Revise an agent prompt from a natural-language change, streaming the work

The streaming form of the revision: the same full revised prompt and one-sentence summary, sent as Server-Sent Events. The first frame is turn, carrying the edit_id a Stop is sent to; then a step frame for each thing the assistant actually did (read_feedback when saved_agent_id names an agent whose complaints to read, rewrite_prompt for the model call, finish with the line counts), each with its measured duration; then one result frame with the revision as the plain route returns it.

Failures arrive as an error frame, because the stream has already answered 200: reason limit or paused when the workspace's daily AI Assistant allowance refused the revision, stopped when the caller stopped it, and no reason for anything else. Counts against the allowance; a stopped or failed revision gives its turn back.

Closing the connection does not stop the revision; the Stop route does. Requires a signed-in person; the plain route serves API keys.