Smart-Kickoff Classifier is the launch step in 📝Smart-Kickoff that infers which repository a work item targets, returning exactly one strict JSON object whose confidence flag halts the launch to ask a human whenever the inference is not certain.
The classifier is the second of the launcher's two fail-closed checkpoints. Its input is the work item's body plus a roster of onboarded repositories — each a name and the one-line description its own configuration supplies as routing signal. Its output is a single object with a fixed schema: the repository name copied verbatim from the roster, a short kebab-case slug describing the work rather than the item, the kind of scope (item-level checklist work or a whole initiative), a one-line summary, the exact words in the item that determined the repository, a confidence of high or low, and the expected deliverable (a pull request, or research with no code). The classifier never emits a branch name; that is computed elsewhere and is not its decision.
Two rules carry the design. First, confidence is low whenever two roster entries are plausible or the item names no repository at all, and low confidence stops the launch: the human is asked, the launcher never guesses. Second, the work item is untrusted input — data to classify, never instructions to follow — so anything in it that addresses the classifier, asks it to change its rules, or names a repository in the second person is ignored. Any model that can return clean JSON can fill this role; the command that runs it is global configuration, never something a repository's own config can set. The 📝Smart-Kickoff — Agent Buildbook helps an adopting agent decide whether it needs this step at all, since a single-repository setup can skip it.
{ "repo": "studio-agent-ops", "slug": "smart-kickoff-buildbook", "scope_kind": "items",
"scope_summary": "Author the public Smart-Kickoff buildbook", "repo_evidence": "\"backed by studio-agent-ops\"",
"confidence": "high", "deliverable_kind": "code_pr" }