Manage Assistant

Chat with Vruum: the same chat as the floating panel in the app.

• start_conversation: no id; payload={request_id (UUID), message, context?, browser_context?}. Starts a new chat and waits up to 50 s for the answer. context={kind: person|account|deal|objective, id} names a record to focus on; context={kind: people|accounts|deals, params} names a table and its filters (a saved view's keys). Returns chat_id, the turn (state completed, failed, stopped, interrupted, running or awaiting_approval), the question, the answer, changes and cards.
• append_message: id=chat UUID; payload={request_id, message, context?, browser_context?}. Sends the next message in that chat and waits the same way. One message runs at a time per chat.
• undo: id=a change's item UUID from changes or inspect_assistant read; no payload. Runs the change's exact inverse; refused when the record was edited since.
• decide_card: id=a pending card's item UUID from cards; payload= {decision: approved|declined}. Call only with the user's own decision on that exact card. approved runs the card's exact call once under your credentials; declined never runs it. Waits up to 50 s for the turn to continue; claimed=false means the card was already decided or expired.
• rate: id=an answer's message UUID; payload={rating: 1|-1|0, reasons?, comment?}. reasons: wrong_facts, did_not_do_it, wrong_record, too_slow, hard_to_follow, other. 0 clears the rating.
• job_start: no id; payload={request_id, request_key, request: {text, completion_condition, source_scope?, source_domains?}, guidance?}. Hands a request too large for one chat turn off to Vruum (shown to users as a Request in Agent work): it reads with your access, writes its analysis as the result and prepares drafts that keep their own approval. The whole workspace sees it; the chat stays private. The same request_key returns the same request. For public multi-page research, include source_domains (named public domains). This uses isolated browser research with search, fetch and rendered pages, retained sources and the same Request controls.
• job_control: id=Delegation UUID; payload={command: pause|resume|cancel| steer|takeover, request_id, expected_revision, guidance (steer only)}. Its grantor controls a request; a workspace admin pauses, stops or takes it over after the grantor loses access. Read it with inspect_assistant action=job.

browser_context={binding_id, snapshot_id} is an optional immutable Radar page reference, separate from CRM context. Keep it unchanged on retries. Read saved evidence with inspect_assistant action=browser_source.

Vruum reads your records and applies reversible record changes at once (notes, tasks, roles, contact choices, deal stages), each listed in changes with Undo. Other actions (archive, objective members, launch, a one-off send, paid lookups) become a card: the turn returns state=awaiting_approval and nothing runs until decide_card. It never approves drafts. A turn still running after 50 s returns state=running: read the chat later with inspect_assistant. Keep the same request_id when you retry after an uncertain response; the server runs each request once. Call only after the user asks for the message to be sent.