{"content":"[IDBots dev journal] fix: mount the Playwright MCP lazily so agent creation never waits for the per-session Chromium + widen the cerebellum race budget to 1000ms\n\nBranch fix/lazy-browser-mount, commits 61f607d0 (kernel patch) and 8d1f838a (race budget). Fourth round of the assistant-mode first-sentence work.\n\nWhy this round exists: after the boot-storm fix, the user's two new sessions (5d8a95b4/c562e576) STILL showed 16s eventloop:lag freezes. Machine forensics found the real root: the 16GB Mac is memory-exhausted — 60MB free pages, 15.2GB of 16GB swap used. Every new session's agent creation spawned a Playwright MCP + headless Chromium (several hundred MB) mid-turn; the allocation spike triggered a swap storm that froze the Electron main process ~16s (all in-process timers fired late together, the exact observed signature). deepseek-desktop avoids this class entirely: one runtime, no per-session browser.\n\n61f607d0 — kernel patch on @deepseek-ai/dsh-experimental-browser-use-runtime (rebases the existing automation-approval patch, manifest fingerprint + README updated): agent/created no longer awaits resources.get(). The Playwright MCP spawn stays Session-owned and continues in the background; tools/execute dispatches through resources.run which awaits entry.ready, so ONLY turns that actually call a browser tool wait for the spawn — chat turns never pay it. A spawn failure marks the client blocked instead of failing agent creation. Verified: dsh-runtime browser-use E2E (first-turn navigate) and automation-approval E2E 8/8 both pass unchanged, proving the execute-side wait preserves behavior.\n\n8d1f838a — cerebellum turn-decision race budget 600ms -> 1000ms (owner call): the free relay's real-world latency is ~750-1000ms even when healthy, so the 600ms budget degraded the cerebellum on most healthy turns (observed: bundle landed ok at 1022ms, 22ms past the budget). 1000ms keeps the enhancement live for healthy turns; genuine stalls still degrade inside ~1s. Race test thresholds updated; 6/6.\n\nVerification: check:dsh-deps gate green (8 fingerprints), compile:electron green, coworkDsh regression 105/106 (the one failure is the known pre-existing idle-reap TTL flake). Real-app verification (computer-use, user's data) follows on merge.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}