Every generated Perl VFS package includes a local conversational assistant consisting of a pure-Perl neural intent classifier, a hashed TF-IDF vectorizer, and reviewed answer articles. It can answer Perl, Federation Image v2, Monk policy, Monk Loom, Monk Spindle, generated runtime, and branch-agent questions without Python or a network.
Three separate roles
| Component | Purpose | Authority |
|---|---|---|
| Main Perl assistant | Classifies questions and retrieves reviewed guidance. | Advisory only. |
| Specialist branch agent | Answers a narrower user-selected domain under /agents/ID; a reviewed article may return a fixed workspace-action recipe. | Training and publication require an explicit host command. Workspace actions remain proposal-only. |
| Monk Loom attachment | Performs deterministic fixed-point inference bound to a Monk profile in Federation Image v2. | Advisory evidence; it cannot rewrite policy. |
Saving AI settings does not train a model. Building a package does not silently train a model. A custom main candidate must be trained, measured, and explicitly approved first.
Conversation behavior
The classifier produces intent probabilities. Retrieval combines the intent result with article similarity. When confidence, winner margin, or retrieval score is too low, the assistant asks for more detail rather than inventing an answer.
Workspace action behavior
The model does not generate commands from arbitrary text. A selected knowledge article may carry an exact, reviewed workspace_actions array. The response exposes that array as data. Before registration, Studio verifies the branch bundle's trusted unexpired Ed25519 signature, independent benchmark, accuracy floor, and regression limit. It then binds the action plan to one workspace configuration, dry-runs it, and requires separate named acceptance and application. Follow Train an agent to propose workspace changes end to end.