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

CLAUDE.md

@@ -38,7 +38,7 @@ pnpm install --frozen-lockfile   # 安装依赖
 pnpm run dev                     # 启动开发服务器（含热更新）
 pnpm run build                   # 生产构建（执行 bin/build.sh）
 pnpm run build:single            # 打包为单个 HTML 文件
-pnpm run lint:ts                 # oxlint + tsc --noEmit（CI 必跑）
+pnpm run lint:ts                 # oxlint + workspace rebuild + tsc --noEmit（CI 必跑）
 pnpm run format:check            # prettier 格式检查（CI 必跑）
 pnpm run test                    # 单元测试（Vitest）
 pnpm run test:coverage           # 带覆盖率的单元测试
@@ -204,7 +204,7 @@ test/setup/vitest.ts          # 全局 mock：matchMedia、URL.createObjectURL
 ```
 
 覆盖率统计 `lib/**/*.ts` 与 `packages/*/src/**/*.ts`，排除类型声明。
-全局最低门槛保留在 `vitest.config.ts`，历史、shared、converter、agent-core 与 chat-ui
+全局门槛为 lines/functions/statements 80%、branches 75%；Agent 应用层、历史、shared、converter、agent-core 与 chat-ui
 另有独立模块门槛。实际数字以 `pnpm run test:coverage` 为准。
 
 开发与 Vitest 通过 `bin/workspace-aliases.mjs` 从 workspace 的 `exports` 派生源码别名，
@@ -1140,3 +1140,10 @@ E2E 回归套件覆盖**（`pnpm run test:e2e` 的 embed-regression.spec：打
 ### AI assistant product entry
 
 `lib/agent-entry.ts` owns the discoverable entry and browser-local default-off preference. URL `agent` flags do not enable it. Opening restores only UI/preferences; loading a model or connecting a service requires an explicit user action. Generated writing must be reviewed before applying. Developer transport and provider details belong in developer references, not product onboarding.
+
+## Runtime and storage maintenance (2026-10-13)
+
+- `bin/offline-assets.mjs` derives mandatory app dependencies from Vite's emitted manifest. Inference chunks/WASM are optional; the small assistant UI static graph stays in core for live-page compatibility. `PREPARE_ASSISTANT` follows explicit or persisted opt-in, and a prepared marker refreshes runtime assets before later worker activation. Model weights and external endpoints remain outside app caches. Preserve the vendor precache for first offline export.
+- `pnpm run tsc` rebuilds workspace declarations before checking the app; `lint:ts` calls it. Do not rely on an earlier production build to refresh package contracts.
+- History v3 uses key-only revision/size indexing for retention; latest snapshot reads only one value. Conversation storage keeps explicit opt-in and atomic revision conflicts while updating individual sessions.
+- Panel context synchronization is lifecycle-owned: no periodic reads while closed, disabled or hidden, and BFCache restoration reinstalls observers. Keep extracted settings/context owners rather than adding them back to `panel.ts`.

bin/offline-assets.mjs

@@ -1,18 +1,67 @@
 import { readdir, readFile, writeFile } from 'node:fs/promises';
 import { join } from 'node:path';
 
-/** Build outputs only: no model weights, documents, source maps or external URLs. */
+const ASSISTANT_ENTRY = 'lib/agent-plugin/agent-plugin.ts';
+const GENERATED_ASSET = /^assets\/[A-Za-z0-9_.-]+\.(?:js|css|wasm|woff2)$/;
+
+/** Follow both static and lazy app edges, stopping at the optional assistant. */
+function collectAssets(manifest, roots, excluded = new Set(), followDynamic = true) {
+  const visited = new Set();
+  const assets = new Set();
+  const visit = (key) => {
+    if (visited.has(key) || excluded.has(key)) return;
+    visited.add(key);
+    const chunk = manifest[key];
+    if (!chunk) throw new Error(`Missing offline manifest dependency: ${key}`);
+    for (const asset of [chunk.file, ...(chunk.css ?? []), ...(chunk.assets ?? [])]) {
+      if (!GENERATED_ASSET.test(asset)) throw new Error(`Unsupported offline asset: ${asset}`);
+      assets.add(`/${asset}`);
+    }
+    for (const imported of [...(chunk.imports ?? []), ...(followDynamic ? (chunk.dynamicImports ?? []) : [])])
+      visit(imported);
+  };
+  for (const root of roots) visit(root);
+  return assets;
+}
+
+/** Core app dependencies install first; optional inference assets need explicit opt-in. */
 export async function stampOfflineAssets(directory) {
-  const chunks = (await readdir(join(directory, 'assets'))).filter((name) => /\.(?:js|css|wasm)$/.test(name));
+  const manifest = JSON.parse(await readFile(join(directory, '.vite', 'manifest.json'), 'utf8'));
+  const entries = Object.entries(manifest)
+    .filter(([, chunk]) => chunk.isEntry)
+    .map(([key]) => key);
+  const assistant = Object.keys(manifest).filter(
+    (key) => key === ASSISTANT_ENTRY || manifest[key].src === ASSISTANT_ENTRY,
+  );
+  if (!entries.length || assistant.length !== 1) throw new Error('Missing offline app or assistant manifest entry');
+  const core = collectAssets(manifest, entries, new Set(assistant));
+  // Retain the small UI graph for old live pages whose lazy import URLs must
+  // survive a deploy. Inference engines, Workers and WASM remain optional.
+  for (const asset of collectAssets(manifest, assistant, new Set(), false)) core.add(asset);
+  const optional = collectAssets(manifest, assistant);
+  // Worker outputs may be emitted outside the main manifest's dependency graph.
+  // They belong to the optional inference runtime, never to the first visit.
+  for (const name of await readdir(join(directory, 'assets'))) {
+    if (/\.(?:js|css|wasm)$/.test(name)) optional.add(`/assets/${name}`);
+  }
   const fonts = (await readdir(join(directory, 'ran-fonts'))).filter((name) => name.endsWith('.woff2'));
-  const assets = [...chunks.map((name) => `/assets/${name}`), ...fonts.map((name) => `/ran-fonts/${name}`)].sort();
+  for (const name of fonts) core.add(`/ran-fonts/${name}`);
+  const assets = [...core].sort();
+  const assistantAssets = [...optional].filter((asset) => !core.has(asset)).sort();
   const path = join(directory, 'sw.js');
   const source = await readFile(path, 'utf8');
   const marker = 'const APP_ASSETS_TO_CACHE = []; // BUILD_APP_ASSETS';
-  if (!source.includes(marker)) throw new Error('Missing offline app asset build marker');
+  const assistantMarker = 'const ASSISTANT_ASSETS_TO_CACHE = []; // BUILD_ASSISTANT_ASSETS';
+  if (!source.includes(marker) || !source.includes(assistantMarker))
+    throw new Error('Missing offline asset build marker');
   await writeFile(
     path,
-    source.replace(marker, `const APP_ASSETS_TO_CACHE = ${JSON.stringify(assets)}; // BUILD_APP_ASSETS`),
+    source
+      .replace(marker, `const APP_ASSETS_TO_CACHE = ${JSON.stringify(assets)}; // BUILD_APP_ASSETS`)
+      .replace(
+        assistantMarker,
+        `const ASSISTANT_ASSETS_TO_CACHE = ${JSON.stringify(assistantAssets)}; // BUILD_ASSISTANT_ASSETS`,
+      ),
   );
   return assets;
 }

docs/explorations/2026-10-13-follow-up-performance.md

@@ -0,0 +1,52 @@
+# Follow-up performance verification
+
+This pass starts from 0131b526 and adds one commit per feature to PR #266. Operation counts establish reduced work; they do not imply a specific browser latency improvement.
+
+## Browser baseline
+
+Chrome DevTools inspected a production local editor preview on 127.0.0.1:4203. The reload trace was unthrottled and reported CLS 0.06; no LCP or field metrics were returned. The network snapshot confirmed native app/SDK/font resources, and the accessibility snapshot confirmed the real editor was interactive. This local reload is not a cold production bandwidth benchmark. No vendor-loading optimization is claimed from it.
+
+## Selected improvements
+
+- Chat rendering: repeated full-history button-state scans and per-row scroll measurements during restoration; repeated full streamed source attribute writes.
+- History expiry: complete live metadata read/sort before expiry filtering despite an existing updatedAt index.
+- GGUF deletion: per-shard SDK deletion repeats cache enumeration, including unrelated cached files.
+
+Verification results will be recorded alongside each independently committed feature.
+
+## Chat rendering
+
+The new appendMany API restores completed messages through the existing renderer, then performs one final scroll update. Ordinary append keeps its immediate behavior. New action controls initialize within their row; existing controls update only when running or draft-emptiness changes. Streamed original source is published once at completion/interruption before the row becomes a completed answer.
+
+| Deterministic fixture                               |                           Before | After |
+| --------------------------------------------------- | -------------------------------: | ----: |
+| Disabled-control writes for 100 agent messages      |                            5,050 |   100 |
+| Disabled-control writes for 500 agent messages      |                          125,250 |   500 |
+| Scroll calls for a 500-message restore              |                              500 |     1 |
+| scrollHeight reads for a 500-message restore        |                            1,500 |     3 |
+| data-source updates across 20 streaming frames      |                               20 |     0 |
+| Final source publication on completion/interruption | Per-frame source already present |     1 |
+
+Validated with 120 focused chat/controller/storage tests, package type generation, and a mounted Chromium case restoring 400 persisted messages, switching to a short conversation and back, checking source/action state and final scroll position. Operation counts are measured in deterministic DOM fixtures; no end-user milliseconds or memory reduction is asserted.
+
+## Indexed history expiry
+
+Use the existing updatedAt index to fetch only snapshots at or before the inclusive expiry cutoff. The precise expiresAt predicate still renews retention after reopening, and a stable descending sort of expired candidates preserves returned ID ordering, including equal timestamps. Deletions and save-target cleanup retain their original transactional behavior. No database schema change.
+
+A fixture with 2,000 recent documents and five old metadata rows reads five candidates instead of all 2,005. Snapshot writes still perform the required full budget check: total metadata results fall from 4,008 to 2,008 (five expiry candidates and 2,003 budget rows); full metadata scans fall from two to one. Eleven focused history suites / 98 tests passed, including exact boundaries/ties, reopened retention, blob/handle cleanup and cancellation rollback.
+
+## Batched GGUF deletion
+
+Resolve the requested cache group, then delete its exact known filenames in one SDK deleteMany call. The installed SDK still owns body/metadata deletion and error propagation. Unrelated models, shared runtimes and the missing-group error remain unchanged.
+
+Real installed SDK/storage-boundary tests show three-shard deletion uses two inventory scans instead of four. An eight-shard, four-unrelated-file small-body adapter probe shows nine scans become two, with identical deleted and remaining filenames; Blob reads fall from 240 to 72. These are aggregate operation counts on small fixtures, not a browser heap or transfer measurement. Eleven focused model-cache/storage tests pass, including metadata cleanup and failure propagation.
+
+## Final assembled verification
+
+- Full repository-pinned unit coverage: 5,394 tests / 181 files pass; 85.75% lines, 80.12% branches, all coverage gates pass.
+- Thirty-eight impacted Chromium cases pass: long saved conversation restore, context notifications, assistant entry/native writes, history pagination/search/retention/export and autosave recovery.
+- Lint, workspace declaration rebuild, application/package type checking, Docker Compose configuration, complete format check and production Vite build pass.
+- Independent read-only review approves all three new feature scopes with no unresolved findings.
+- New commits stay separate by feature. Earlier optional-cache/storage/UI changes remain in the previously reviewed commit; no dependencies, schemas, model defaults or vendor code changed in this pass.
+
+The previous revision was separately checked with the user's pending upgraded dependencies. This follow-up was verified with repository-pinned dependencies; that distinction remains explicit in the PR. Real-user field metrics, production-network timing and deployment are outside the local evidence.

docs/superpowers/plans/2026-10-13-runtime-storage-optimizations.md

@@ -0,0 +1,91 @@
+# Runtime and storage optimization implementation plan
+
+> For agentic workers: execute the independent domains using dispatching-parallel-agents; the coordinator owns cache integration and final verification.
+
+**Goal:** Implement all audit findings and deliver one reviewed GitHub PR.
+**Architecture:** Preserve existing browser APIs and offline editor behavior. Defer optional assistant assets, reduce storage byte copies, and extract bounded UI lifecycles.
+**Tech Stack:** TypeScript, Vite, IndexedDB, Service Worker, Vitest, Playwright.
+**Spec:** docs/superpowers/specs/2026-10-13-runtime-storage-design.md
+
+## Global constraints
+
+No new dependencies or vendor changes. Preserve first offline save, stored data, explicit cloud/history opt-in, stale-tab conflicts, cancellation, public panel exports and BFCache. Do not include the root checkout's unrelated dependency edits.
+
+## Review focus
+
+- Offline reopen/export after one online editor initialization, with a cold landing that defers inference assets.
+- Enabling an already cached assistant offline after reload.
+- Database upgrades with old data and another tab open.
+- Erasure/cancellation while a save is queued or in flight.
+- Hidden panel, replaced editor and BFCache restoration.
+
+### Task 1: Optional assistant precache
+
+Files: bin/offline-assets.mjs, declaration; vite.config.ts; public/sw.js and public/sw-register.js; lib/agent-entry.ts; new lib/assistant-offline.ts; manifest/worker/entry/offline tests.
+
+- [x] Assert the core manifest retains the static assistant UI graph and excludes inference runtime assets and includes static/dynamic editor dependencies.
+- [x] Assert explicit worker preparation caches only the local allowlist, rejects failures, and preserves first native offline export.
+- [x] Enable Vite manifest; derive graph closures and stamp two asset lists; request optional preparation after explicit enablement.
+- [x] Run related unit tests and production build; measure core bytes before/after.
+
+### Task 2: Bounded document history reads
+
+Files: lib/history/{store,db,types}.ts and history unit tests.
+
+- [x] Add regressions rejecting bulk blob reads while returning newest bytes/accounting correctly.
+- [x] Implement reverse latest cursor and lightweight revision metadata/key trimming; migrate legacy records as needed.
+- [x] Verify retention, quota, cancellation and source/handle cleanup tests.
+
+### Task 3: Incremental conversation storage
+
+Files: lib/agent-plugin/ui/{sessions,conversation-db}.ts and their unit tests.
+
+- [x] Test burst coalescing, unchanged-session writes, v1 migration, stale-tab and erase races.
+- [x] Implement a bounded pending save and per-session persistence under atomic index revision, keeping repository API.
+- [x] Run the conversation/history controls/controller test suites.
+
+### Task 4: Panel lifecycle and decomposition
+
+Files: lib/agent-plugin/ui/panel.ts; extracted context/settings/history modules; related unit tests.
+
+- [x] Test no hidden/disabled/background polling, immediate resume, SDK callbacks and BFCache cleanup.
+- [x] Extract context scheduler and at least two focused presentation responsibilities; preserve public panel functions and request semantics.
+- [x] Run all panel/composer/context/sidebar tests.
+
+### Task 5: Quality gates and review
+
+Files: failing iframe/focus tests; package.json scripts; vitest.config.ts; shared/root CLAUDE guides; workflow contract tests.
+
+- [x] Fix synthetic iframe/focus fixtures without weakening production behavior assertions.
+- [x] Make type checking rebuild declarations, verify from deleted stale package output.
+- [x] Tighten useful application-owned coverage gates and update outdated guide statements.
+- [x] Run full unit coverage, lint, format, production build and impacted E2E.
+- [ ] Fresh review of the complete diff; resolve findings; commit, push and create PR.
+
+## Progress
+
+- Scope authorized by the user's instruction to implement all audit recommendations and submit a PR.
+- Phase 1 branch prepared; PR phases 2-7 pending until implementation and validation complete.
+
+## Implementation evidence
+
+- History: 11 suites / 112 tests passed, including legacy v1/v2 upgrade, bounded reads, exact gapped-revision accounting, blocked upgrade retry and cancelled trim rollback.
+- Conversation: 81 targeted tests passed; v2 migration, stale-tab conflicts/erasure and coalescing verified. Full-snapshot validation/serialization remains proportional to input size; unchanged records are neither cloned nor rewritten.
+- Panel: 176 targeted tests passed; three modules extracted, panel reduced from 2905 to 2633 lines. SDK bursts coalesce and hidden/disabled/BFCache lifecycles release resources.
+- Integrated upgraded-toolchain run: 5360 unit tests passed, coverage gates passed (85.39% lines, 79.84% branches), 34 impacted Chromium E2E passed.
+- Default app precache: 1,453,207 bytes (~1.39 MiB), down from 57,481,875 bytes (~54.8 MiB); optional inference runtime: 56,063,311 bytes (~53.47 MiB). These are decoded/build file bytes, not network transfer measurements.
+- Fresh converter declaration output was moved away before running the type-check script; the script rebuilt packages and passed.
+- Independent review found and fixed prepared-runtime update migration/races. Final scoped review reports no unresolved Critical/Important findings.
+- The first browser probe assumed a landing-only visit fully warms the editor. Missing native vendor assets disproved that assumption; the acceptance flow now initializes the native editor online before offline reopen/export, preserving the existing contract.
+
+Final pinned-toolchain verification: 179 unit files / 5360 tests passed with coverage gates; 34 impacted browser cases plus the isolated worker-update case passed (35 total). Lint/type/build/format and fresh-declaration checks passed. The same pinned update case also passed against unmodified main. Upgraded Playwright update timing remains inconsistent in local probes; dependency updates are outside this PR.
+
+## Follow-up: event-driven assistant UI
+
+- Replace panel context intervals with disposable SDK/native-input subscriptions and explicit readiness/open/visibility notifications.
+- Replace sidebar rail scans with filtered DOM/iframe-load/media-query observers.
+- Replace pending-preview validity intervals with the context event lifecycle; preserve pre-apply validation.
+- Refresh model release controls and translated context explicitly from existing runtime/language updates.
+- Verify focused no-polling regressions, full quality checks, real editor sidebar/selection interactions, review and update PR #266.
+
+Follow-up verification: 5378 tests across 179 files pass with coverage gates (85.51% lines, 79.90% branches); lint/type checks, format and production build pass. Twenty impacted assistant Chromium cases pass, including three new event-driven cases; the final rebuilt three-case suite additionally verifies a native sheet rename and same-selection history mutation. Independent read-only review found missing spreadsheet mutation/sheet notifications, now covered and rechecked. The tracking fixture models SDK multiple-listener ownership and checks complete pagehide cleanup. No recurring timers remain under lib/agent-plugin/ui.