perf(desktop): Bot Mode wakes paint instantly — paint-first hydration + durable transcript cache (#89206 class)

[teknium1]

Summary

Bot Mode wake is now instant-feeling and spawn-proof, without raising any timeout — two layers:

1. Paint-first hydration: a history-bearing chat counts as hydrated the moment its persisted transcript is painted; the full runtime resume (agent build, MCP discovery, skill load) warms in background and binds the composer when ready.
2. Durable transcript-tail cache: a bounded tail (40 msgs / 256KB / 50-session LRU, localStorage, keyed by durable stored id) of every reconciled transcript persists across app launches and backend reaps — a cold wake paints at ~0ms before any network I/O, then the REST prefetch/runtime resume replace it with authoritative content. Display-only: reconciliation and failure latches treat the cached paint as empty, so it can never mask a stranded resume or graft onto stale rows; deletes drop entries, gateway re-homes wipe the cache.

Root cause (zero trust's third bundle, on ae6578af with both prior #89206 fixes present): cold profile backends on slower Windows machines take 47–120s to fully boot, and the wake path's fixed budgets (20s hydration + ~15s resume retries) raced the whole boot and lost — "errors waking up BOTS" while the backend came up healthy moments later. The transcript itself was available seconds in via the REST prefetch; the wait was serialized behind the slowest part of boot for no user-visible benefit.

Changes

• apps/desktop/src/sdk/index.ts — waitForFocusedSessionHydration: paint-first predicate (expectHistory ? transcriptPainted : runtimeReady). Expected-empty chats still wait for the runtime (nothing to paint; a bound runtime is the only proof the surface is real). On timeout, a [bot-wake] phase breakdown (activation ms, hydration ms, unmet conditions) is logged to the renderer console → desktop.log, so the next support bundle names the slow phase directly.
• hermes_cli/web_server.py — flush the headless "listening" line; block-buffered on Desktop's piped stdout it surfaced minutes late, inflating apparent boot times in support bundles (part of the 120s "gap" in this bundle was this artifact).
• apps/desktop/src/sdk/profile-routing.test.ts — paint-first regression test (slow-runtime shape).

Validation

Check	Result
Sabotage: restore runtime-gated wait	paint-first test FAILS by timing out (field shape)
profile-routing.test.ts	15/15
transcript-tail-cache tests (incl. sabotage: cache disabled fails 5/7)	7/7
session hooks + sdk full	654/654
sdk + session hooks + router suites	654/654
hermes-bots plugin tests	270/270
npm run check:lint (3 tsconfigs + eslint)	0 errors
tests/hermes_cli/test_web_server.py	failures pre-existing on main (env-dependent; verified identical with change stashed)

Third and final layer of the #89206 wake-failure class (gate fix #89394, routing fix #89475).

Infographic

[github-actions]

૮ >ﻌ< ა ci review

_{ran on a7bf3df — perf(desktop): durable transcript-tail cache — bot wakes pai}

⚠️ Warnings

OSV vulnerability scan · View job

7 known vulnerabilities found in pinned dependencies.

• CVE-2026-67213 in package-lock.json
• CVE-2026-67213 in website/package-lock.json
• CVE-2026-71554 in uv.lock
• CVE-2026-70608 in package-lock.json
• CVE-2026-56876 in package-lock.json
• CVE-2026-70606 in package-lock.json
• CVE-2026-71554 in uv.lock

How to fix:

Review the findings in the Security tab. Update the affected dependencies if a patched version is available.

---

debug info

CI timings

CI timings · View report · View job

Wall time 3m45s vs 3m51s (-2.6%). 15 job(s) slower, 19 faster, 4 unchanged.

• JS & TS checks / apps/desktop / check:test:ui:shard-1of3: +42.0s
• Python tests / Run tests slice 12/12: -25.0s
• Python tests / Run tests slice 3/12: -24.0s
• Python tests / Run tests slice 7/12: -19.0s
• Python tests / Run tests slice 10/12: -14.0s

[Enough1122]

AI code review — automated review for reference; please use your judgment.

• correctness: The display-only contract is handled carefully — viewMessagesForReconcile() reports empty while the provisional paint is on screen, the stranded-resume latch and session-gone checks use it, and the rollback branch drops the cache entry when authoritative sources say empty. Good.
• correctness (nit): viewMessagesForReconcile() relies on reference identity (current === cachedTailPaint) surviving the setMessages round-trip through the store. If the store ever clones/wraps arrays on set, the provisional paint silently becomes "real" view content and stale rows get grafted — the exact failure the comment warns about. Worth an assertion somewhere that identity is preserved.
• robustness (nit): loadTranscriptTail validates only that messages is a non-empty array, not element shape; a valid-JSON-but-wrong-shape entry paints garbage rows until reconcile replaces it. Low impact given the self-evict on parse failure.
• security: Transcript tails in localStorage are plaintext, but the "same trust domain as state.db on the same disk" rationale holds; gateway re-home wipe and delete-drop prevent cross-backend id reuse. Sound.
• tests: Behavior-based (round-trip, bounds, LRU eviction, sabotage-checked paint-first regression) — no snapshot/source-reading antipatterns.