effort max, so they hit max regardless of the session's inherited effort. (The Agent tool / agent-teams can't set a per-call effort, so the seats don't go through them.) The one thing that still wins over this is a settings.json env block pinning a different CLAUDE_CODE_EFFORT_LEVEL. The GPT-5.6 Sol panelist and the GPT-5.6 Sol judge are locked at ultra in run_codex.sh, pinned to the gpt-5.6-sol model (-c model=…) regardless of ~/.codex/config.toml. Every seat is also time-bounded — each runner wraps its CLI call in a per-seat timeout (FUSION_TIMEOUT, default 1800s) so one stuck seat can't hang the panel.