Agent yield-and-resume flow for intent veto window

[marcus-sa]

Problem

When an agent creates an intent that requires human review (veto window), the agent has no protocol for pausing and resuming. The current options are all suboptimal:

• Polling: Burns tokens/compute while waiting minutes to hours
• Blocking: Keeps an LLM session alive unnecessarily
• Fire-and-forget: Agent loses context about why it initiated the action

Design

Yield-and-resume within the same agent session — the agent yields its current turn when an intent enters pending_veto, and Brain re-invokes it in the same session (new turn, full conversation context preserved) when the human approves or rejects.

Flow

Agent: creates intent → gets { status: "pending_veto", intent_id: "abc" }
Agent: yields turn, tells user "Awaiting approval for Stripe refund of $50"
Human: approves/rejects in feed UI
Brain: re-invokes agent in same session with { event: "intent_authorized", intent_id: "abc" }
Agent: exchanges intent for RAR token → calls MCP tool → completes

On rejection:

Human: rejects in feed UI
Brain: re-invokes agent in same session with { event: "intent_vetoed", intent_id: "abc", reason: "..." }
Agent: acknowledges rejection, communicates to user, adjusts approach

Key decisions

• Same session, new turn: Agent retains full conversation context (why it was refunding, what the user asked for). No need to re-explain.
• Re-invocation trigger: SurrealDB EVENT on intent status transition (pending_veto → authorized or pending_veto → vetoed) fires a callback that pushes a new turn into the agent session.
• No polling: The agent does not check intent status. The platform pushes the resolution event.
• Structured yield response: The intent creation tool returns a structured result that the LLM recognizes as "pause here" — not an error, but a signal that this action requires async approval.

Open questions

• How does the re-invocation interact with SSE streaming? Does the agent session's SSE connection stay open during the veto window, or does the client reconnect?
• Should the agent be able to do other work while waiting (e.g., answer follow-up questions), or is the session paused entirely?
• What's the maximum veto window duration before the session context is considered stale?
• How does this interact with composite intents (multi-step tool chains where step 2 needs approval)?

Context

• Research doc: docs/research/intent-rar-mcp-tool-gating.md
• Intent state machine: app/src/server/intent/status-machine.ts
• Veto manager: app/src/server/intent/veto-manager.ts
• Agent session model: app/src/server/agents/