genius

# GENIUS — debate pipeline entry **Plugin paths.** This is a globally-installed plugin, so its bundled files live under the plugin root, NOT the current working directory. Before launching the workflow, resolve the absolute plugin root once: run `echo "$CLAUDE_PLUGIN_ROOT"` (Bash). Call the resolved value `PLUGIN_ROOT`. The bundled workflow is `PLUGIN_ROOT/workflows/debate-engine.js`. **Run artifacts (`runDir`) go in the user's CURRENT project** as a RELATIVE path (the workflow's safety guard rejects absolute paths and `..`). When invoked with an idea (the user's text after `/genius`): 1. Make a slug from the idea (lowercase, alnum → hyphen, collapse repeats, trim to ≤40 chars). Set `today` to today's date `YYYY-MM-DD` and `runDir` to a relative path in the current project — `genius-runs/<today>_<slug>` (or, if the project has a `results/` dir, `results/genius/<today>_<slug>`). 2. Optional knobs the user may state: `roster` (a subset/superset of families `claude`/`codex`/`kimi`) and `cap` (debate rounds, default 3). Otherwise use the default roster `[{ "family": "claude" }, { "family": "codex" }, { "family": "kimi" }]` and `cap` 3. 3. Resolve `PLUGIN_ROOT` (`echo "$CLAUDE_PLUGIN_ROOT"`), then launch the Workflow tool with `scriptPath: "<PLUGIN_ROOT>/workflows/debate-engine.js"` (the absolute resolved path) and `args: { idea, runDir, today, roster, cap }`. (The workflow creates `runDir` during its persist step, so you do not need to pre-create it. `args` is delivered to the workflow as a JSON string; that is expected — the workflow parses it.) 4. If the workflow result has `needsSharpening: true`, the idea was too vague — ask the user the single clarifying question implied by `storyLock` (what concrete domain / phenomenon / believed-but-unproven claim they mean), then re-launch with their answer folded into the idea. 5. Otherwise read `<runDir>/portfolio.md` and present it to the user. Note whether the debate `converged`, how many `rounds` ran, and where the run artifacts live (`portfolio.md`, `board.jsonl`, `concept_map.json`, `transcript.md`). The run is heavy (≈14 agents, paper-search-driven, tens of minutes). It runs in the background; if a transient API error or stall aborts it, simply re-launch (optionally with `resumeFromRunId`). ## EVOLVE gate (v1b — gated; runs after the debate portfolio) After presenting `<runDir>/portfolio.md`, offer EVOLVE for a **computable** candidate: 1. Ask the user which portfolio item (consensus pick or a protected outlier) to evolve, or to skip. EVOLVE only fits a candidate that reduces to runnable code + a metric; if none, stop (debate portfolio is final). Prefer self-contained, single-file, fast-running candidates first (the loop is proven on those; multi-file ML/bio pipelines are a later validation rung). Confirm prerequisites exist (see `PLUGIN_ROOT/docs/EVOLVE-SETUP.md`). 2. In `<runDir>/evolve/<slug>/`, draft two files: - `initial_program.py` — the candidate mechanism as runnable code, with the mutable region wrapped in `# EVOLVE-BLOCK-START` / `# EVOLVE-BLOCK-END`. - `evaluator.py` — a frozen metric exposing `def evaluate(program_path) -> {"combined_score": float, ...features}` (plus a `__main__` that prints it as one JSON line). 3. Show BOTH to the user; have them review/edit until the evaluator measures the right thing. **Do not proceed without explicit approval** — the frozen metric is the disposer and is only trustworthy if it is correct, which is the user's call. Warn that candidate code will execute locally (openevolve runs it in a process pool, not a sandbox). 4. Freeze: `shasum -a 256 <runDir>/evolve/<slug>/evaluator.py | awk '{print $1}' > <runDir>/evolve/<slug>/frozen.lock`. 5. Run: `bash "$CLAUDE_PLUGIN_ROOT/scripts/run-openevolve.sh" <runDir>/evolve/<slug> --iterations <N>` (default N=30; raise per appetite; the run is wall-clock bounded by the wrapper). If it exits non-zero, report the reason (frozen mismatch / missing venv / no best produced) and stop. 6. On success, read `<runDir>/evolve/<slug>/out/best/best_program_info.json` and merge a `## 📊 Measured winners` section into `portfolio.md`: the evolved artifact path (`out/best/best_program.py`), its `combined_score`, the metrics, the iteration, and the evaluator hash (`frozen.lock`). Unmeasured items stay as they were.