research-classifierlisted

# Research classifier The cheapest, most-aggressive-to-skip gate before research fires. Bias toward SKIP. Over-triggering kills UX; under-triggering makes the feature invisible. Default outcome on most tasks is `RESEARCH_SKIP`. Two reference files carry the detail this body points to: - [`STRATEGY.md`](./STRATEGY.md) — aggressiveness tuning, the stakes ladder, the capability survey, and the MCP-nudge rules. - [`EXAMPLES.md`](./EXAMPLES.md) — the curated input→outcome table, kept in sync by hand (see Examples). ## When to invoke Two fixed trigger points only — never every turn: - **Trigger A — pre-spec**: between `brainstorming` step 1 (read the room) and step 2 (clarifying questions). Input: the user's raw task text. - **Trigger B — pre-plan**: at `writing-plans` step 0.5, before any plan file is written. Input: the spec's `## Approach` section. Don't invoke for arbitrary user messages, for `/llm-orchestrator:debug`, `/llm-orchestrator:verify`, `/llm-orchestrator:finish`, or any non-design conversation. The hook layer does not enforce these trigger points — `orch-research-gate.sh` is content-addressed: it fires whenever a user prompt matches the shared signal patterns (design verb plus library/version/security token), wherever in a conversation that happens, and compels the classifier to run. The two fixed trigger points are a discipline held by this skill and its callers (`brainstorming` step 1.5, `writing-plans` step 0.5); the classifier itself only checks signals aga