Multi-Model Team

Set up multi-model-team — create your personal roster at ~/.claude/mmt-roster.json (seeded from the shipped default) so your tuning survives plugin updates, and print the next steps (roster precedence + statusline HUD).

Run a question through the multi-model Fusion pipeline — fan it out to a configurable panel of models in parallel (default panel is the user-configurable set from `reasoning.panel` in roster.json, defaulting to Opus+Sonnet+Gemini; override per-invocation with a spec like "2:gemini,opus,codex"), have a judge compare their answers into structured analysis (consensus / contradictions / unique insights / blind spots), then synthesize one unified answer that is better than any single model's.

Dry-run the multi-model-team router on a task and print the decision JSON + score breakdown. No model call — pure routing logic, for tuning.

Run a task through the multi-model team pipeline — decompose into backend-assigned subtasks (commodity → parallel agy/Gemini, judgment/hard-line → native Claude), dispatch dependency-aware, verify each result, fix failures in a bounded loop, then synthesize. Optional caps like "5:gemini,2:claude". Add --writable to give each agent its own git worktree and merge into an integration branch.

Dispatcher for the agy (Gemini) CLI backend — a configurable, equal tool, not a fixed task bucket. WHERE work routes is decided by config/roster.json (routes + tags.txt) and per-subtask /team assignments, not hardcoded here. By the shipped routing config, agy is the default lane for fast, commodity, easily-verifiable work and Gemini's edges: new UI/React components, CSS, SVG/animation, scaffolding, CRUD, REST endpoints, scripts, CLI tools, glue code, SQL, regex, configs, Dockerfiles, unit tests, fixtures, data transforms, web-search/doc summarization, bulk text, and audio/video (Gemini-exclusive). The shipped rules keep reverse-engineering, IL2CPP/protobuf-RE, disasm, FFI/unsafe, injection/hooking, concurrency, protocol/KCP design and other systems-hard work on native Opus — but that is roster policy you can retune, not a property of this agent. Spawning this agent is an explicit choice to run on agy: it relays the task to the CLI and returns the result verbatim; it does not re-route or refuse based on content.

Dispatcher for the codex (OpenAI Codex CLI) backend — a configurable, equal tool, not a fixed task bucket. WHERE work routes is decided by config/roster.json (routes + tags.txt) and per-subtask /team assignments, not hardcoded here. By the shipped routing config, codex is the default lane for code review, test-writing, and verification — review a diff/file for correctness, bugs, edge cases and regressions; write or extend unit/integration/e2e test suites; verify an implementation meets its spec — and it serves as the default /team verifier. It runs full-auto in its own git worktree under /team --writable, so it can also make real code changes there. The shipped rules keep reverse-engineering, IL2CPP/protobuf-RE, injection and other systems-hard work on native Opus, but that is roster policy you can retune, not a property of this agent. Spawning this agent is an explicit choice to run on codex: it relays the task to the CLI and returns the result verbatim; it does not re-route or refuse based on content.

3 hooks across 2 events