Included in this PR

Resolves #798

active_run handle on Runner.run and Runner.run_streamed

cancel(reason) for a clean stop

inject(items) included on the next turn

state() for a tiny debug snapshot

Streamed runs emit a cancelled status event and close the queue; non-streamed return a terminal result without final output.

Public APIs stay the same; tracing and hook order are preserved for normal completions.

Tests

New tests cover cancel, inject, and status emission in both streamed and non-streamed paths.

RunState + run_single_step (next step)

What I'm thinking is that this PR lays the groundwork for a proper step API while solving the immediate problem of cancellation and live input injection. The follow-up I’m proposing could perhaps follow this path (all open to discussion!):

Introduce a light RunState that captures current agent, turn, items, and bookkeeping.

Expose run_single_step(state) -> new_state that does “call model → run tools/handoffs → decide next”.

Keep Runner.run* as a thin loop over run_single_step for backward compatibility.

Add step-level tests and examples so folks can drive agents interactively.

If this direction sounds good, I can open a focused PR right after this one that sketches RunState and a minimal run_single_step while maintaining today’s surface.

Happy to make changes to better align with roadmap.