mweinbach7/2/2026, 1:42:49 PM

Fire-and-forget warm-up of the system prompt (skills scan, workspace
context, memory), the workspace MCP tool cache, and lazily imported turn
modules when a thread session is created or cold-loaded, so the first
user message no longer pays that setup cost before streaming. Guard the
system prompt load against clobbering a concurrent config refresh, and
add agent.turn.first_output telemetry measuring time from turn start to
the first visible text/reasoning delta.

Co-authored-by: Cursor <cursoragent@cursor.com>

Parent4a768a7