Module agent 

The agent loop.

Ask the model, run whatever tools it asks for, feed the results back, repeat until it stops asking. Everything interesting — which provider, which tools, who approves side effects — is injected, so the same loop drives the REPL, a one-shot run, and a batch worker.

Structs

Agent

Budget

Per-run ceilings. Every None falls through to the agent’s own config, so a caller overrides only what it actually means to change.

Conversation

A conversation, and what has entered it.

RunContext

What the loop consults that is properly per-run rather than per-agent: what tools may touch, who approves the ones that aren’t read-only, and what this particular run is allowed to spend.

RunOutcome

Taint

What has entered this conversation so far.

ToolCallTrace

One tool call as it actually happened. The trace is what you grade a model on — final text alone can’t tell a lucky guess from correct tool use.

Enums

AgentEvent

Everything the loop wants to tell an observer. The CLI renders these; a batch runner ignores all but the last.

Phase

Which half of the work a run is doing.

StopCause

Why the loop stopped. Completed is the model deciding it was done; everything else is the harness cutting it short.

Functions

turns_phrase

“1 turn”, “3 turns”. These strings are read by people.