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)

Seat	Model	CLI	Effort	Role
Panelist 1	Claude Fable 5	claude -p (run_claude.sh)	max (locked)	independent answer
Panelist 2	GPT-5.6 Sol	codex (run_codex.sh)	ultra (locked)	independent answer
Judge	GPT-5.6 Sol	codex (run_codex.sh)	ultra (locked)	structured analysis — does NOT write the final answer
Synthesizer	Claude Fable 5	claude -p (run_claude.sh)	max (locked)	writes the ONE final answer from the judge analysis

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.

The judge and the synthesizer are deliberately different seats. GPT-5.6 Sol (codex) analyzes the panel; a separate Claude Fable 5 commits the final answer. Splitting "analyze" from "write" — and crossing model families between them (Sol judge / Fable synth) — keeps the final answer honest rather than one seat defending its own draft. The judge can't be skipped and the synthesizer always runs last.

Step 0 — Verify the panel

bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/detect_panel.sh"

This panel needs both CLIs installed and authenticated: claude (the Fable 5 panelist and the Fable 5 synthesizer run as claude -p subprocesses, locked at max) and codex (the GPT-5.6 Sol panelist and the GPT-5.6 Sol judge, locked at ultra — needs a codex build and account with GPT-5.6 Sol access; tested against codex-cli 0.144). The detector prints CLAUDE=ready|missing, CODEX=ready|missing, and a final PANEL=ready|degraded. Handle a degraded panel explicitly (don't silently change it):

• If claude is missing from PATH: the Fable 5 panelist and synthesizer can't run at locked max — fall back to spawning the panelist with the Agent tool (session effort, not max) and have the orchestrator write the final answer itself (the Step 4 synth fallback).
• If codex is missing: the GPT-5.6 Sol panelist and judge can't run — the judge falls back to the orchestrator (the Step 3 judge fallback) and you lose the Sol panelist; tell the user and offer to proceed with the Fable-only seats.
• GPT-5.6 Sol access is NOT detectable up front; if the account can't run it, the codex seats fail at run time and are treated as absent (the panel degrades and says so).

Effort: every reasoning seat is a locked subprocess, so the Fusion session's own effort no longer affects output. The Fable 5 panelist and synthesizer are set to max by run_claude.sh (env knob + flag, as above). The GPT-5.6 Sol panelist and judge are locked at ultra in run_codex.sh. The orchestrator only coordinates — it writes the prompt files, launches the runners, and presents results; it does no judging or synthesis reasoning itself.

Step 1 — Preflight (informational, never a gate)

Write the user's task verbatim to a scratch file (you'll reuse it for the panel and the provenance record), then show the preflight estimate:

bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/preflight.sh" /tmp/fusion_question.txt

Show its output to the user (a rough token estimate for the fixed 4-seat panel, the per-seat timeout, and a Codex cap reminder), then proceed. It never blocks — it always exits 0, it only informs. Each seat is bounded by a per-seat timeout (FUSION_TIMEOUT, default 1800s) baked into the runners; raise it for heavy deep-research questions or big code merges (prefix the runner calls with FUSION_TIMEOUT=3600).

Step 2 — Fan out: 1 Fable 5 (max) + 1 GPT-5.6 Sol (ultra), parallel and blind

Read references/panel.md. Build the panelist prompt as the user's task verbatim plus a short instruction to research with web + bash and return a complete, self-contained answer as one of several independent experts who won't see the others' work, reasoning at maximum depth and answering it themselves (no delegation). Do not assign lenses; do not pre-digest the task. The same prompt goes to both seats — diversity comes from independence, not from different prompts.

Allocate one scratch dir, write the panelist prompt once, then launch both seats in the same turn so they run concurrently:

mktemp -d "${TMPDIR:-/tmp}/fusion-panel.XXXXXX"   # prints the panel dir — call it PDIR

Take the literal path it printed (a shell variable would not survive into your next Bash call) and use it everywhere below as PDIR. Write the verbatim panelist prompt to PDIR/prompt.md with the Write tool, and also write the user's verbatim question (without the panelist instruction wrapper) to PDIR/question.md — Step 5 records it. Then launch the two runners as background tasks — a panel at max/ultra routinely runs many minutes, well past a foreground Bash timeout, so do not block on them. Start both in the same turn (each its own backgrounded Bash call) so they run concurrently:

# run each of these as a separate background Bash task, in one turn (substitute the real PDIR):
bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/run_claude.sh" "PDIR/prompt.md" "PDIR/fable.md" max   claude-fable-5
bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/run_codex.sh"  "PDIR/prompt.md" "PDIR/gpt.md"   ultra gpt-5.6-sol

Each runner wraps its CLI call in a per-seat timeout (_fusion_lib.sh's _run_with_timeout, FUSION_TIMEOUT default 1800s), so no single seat can hang the panel. You'll be notified as each background task finishes — whether it completed, exited non-zero, or hit its timeout (exit 124). Read that seat's output file then. While they run, keep the user posted on which seats are still going so the run never looks frozen. Proceed to judging once both have finished or are confirmed absent. For a heavy deep-research question, raise the budget before launching (FUSION_TIMEOUT=3600 bash …/run_*.sh …) so a slow-but-valid seat isn't killed.

• Fable 5 panelist (×1) → run_claude.sh runs claude -p --model claude-fable-5 against a throwaway copy of the repo, exporting CLAUDE_CODE_EFFORT_LEVEL=max (highest-precedence effort knob) and passing --effort max, so the panelist hits max regardless of the session — this is the whole point of the runner (the Agent tool / agent-teams can't set per-call effort, so teammates just inherit config). No fallback model: if the seat fails permanently, it is absent.
• GPT-5.6 Sol panelist (×1) → run_codex.sh runs codex sandboxed (-s workspace-write) against its own throwaway copy, pinned to gpt-5.6-sol at ultra on codex's priority (fast) service tier — sees the project for context but never touches your live checkout. A different model family broadens the panel.

Never use fixed paths like /tmp/fusion_out.md; concurrent Fusion runs would clobber each other. Each out file is that seat's answer — a non-empty file is a success. A seat whose runner exits 124 (timed out), exits non-zero, or leaves an empty file is absent (note it; never treat it as silent agreement) — judge with the seats you have. Keep panelists isolated: never paste one panelist's output into another's prompt. Both panelists are independent cold runs in separate processes, and the judge and synthesizer are separate processes again, so every answer is read fresh downstream rather than any seat grading its own draft.

Step 3 — Judge (GPT-5.6 Sol via codex, ultra)

Once both panelist files are written (PDIR/fable.md, PDIR/gpt.md), hand them to a fresh GPT-5.6 Sol judge. Read references/judge_rubric.md, then you (the orchestrator) write a judge prompt file containing, in order: the original task verbatim, both panel answers (each labeled by seat — "Fable 5", "GPT-5.6 Sol"), and the judge instructions from references/judge_rubric.md. Allocate a fresh scratch dir:

mktemp -d "${TMPDIR:-/tmp}/fusion-judge.XXXXXX"   # prints the judge dir — call it JDIR

Take the literal path it printed (a shell variable would not survive into your next Bash call) and use it everywhere below as JDIR. Write the judge prompt to JDIR/judge_prompt.md with the Write tool, then run it as a background task (the judge reasons at ultra and routinely runs minutes — don't block on it):

bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/run_codex.sh" "JDIR/judge_prompt.md" "JDIR/judge_out.md" ultra gpt-5.6-sol

judge_out.md is the structured analysis: the judge classifies the deliverable (artifact → Track A, research → Track B) and produces the analysis, attributing every point to its seat ("Fable 5", "GPT-5.6 Sol"). The judge does analysis only — it does not write the final deliverable; the synthesizer does. A panelist that failed or was dropped is treated as absent, never as silent agreement. The judge is bounded by FUSION_TIMEOUT too; if the judge runner fails or times out (exit 124), fall back to producing the analysis yourself from the panel answers and say so.

Step 4 — Synthesize (Claude Fable 5 via claude -p, max)

Wait for the Step 3 judge background task to finish (i.e. JDIR/judge_out.md is written) before starting — the synth depends on the analysis. Then hand the judge analysis to a fresh Fable 5 synthesizer. Write a synth prompt file containing, in order: the original task verbatim, the full judge analysis (judge_out.md), and both panel answers (labeled by seat); end it with the synthesizer instruction from references/judge_rubric.md (Track A: run/merge the candidates in the trusted copy until they pass; Track B: derive the answer from the judge's five sections). Allocate a fresh scratch dir:

mktemp -d "${TMPDIR:-/tmp}/fusion-synth.XXXXXX"   # prints the synth dir — call it SDIR

Take the literal path it printed and use it as SDIR. Write the synth prompt (task + judge analysis + panel answers + synth instruction) to SDIR/synth_prompt.md with the Write tool, then run it as a background task (the synth reasons at max and routinely runs minutes — don't block on it):

bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/run_claude.sh" "SDIR/synth_prompt.md" "SDIR/synth_out.md" max claude-fable-5

synth_out.md is the final answer. For a code/artifact task the Fable 5 synthesizer runs and fixes the merged result inside its own throwaway repo copy (run_claude.sh works in a copy with bypassPermissions, so it can build/run/test) before emitting it; for research it derives the answer from the judge's sections. The synth ingests the pipeline's largest prompt (task + analysis + both answers) and is bounded by FUSION_TIMEOUT too, so for a big Track-A merge raise it (e.g. FUSION_TIMEOUT=3600) and, if a 1M-context variant is available on your account, set FUSION_CLAUDE_MODEL accordingly to give this seat the larger window. If the synthesizer runner fails or times out (exit 124), the orchestrator falls back to writing the final answer itself from the judge analysis and says so.

Step 5 — Save provenance

Record the run to an internal provenance file under ~/.claude/fusion-runs/ (the verbatim question + all raw panelist answers + the judge analysis + the final answer, timestamped, for auditing). Substitute the literal PDIR/JDIR/SDIR paths from the steps above:

FUSION_PANEL_NOTE="<degradation note, or empty>" \
bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/save_run.sh" \
  fable5-gpt5.6sol \
  "PDIR/question.md" "JDIR/judge_out.md" "SDIR/synth_out.md" \
  "fable5=PDIR/fable.md" "gpt5.6sol=PDIR/gpt.md"

save_run.sh substitutes a placeholder for any answer file that is missing or empty, so a degraded panel (an absent or timed-out seat) still produces a complete record. It prints the path it wrote — keep it for Step 6. Privacy: the file holds the raw question and raw answers in cleartext; it's written 0600 under a 0700 dir. For sensitive or confidential work, set FUSION_NO_SAVE=1 (it skips the record entirely, nothing hits disk) — surface that option if the task looks confidential. Set FUSION_PANEL_NOTE to the one-line degradation note when a seat dropped (e.g. "gpt5.6sol dropped: codex timed out -> fable-only"), or leave it empty for a clean run. If the judge ran on a fallback CLI (e.g. the Sol judge failed and you produced the analysis on Fable 5 instead), set FUSION_JUDGE_LABEL to the judge that actually ran so the record doesn't mislabel it; it defaults to GPT-5.6 Sol judge (the normal codex judge). Use a slug that reflects what actually ran (fable5-gpt5.6sol for the full panel, or e.g. fable5 if the Sol seat dropped, gpt5.6sol-only if the Fable seat dropped).

Step 6 — Present

Lead with the final answer (the synthesizer's synth_out.md), then the audit trail beneath it: the per-seat attribution and the judge analysis (Track A: what each candidate did when run + what was merged and verified; Track B: the five sections). Name the panel you ran (1× Fable 5 panelist at locked max + 1× GPT-5.6 Sol panelist at locked ultra, judge = GPT-5.6 Sol (codex, ultra/fast tier), synth = Fable 5 (claude -p, max)) — every seat is a locked subprocess, so there's no session-effort caveat to flag. If a seat dropped (missing CLI, non-zero exit, or FUSION_TIMEOUT exit 124) and you ran a degraded/fallback panel, say so and how to enable the full panel. Unless FUSION_NO_SAVE was set, mention the provenance record path save_run.sh printed (under ~/.claude/fusion-runs/) so the user knows where the audit trail lives.

Max panel variant — /fusion-max

The fusion-max command runs the SAME pipeline with one extra Claude seat. Only the panelist count changes; efforts are identical and stay locked (Claude seats max, codex seats ultra), the GPT-5.6 Sol judge and the Fable 5 synthesizer are unchanged, and every rule in this file (independence, timeout = absent, provenance) applies as written. Differences:

Seat	default (/fusion)	max (/fusion-max)
Panelist 1	Fable 5	Fable 5 #1
Panelist 2	GPT-5.6 Sol	Fable 5 #2
Panelist 3	—	GPT-5.6 Sol
Judge	GPT-5.6 Sol	GPT-5.6 Sol
Synthesizer	Fable 5	Fable 5

Concretely, in Step 2 launch two independent Fable 5 cold runs (use fable1.md / fable2.md) alongside the Sol seat:

bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/run_claude.sh" "PDIR/prompt.md" "PDIR/fable1.md" max   claude-fable-5
bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/run_claude.sh" "PDIR/prompt.md" "PDIR/fable2.md" max   claude-fable-5
bash "${CLAUDE_PLUGIN_ROOT}/skills/fusion/scripts/run_codex.sh"  "PDIR/prompt.md" "PDIR/gpt.md"    ultra gpt-5.6-sol
# ... judge exactly as in Step 3 (GPT-5.6 Sol, ultra), synth exactly as in Step 4 (Fable 5, max)

Label the seats "Fable 5 #1" / "Fable 5 #2" / "GPT-5.6 Sol" in the judge and synth prompts. Two cold Fable 5 runs still diverge (different reasoning paths, tool calls, sources) — the judge weighs their agreement as same-model agreement, and Fable-vs-Sol agreement as the stronger cross-family signal. Step 5 slug: fable5x2-gpt5.6sol for a clean run, with answer labels fable5-A=PDIR/fable1.md fable5-B=PDIR/fable2.md gpt5.6sol=PDIR/gpt.md. The codex-side runner selects this variant with FUSION_PANEL_PRESET=max.

Cost & latency note

A panel costs roughly N× a single answer in tokens and runs as slow as its slowest seat, plus a judge and a synthesizer pass. Each seat is bounded by FUSION_TIMEOUT (default 1800s) so a stuck seat degrades the panel instead of hanging it; raise it for deep research or big merges. Every reasoning seat works against its own throwaway copy of the repo (an rsync per seat) — the panelists, the Sol judge, and the Fable synthesizer — so a run makes up to four copies (five with /fusion-max), bounded by repo size. That's the deliberate trade: you spend more to stop being confidently wrong where that is expensive. For quick or low-stakes questions, a single direct answer is the right call.