perf(editor): optimize caching, history and conversation rendering

[chaxus]

Merge Verdict

[APPROVE] — The original runtime/storage optimizations and three additional performance features are implemented and independently reviewed, with no unresolved blocking findings.

46 files · +3663 / -485 · offline caching, local history, conversations, assistant UI, model cache and quality gates

Summary

• Reduce mandatory app precache from 54.8 MiB to 1.39 MiB through Vite's core/static UI versus optional inference dependency graphs. Preserve explicit/persisted opt-in and optional assets across worker updates, including legacy users and preparation races. Figures are build-file bytes, not network-transfer measurements; native vendor precache remains unchanged.
• Bound latest snapshot reads and revision trimming, coalesce conversation saves and persist changed sessions atomically, and remove assistant context/sidebar/preview polling using disposable SDK/input/DOM/media notifications.
• Add three independently committed performance features: linear action initialization and batched chat restoration, indexed history-expiry candidates, and exact-group GGUF batch deletion. Preserve interaction, retention and deletion semantics.
• Repair jsdom fixtures, rebuild workspace declarations before type checks, tighten coverage gates and update package/project guidance.

Risk Analysis

Risk	Level	Mitigation
Stored-data upgrades and retention	Medium	Transactional legacy migrations remain covered. Indexed expiry keeps reopening renewal, inclusive cutoff, stable timestamp ties and cancellation rollback; byte/handle accounting tests pass.
Offline behavior across deploys	Medium	Retain native first-export precache and static UI dependencies. Prepared markers, legacy handoff, install/activation rechecks and controller-change preparation cover optional runtime updates.
Chat controls and scrolling	Low	Existing renderer and normal append behavior retained. Tests cover action outcomes, draft transitions, grouping, partial answers and reader scroll intent; Chromium restores 400 saved messages and switches sessions.
Model-cache deletion	Low	Delegate exact filenames to installed SDK deleteMany; real SDK tests verify body/metadata cleanup, unrelated models, missing groups and error propagation.

Design Decisions

• Keep the small static assistant UI graph in mandatory core so old live pages retain lazy UI imports after deployment. Large inference assets remain optional; default-off visits do not prepare models or connect services.
• History adds a compound revision/byte-length index in the original optimization. Latest retrieval stops after one value, trimming reads keys, and this follow-up uses the existing updatedAt index for expiry without another schema change.
• Preserve complete conversation input validation and full-snapshot repository APIs. Identity-only serialization caching would miss callers that mutate loaded messages in place. Unchanged sessions skip IndexedDB writes.
• Assistant UI notifications coalesce into one queued refresh. No recurring context/sidebar/preview timers remain; stale targets are still checked before applying. Autosave timing and bounded native initialization/font waits retain their behavior.
• Chat history uses an additive appendMany API rather than virtualization. Streaming continues to append text per frame and parse Markdown on completion; original source metadata is finalized once at completion/interruption.
• Follow the user's feature-per-commit request: retain the original reviewed commit and three separate new commits; do not squash this PR. No dependency/lock versions, vendor code, model defaults or external services change in this follow-up.

Follow-up commits and measured work reduction

Commit	Feature	Deterministic comparison
79034c66	Chat rendering and conversation restoration	For 500 messages, action-state writes 125,250→500, scroll calls 500→1, scrollHeight reads 1,500→3. Twenty streaming frames write no full source attributes; completion publishes once.
60bad002	Indexed history expiry	2,005 metadata rows become 5 candidates. Normal snapshot writes retain one budget scan instead of two full metadata scans.
c891d820	GGUF group deletion	Eight-shard adapter/SDK probe: 9→2 inventory scans, identical deleted/remaining filenames. Regression tests exercise actual installed SDK methods.

These are operation counts on deterministic fixtures and local SDK/storage probes, not claimed browser latency, heap or transfer improvements. Details: docs/explorations/2026-10-13-follow-up-performance.md.

Code Concerns

No unresolved findings from independent reviews. Offline-update races and missing spreadsheet content/sheet notifications were fixed and rechecked. The follow-up review traced real callers, installed SDK deletion semantics, expiry ordering and chat action/stream ownership without finding additional correctness issues.

Verification

Check	Result
Final full unit coverage, repository-pinned toolchain	5394 passed, 181 files; lines 85.75%, branches 80.12%; all tightened gates pass.
Final follow-up Chromium E2E, repository-pinned toolchain	38 passed: 400-message restoration and session switching, context notifications, assistant entry/native writes, history search/pagination/expiry/export, autosave recovery.
Earlier offline/storage Chromium E2E, pinned toolchain	35 passed, including the separately run silent SW update case.
Earlier optimization revision, pending upgraded toolchain	5360 unit tests passed, coverage gates and 34 affected E2E cases passed. The isolated SW update case times out on both branch and unmodified main with Playwright 1.64; pinned Playwright 1.62.1 passes on both. These dependency upgrades remain outside the PR; the newest follow-up was tested with pinned dependencies.
Lint, workspace declaration rebuild and TypeScript	Passed, including package and application declarations.
Formatting and diff whitespace	Passed.
Production build	Passed; current mandatory app precache is 1,456,339 bytes / 1.39 MiB.
Docker Compose configuration	Passed.
Independent code review	All new feature scopes approved; no unresolved Critical/Important findings.
Git boundaries	Clean worktree; current origin/main is the base; original commit plus three separate feature commits.

Conclusion basis

• Verified: production builds and manifests, full unit/coverage suites, actual installed SDK/cache behavior, deterministic operation counts and affected real-editor Chromium flows.
• Observed locally: unthrottled editor reload trace reports CLS 0.06; no LCP or field data was returned. This is not a cold production benchmark and prompted no vendor rewrite.
• Not covered locally: the entire multi-host/browser CI matrix, real-user timing or production deployment. No merge or deployment was performed.

[cloudflare-workers-and-pages]

Deploying document with    Cloudflare Pages

Latest commit:	c891d82
Status:	✅  Deploy successful!
Preview URL:	https://760c47e8.document-7hm.pages.dev
Branch Preview URL:	https://perf-runtime-and-storage.document-7hm.pages.dev

View logs