Prewarm Cloud machines while typing
HOLD: F1 passive wake retry regression requires final native verification and independent review.
An existing Cloud chat with a retained session ID now sends a best-effort wake hint on its first non-empty composer edit, overlapping machine resume with typing. Pending hints are deduplicated; accepted hints retain the two-second dedupe window. A hint refused while the inventory loads or rejected by transport can retry on the next edit. Successful native null replies count as accepted. The native bridge accepts a paused machine from the current signed-in inventory and preserves script origin; it consumes no gesture credit. Unsupported or failed hints leave typing and Send working normally.
This implements the existing-chat portion of the request to prewarm while typing before Send. Fresh drafts and missing-session recovery are tracked in #34828; deduplication across A ā B ā A chat switches is tracked in #34829. No visual behavior changes. The change is on by default and can be reverted independently.
Performance:
| metric | before | after | delta | harness or run link |
|---|---|---|---|---|
| wake hints emitted for the first non-empty edit of an existing Cloud chat | 0 | 1 | +1 | cloudTypingPrewarm.test.ts deterministic bridge-port harness |
This measures request emission, not machine readiness. First-keystroke-to-agent-ready p50/p95 is measured separately by the Cloud readiness probe; this PR does not claim those results.
Validation: nine focused typing-hint tests, including declined/rejected retry and pending/stale completion handling; native request decoding/passive-gesture tests and script-origin propagation coverage. Exact-head CI and review receipts are linked in the checks and review comments.
š¤ Generated with Codex
Agent session: agent session