feat(agent): add single-step execution and mid-turn resume

[moedash]

What changed?

Step-level: runLoop's inner body moves into runSingleTurn, and agentStep runs exactly one turn without adding a message. Agent.step() and AgentSession.step() expose it. A step is a turn, so it emits no agent_start, and the session still does its post-run handling each step, so retries and compaction keep working.

Starting and recovering a turn without running it: AgentSession.recordPrompt() builds the turn's messages the way prompt() does and records them, with no model call. AgentSession.prepareStep() settles the tool calls a stopped turn left with no result (findDanglingToolCalls() finds them) and drops a trailing assistant message that holds no answer, returning whether the turn still has work.

Resume: AgentSession.resumeInterruptedTurn() is prepareStep() plus a run to the end of the turn, for a caller that wants the whole turn back in one call. It never re-adds the prompt.

Why?

Driving pi from an external durable executor means checkpointing between steps: start a turn, advance it one step, and pick up a half-finished one, without ever running a whole turn. prompt() runs to agent_end, so it could only do the first of those.

After a hard kill during a tool call, the session ends with an assistant tool-call and no result. That payload is still valid, transformMessages handles it. What you can't do is finish that turn: continue() refuses a trailing assistant message, so the only way forward is a new prompt, which changes the conversation.

Two details worth flagging. Only the trailing message counts, because an aborted tool batch leaves older calls unresolved on purpose and a result appended at the tail would attach to the wrong call. And a settled call reports the outcome as unknown, not failed, since a crash after the tool ran but before its result landed means the side effect may already have happened.

This can't be an extension. The guard that blocks the resume is in the core loop.

How did you test it?

• Unit Tests
• Staging
• End to End Tests

npm run check passes. ./test.sh has pre-existing find tool failures that need the fd binary; they fail the same way on main.

End to end: ran pi under a Temporal worker, killed it with kill -9 mid tool call, and the retry finished the turn with the prompt not re-added. That run drove resumeInterruptedTurn(); recordPrompt() and prepareStep() are covered by unit tests so far.