What problem does it solve? Onboarding a new batch inference provider (e.g. Anthropic, Gemini, xAI) requires exhaustive, evidence-backed knowledge of its API before any adapter code is written. This Skill replaces scattered doc-reading and ad-hoc cURL probing with a structured research phase that produces a single committed note serving as the contract for all downstream implementation and audit work. ## Core Features & Use Cases - Cited capability matrix: Produces a 14-row research note covering auth, endpoints, status models, output/error shapes, limits, retention, pricing, credential shapes, and remote URL support, with every claim backed by an official doc URL or live capture. - Live capture workflow: Runs a canonical capture matrix of batchable-request scenarios (text, tool-calls, reasoning, structured-output, truncated) against the real provider, redacts transcripts, and promotes load-bearing shapes into committed fixtures. - Completion and failure semantics: Answers how the provider signals batch completion, how failed sub-requests appear, and maps each fact to a concrete adapter seam such as pollBatch, parseResult, and parseUsage. - Use Case: Before adding Gemini batch support, run this Skill to generate docs/batch-research/gemini.md with the status mapping, skin reuse decision, fixture plan, and deferred captures, then present it for human sign-off. ## Quick Start Ask the agent to research a new batch provider by name, for example: run the research-batch-provider skill for Anthropic and produce the committed research note with live captures.
