v0.4.0

or/agent — agent-core complete

This release rounds out the or/agent package: the stateful Agent now exposes live run state, accepts more per-request options, and resumes more flexibly, plus hardening for misbehaving hooks and tools. The or/llm package is unchanged and backward compatible.

Highlights

• Live run state — Snapshot() now reports StreamingMessage and PendingToolCalls, and the transcript grows incrementally during a run, so a concurrent reader sees progress without replaying events.
• Flexible resume — when the transcript ends with an assistant message, Continue falls back to queued messages: it drains the steering queue first, then the follow-up queue, instead of erroring.
• Request tuning — Options.StreamOptions passes Temperature, MaxTokens, Headers, and the OnRequest / OnResponse observers through to every turn.
• Multimodal prompts — new agent.UserMessage(text, images...) builds a user message from text and images for Prompt, Steer, or FollowUp.
• Robustness — a panic in a hook or tool is now recovered into an error result / terminal error event instead of crashing the process; a cancelled run stops executing between sequential tools.

Behavior change

• The default steering and follow-up queue mode is now QueueOneAtATime (previously QueueAll). Each drain point injects only the oldest queued message; set SteeringMode / FollowUpMode to QueueAll for the previous behavior.