Module agent_loop 

Mirrors packages/agent/src/agent-loop.ts — the provider-agnostic agent loop.

Two public free functions, run_agent_loop (new prompt) and run_agent_loop_continue (no new prompt), drive the outer follow-up loop and inner steering+tool loop. The only LLM boundary is StreamFn (sync return → AssistantMessageEventStream); everything else talks in AgentMessage.

Critical invariants enforced here (plan §5):

• Tool-execution ordering: in a parallel batch, ToolExecutionEnd fires in completion order; tool-result MessageStart/MessageEnd fire later in source/ordinal order. Implemented by collecting completion signals into a queue, then walking the finalized vec by index for the result messages.
• Truncate-fail: stop_reason == Length → every tool call in the message fails-in-place with is_error:true and is not executed.
• Late-update suppression: on_update after execute resolves is a no-op via an accepting_updates: Arc<AtomicBool> flipped false on settle.

The loop is fully testable without crate::Agent — run_agent_loop takes plain AgentContext/AgentLoopConfig + an AgentEmitter.

Enums

LoopOutcome

Why a run ended. Completed is the normal exit; Aborted/Failed are set when the terminal assistant message carried Aborted/Error.

Functions

run_agent_loop

Run an agent loop starting from a new prompt. Mirrors TS runAgentLoop.

run_agent_loop_continue

Continue an agent loop from the existing context (no new prompt). Mirrors TS runAgentLoopContinue. Errors if the context is empty or its last message is an assistant message (the provider would reject that).

Type Aliases

NewMessages

The messages a single run_agent_loop invocation added (prompt + assistant