researchlisted

# research One axis: **where the answer lives.** | the answer is in | invoke | reference | |---|---|---| | this repository | *(default)* | `references/codebase.md` | | documentation, an API, a spec, a standard | `--web` | `references/web.md` | ## Default — the codebase Splits the goal into 3-7 staged hypotheses, routes each stage by complexity, fires up to 16 agents concurrently, then cross-validates the returns against each other for contradictions and gaps. Local only: file reads, grep, bash, git. **It does not search the web**, and that limit is the point — a codebase question answered from a blog post is worse than no answer. ## `--web` Decomposes one question into 2-5 parallel lookups against documentation, APIs, specs and standards. Use it when the answer cannot be in this tree: a library's current behaviour, a protocol's wording, whether an API still has the shape your code assumes. ## Not this skill - **"Where is the important code" / "what is this repo"** �� `orient`. It ranks symbols by PageRank over tree-sitter tags and returns immediately with no reasoning. `research` forms hypotheses and dispatches agents; using it to get oriented is paying for a cross-validation you did not need. - **"What could we do about X"** → `plan --stage ideas`. That is idea generation with research rounds inside it, not research with an idea at the end. - **"Why is this broken"** → `diagnose`. Competing hypotheses against each other and one decisive probe, not a survey.