What problem does it solve? AI agents often write Python code from training-data memory, which breaks when the installed library version has renamed, re-signatured, or deprecated a symbol. This Skill forces every symbol lookup to happen against the actually installed version and caches the verified result so the work is never repeated. ## Core Features & Use Cases - Four lookup shapes by question type: cache hit first, then inspect.signature + pydoc.render_doc symbol cards, dir/pkgutil module surface dumps, LSP hover via Pyright, and WebSearch/WebFetch of version-pinned narrative docs. - Version-keyed cache: every finding lands in scratch/api/<lib>/<version>/<topic>.md with a source line, verbatim extracts, and an agent-synthesized Usage section (Call / Don't call / Trap / Returns). - Named version traps: ships a list of known renames and footguns (e.g. skrub tabular_learner → tabular_pipeline, skore Project.get by id not key) so agents check the right way before writing code. - Use Case: Before writing skore.evaluate(learner, data={...}) in an ML experiment script, the agent resolves the installed skore version, checks the cache, runs a probe if needed, and records the verified signature — instead of guessing from memory. ## Quick Start Ask the agent to look up the signature and docstring of a function in the installed version of a Python package and cache the result for future sessions.
