Fusion deck

fusion-deck

A premium decision-and-execution toolkit. It does two things well: think hard (fan a question out to the strongest models in parallel, blind, and let Claude (the session model) judge) and orchestrate (turn a vague ask into a verifiable contract, curate context, decompose into scoped subagent work, verify each step, and hand off cleanly). It is a Claude Code skill — markdown commands plus small bash/python helpers. There is no MCP server, no UI, no OpenRouter, no model-config platform.

The three lineages it fuses

It fuses three lineages — panel→judge (fusion-fable), contract+honesty (goal-meta; not Codex /goal), and context+orchestration (RepoPrompt CE) — see README → "The three lineages".

The panel and its honest-degrade rule

The PREMIUM panel is the full triple. Availability is reported by scripts/detect_panel.sh as a PANEL_STATE:

PANEL_STATE	Panel	Slug
PREMIUM	Claude + GPT-5.6 Sol + Gemini 3.1 Pro (agy by default)	claude-gpt5.6sol-gemini3.1pro
DEGRADED_CLAUDE_GPT	Claude + GPT-5.6 Sol	claude-gpt5.6sol
DEGRADED_CLAUDE_GEMINI	Claude + Gemini 3.1 Pro	claude-gemini3.1pro
CLAUDE_ONLY	two cold Claude runs	claude-x2

The cardinal rule: never silently fake the premium triple. Premium commands call scripts/assert_triple_panel.sh, which hard-fails unless codex and a Gemini backend are present. An operator who knowingly wants a smaller panel sets FUSION_ALLOW_DEGRADED=1 — then the run proceeds but is loudly marked degraded. The rule holds at runtime too: if a panelist fails mid-run (rate limit, timeout of FUSION_PANEL_TIMEOUT, implausibly tiny output), the runner writes the honest manifest and exits 13 unless degrade was explicitly allowed — stop and disclose, never silently continue. Every panel answer must disclose the PANEL_STATE it actually ran. See references/degraded-mode.md.

Commands — and where the panel is worth its cost

A triple panel costs ~N× tokens and runs as slow as its slowest panelist. So fan out only where independent cross-checking changes the answer's risk profile; everywhere else, one model is correct.

Command	What it does	Panel by default?
/fusion	Fan a hard question to the panel; Claude judges and writes the answer. --wide adds a second cold Claude (4 panelists).	Yes
/fusion-auto	Route a task through v2: pick workflow, verify, and escalate only when needed.	Router decides
/fusion-ultra	Max-quality two-round workflow: wide blind panel (Claude ×2 + GPT + Gemini) → contradiction matrix → targeted probes → verifier.	Yes + targeted probes
/fusion-review	Audit code/a plan via the panel; structured, cross-checked findings.	Yes
/fusion-investigate	Evidence-first root-cause investigation; panel adjudicates competing hypotheses.	By exception — only when ≥2 hypotheses survive the evidence; --panel forces it
/fusion-plan	Turn a vague request into a Claude Code Workflow Contract; --deep adds involvement + a critique pass.	No — single model; --panel to escalate a genuinely ambiguous, high-stakes planning question
/fusion-context	Build a RepoPrompt-style Context Pack; --discover adds evidence-gated agentic curation.	No — curation is mechanical
/fusion-orchestrate	Decompose, dispatch scoped subagents, verify each, roll up; --worktrees isolates parallel items.	No — single-model orchestrator; --panel to review a thorny decomposition
/fusion-optimize	Measure→change→re-measure loop; the panel calls continue/stop at each decision point.	By exception — only at the stop/continue decision points
/fusion-refactor	Structure analysis → behavior-preserving plan → steer-one-agent execution.	No — composes review/plan/orchestrate
/fusion-handoff	Emit a Handoff Capsule.	No — summarization
/fusion-remind	Re-anchor a drifting session: cheat-sheet of situation→command + the invariants.	No — pure recall

All twelve install as ~/.claude/commands/<name>.md wrappers (see README → Install); the whole skill is also invocable as /fusion-deck. (If a separate skill named fusion is also installed, that skill takes precedence for /fusion — this skill does not assume one is present.)

Each command file in commands/ is the procedure; it loads the matching references/*.md on demand.

Routing — match the situation to the command

When the user describes a situation instead of naming a command, map it to the fitting /fusion-* and offer it — they shouldn't have to know the catalogue. Suggest, don't silently run: the panel and orchestration commands cost quota and time, so name the command + why and let them say go. For a quick factual question, just answer directly — don't route a trivial ask into a panel.

<!-- SOURCE OF TRUTH for routing. On change: sync README.md EN + 中文 tables and commands/fusion-remind.md. -->

The user is…	Offer
stuck on a hard call / trade-off, or wants one cross-checked	/fusion
asking "choose the right fusion workflow for this" or wanting cheaper/faster unless risk says otherwise	/fusion-auto
asking for maximum quality / strongest possible answer / hard high-risk decision	/fusion-ultra
asking you to review code, a diff, or a plan before it ships	/fusion-review
reporting a bug, or asking "why is it built like this / where does this come from"	/fusion-investigate
handing you a vague, underspecified ask to build	/fusion-plan (--deep for a full design doc)
about to send code to another model/agent, or working an unfamiliar repo	/fusion-context (--discover to auto-curate)
asking for a big, multi-step change done carefully	/fusion-orchestrate (--worktrees for parallel siblings)
asking to make something measurably faster / smaller / cheaper	/fusion-optimize
asking to clean up or consolidate code without changing behavior	/fusion-refactor
wrapping up and passing work on (to another agent or future-them)	/fusion-handoff
drifting in a long session / a fresh agent needs the map and the invariants	/fusion-remind

Compose, don't silo: a feature is usually /fusion-plan → /fusion-context → /fusion-orchestrate → /fusion-handoff; a bug is /fusion-investigate → /fusion-plan → /fusion-orchestrate. If two fit, prefer the cheaper/narrower one and say what the other would add.

Core invariants (do not violate)

1. Never fake premium. Disclose the real PANEL_STATE; degrade only explicitly. (degraded-mode.md)
2. Blind panel. Panelists never see each other's work; only the judge sees all answers, and only after every panelist returns. Claude always judges. (panel-prompt.md, judge-rubric.md)
3. Orchestrator never implements. In /fusion-orchestrate, the orchestrator only plans, decomposes, delegates, and verifies; it reads files only to verify and makes all code changes inside Task subagents. One level of fan-out. (orchestration-rubric.md)
4. Verify, then dispatch fresh. Check each item against its Done-when with a concrete probe (grep / read / test — not a skim) before moving on. Never proceed with unresolved gaps.
5. v2 routes, it does not guess. /fusion-auto chooses a workflow only, writes a ledger entry, verifies when possible, and escalates by explicit policy. It does not replace /fusion's full-panel meaning.
6. Contract, not /goal. /fusion-plan emits a Workflow Contract; scripts/lint_contract.py rejects any /goal reference. (workflow-contract.md, contract-lint-rules.md)
7. Honesty path. Use the status states [todo]/[doing]/[done]/[blocked]/[incomplete]/[abandoned]; [incomplete] carries reason / proof / attempted / impact / next-decision. Report failures plainly.
8. Safety. Never hardcode keys, accounts, or private paths; never leak secrets into a Context Pack; the smoke test never calls paid models (even under FUSION_LIVE=1 — that flag only marks the mode for the commands). Panel prompts embedding untrusted content (a diff under review) run with FUSION_NO_WEB=1 so injected instructions have no exfiltration path. (references/safety.md, references/panel-prompt.md)
9. Honest-degrade beyond the panel. The new helpers obey the same rule: codemap.sh discloses CODEMAP_STATE and falls back tree-sitter→ctags→grep; selection_lint.py gates discovered context on evidence (S007); --deep/--discover/--worktrees are opt-in. Use the best available, disclose what ran, fall back loudly. (references/codemap.md, references/context-discovery.md, references/worktrees.md)

Skill root & paths

When a command or reference says scripts/… or references/…, that path is relative to this skill's root directory (the folder containing this SKILL.md) — resolve it against that directory, e.g. bash <skill-root>/scripts/detect_panel.sh, not against your current working directory. The installed location is ~/.claude/skills/fusion-deck/ (or wherever CLAUDE_SKILLS_DIR points); the generated /fusion-* command wrappers pass you this absolute path.

Quick start

# from the skill root (resolve these against <skill-root> if your shell is elsewhere):
bash scripts/detect_panel.sh        # what panel can this machine run right now?
bash scripts/smoke_test.sh          # offline self-check (no paid calls)