fix(desktop): keep long tool sessions responsive (rebased #68503)

[iwillwill-ALLWILL]

Summary

Rebased and hardened continuation of #68503 by @frizikk, whose branch became non-mergeable after #69019 changed the same thread-list virtualization code.

This keeps Hermes Desktop responsive when opening long, tool-heavy sessions without deleting, truncating, or mutating transcript data.

The first three commits preserve Jakub Wolniewicz's original authorship. The final commit contains review-driven hardening and regression tests from @iwillwill-ALLWILL.

Supersedes #68503.

Fixes #68467.

Root cause and approach

A representative transcript (266 messages, 143 tool calls, 650,451 content characters) could synchronously mount and highlight enough history to block the renderer for seconds even while the backend was idle.

The change:

• mounts at most 20 complete semantic history groups initially and progressively reveals older groups;
• enforces one turn-wide 20-tool budget across interleaved narration and assistant messages;
• preserves oldest-visible identities, scroll position, and keyboard focus while revealing history;
• discovers preview artifacts from authoritative transcript state without mounting hidden heavy tool rows;
• renders complete plain/copyable code immediately, then viewport-gates Shiki enhancement in a shared lazy Web Worker;
• bounds worker retry, cancellation, late-reply handling, generation retirement, and final lease cleanup;
• preserves fix(desktop): stop long-session transcript from drifting to old turns #69019's live-tail content-visibility behavior while resolving the original PR's conflicts.

Additional hardening after independent review

A second review found two correctness bugs in the original candidate. Both now have fail-before/fix-after regression coverage:

1. A lifetime preview-target Set could suppress a same-path artifact after restore/edit cleared the preview store and the regenerated timeline produced that target again. Registration now tracks bounded per-session tool-occurrence keys plus a clear generation, and batches all newly observed targets into one idempotent atom update. This preserves dismiss semantics and prevents historical-target replay churn.
2. Shiki tokens from old code/language props could survive into one committed frame before a passive effect cleared them. Highlight state now carries its source identity and is synchronously ignored unless it matches current props.

The hardening commit also keeps ref mutation out of React state updater callbacks and adds coverage for:

• session-key changes while history is expanded;
• tool-window boundaries at 0, 1, 19, 20, 21, 40, and 41 calls;
• stable oldest-visible tool identity after append;
• removed saved pagination keys;
• restore-and-reproduce preview targets;
• one-write reconciliation for 101 historical targets and sequential eviction semantics;
• dismissed previews remaining dismissed when unrelated targets append;
• stale highlighter props at commit time.

Production Electron A/B

Fresh minified production renderers were compared on the same macOS arm64 machine, with the same 266-turn synthetic mixed-Markdown transcript and three runs per side.

Metric	current main	this branch	delta
transcript mount	144.0 ms	122.7 ms	-14.8%
total renderer long-task time	251 ms	72 ms	-71.3%
max renderer long task	176 ms	72 ms	-59.1%
keystroke p50	9.5 ms	2.7 ms	-71.6%
keystroke p95	16.6 ms	16.2 ms	-2.4%
keystroke p99	17.5 ms	17.5 ms	0.0%
keystrokes over 16 ms	12	7	-41.7%

The final exact-HEAD run also passed the repository's committed performance gate. The large keystroke improvement is treated as host variance rather than claimed as a feature gain; the merge-safety signal is that no same-machine metric regressed materially, while transcript long-task time remained substantially lower.

A separate production Electron probe loaded a real fenced TypeScript block and observed 25 styled token spans after deferred highlighting, proving that the emitted worker asset loaded and returned tokens under the production renderer/CSP path.

Validation

Run after rebasing onto current origin/main:

npm --workspace apps/desktop run typecheck
npm --workspace apps/desktop run lint
npm --workspace apps/desktop test
npm --workspace apps/desktop run build
git diff --check origin/main...HEAD

Results:

• typecheck: passed;
• ESLint: 0 errors (9 pre-existing warnings outside this diff);
• Full Desktop suite: 284 files (283 passed, 1 skipped); 2,550 tests passed, 3 skipped;
• focused changed suites: 63/63 passed;
• production renderer/Electron build: passed, including the emitted lazy Shiki worker;
• git diff --check: passed;
• current-main RED proof: all 7 representative long-session assertions fail without the implementation;
• every review-blocker regression test was observed failing before its fix and passing afterward.

Scope and attribution

• 20 files, all under apps/desktop/src;
• no dependencies, lockfiles, generated output, logs, credentials, fixtures containing private data, or backend/Python changes;
• original three commits remain authored by Jakub Wolniewicz (@frizikk);
• the replacement exists because fix(desktop): keep long tool sessions responsive #68503 is DIRTY against current main; it is not presented as independently authored work.

Type of change

• Bug fix
• Tests
• New feature
• Security fix
• Documentation/config change

Checklist

• Read the contributing guide
• Conventional commit messages
• Searched existing issues/PRs and explicitly linked the superseded PR
• Desktop-only scope with no unrelated changes
• Regression tests added
• Typecheck, lint, complete Desktop test suite, production build, and diff check passed
• Tested on macOS arm64 in a production Electron renderer
• Considered Worker/IntersectionObserver fallback behavior for cross-platform compatibility

[teknium1]

Thanks for the substantial investigation, realistic regression coverage, and preservation of the earlier contributors' attribution.

Automated hermes-sweeper review found that the renderer-responsiveness protection is already implemented on current main:

• apps/desktop/src/components/assistant-ui/thread/list.tsx:45-57 bounds the synchronous initial paint to 20 parts; :241-280 restores retained history through an interruptible transition and keeps the newest weighted groups mounted first.
• apps/desktop/src/components/assistant-ui/thread/list.tsx:425-445 applies content-visibility:auto to older visible turns while preserving the live tail.
• The linked issue Desktop: long tool-heavy sessions freeze composer (typing/send dead) while backend is idle #68467 was closed as implemented on main, and the maintainer-associated review on fix(desktop): keep long tool sessions responsive #68503 likewise concluded that its alternate pagination/worker implementation was no longer needed for the reported root failure.
• Main subsequently retained unconditional retained-history backfill in 7f87b672455c, because suppressing it during streaming broke the real large-session resume invariant.

Closing as implemented on main. This is an automated hermes-sweeper review.