@vendoai/agent

Runs Vendo's streaming, tool-calling conversation loop against any AI SDK LanguageModel, with persisted threads and guard-bound tools.

Threads persist through the composed store; without a configured store the agent defaults to core's in-memory reference StoreAdapter through the same single path. agent.evictSubject(subject) list-and-deletes the subject's thread rows — the umbrella calls it for every subject the store's idle sweep returns (usually a no-op, since the sweep's erase cascade already removed those rows).

stream({ signal }) cancels a turn: the in-flight provider call aborts, no further step starts, and the thread stays consistent and resumable (the umbrella passes the request's own signal, so client disconnect cancels the loop). The per-turn step cap is context.maxSteps (default 20); exhausting it streams a renderable data-vendo-step-limit part instead of ending silently.

Approvals a conversation walks away from resolve guard-side: a fresh user turn marks undecided asks abandoned in the thread and denies them through guard.abandonApprovals when the guard provides it. Client upserts are validated — a new message must be user-role, and an existing assistant message may change only by answering a pending approval.

Read Prompts and the architecture overview.