mirror of
https://github.com/heygen-com/hyperframes.git
synced 2026-11-19 11:26:02 +00:00
Cancelling asks the agent to stop with session/cancel before its process is taken away, so it can put its tools down rather than have them removed mid-write; the kill still follows on a short leash for an agent that does not take the hint. Correcting is where the two transports genuinely differ. Claude Code reads a correction into the turn it is already having. ACP has no way to speak into an open turn, so a correction cancels this one and prompts again on the same session: one job, one session id, everything the agent has read still in hand — but it does start the turn over, and the run says exactly that rather than implying the correction was slipped in. Steering an ACP run also used to build the harness' native resume flags for a command that is not the harness, so that path is closed. Studio's own verdict on a run now outlives the agent acknowledging it: a run failed for an unanswered question was being relabelled cancelled by the stop reason its own cancel produced. Verified against @agentclientprotocol/codex-acp: a live run told to make the text BLUE instead of RED came back as one run, one session, the steer recorded, and BLUE on disk.