feat(assistant): improve conversation continuity and responsiveness

[chaxus]

Merge Verdict

[APPROVE]

Final source review, local regression, deployed assistant acceptance and all reported CI checks passed at exact head bdda0d0a274f7c5dae50d8547874bd7a612c4595. Functional commits remain separate; this PR is not merged.

Summary

• Keep drafting and dictating while an answer runs. Pending questions can be edited, removed or sent immediately; independent Stop preserves the partial answer and next draft, and later questions remain usable. Title navigation finishes dictation into its original chat before restoring another draft.
• Replace the history selector with searchable, grouped pages. Restore the latest 40 messages first, load earlier messages without replacing current nodes or shifting the reading anchor, and preserve the active reply when another chat is deleted.
• Render streamed Markdown progressively with stable completed blocks and bounded formatting work. Native scrolling follows new output until the reader interrupts it; selection disclosure, compact controls and responsive layouts remain aligned.
• Preserve request ownership during delayed history startup. Ordinary conversation survives irrelevant document edits, while document operations retain their original-target freshness and review requirements.
• Avoid unnecessary greeting classification and document scans, move history search off the main thread with a cooperative fallback, and bound repeated exact CPU context counts. Record actual content-free local request stages without inventing model progress.

Risk Analysis

Cross-cutting lifecycle changes retain per-request identity, verified document outcomes, original-document scope checks and physical inference/Apply settlement. Drafts and pending questions remain volatile, and disk history stays opt-in. Model defaults, dependencies, persistence contracts and document review permissions are unchanged.

Design Decisions

The implementation borrows stable incremental presentation and explicit scroll ownership from Streamdown and DeepSeek Harness while retaining the existing TypeScript, safe Markdown renderer and Worker architecture. Oversized automatically formatted content uses exact raw text, with explicit expansion for large restored messages. Long model context trims older complete turns using an exactly verified budget; semantic summarization is not added.

Code Concerns

The integration review and independent focused dictation-repair review found no blocking issues within this PR. Three recognition phases and all four editors cover the title-navigation draft ownership regression. The focused repair, tests and acceptance record are one separate commit, bdda0d0.

Verification

Conclusion basis: Verified

• Exact head bdda0d0: all 5,858 unit tests in 200 files passed with the existing coverage thresholds; lint, workspace/application types, formatting and production build passed. Both archive-aware checks verified all 1,511 immutable evaluation artifacts.
• Thirteen serialized real-editor regressions passed: first microphone permission, dictation in Word/Excel/PowerPoint/PDF with title navigation and late callbacks, history hydration/deletion, and Stop followed by new questions.
• All five dictation cases also passed on the immutable Cloudflare deployment for bdda0d0, including each editor's original and destination draft isolation.
• All reported checks passed at exact head bdda0d0: regular E2E, Docker-image E2E, Cloudflare-semantics E2E, timing budgets, lint/validation, Cloudflare deployment, real-preview smoke and PR title validation. No merge has been performed.

Chromium probes covered delayed history hydration, request preparation, conversation-only replies during editor changes, dictating a queued next draft, stable streaming nodes, Stop with a 108KB partial answer, native scroll interruption and bounded history restoration. English/Chinese layouts were inspected at three widths in light/dark themes, including keyboard and reduced motion. Actual presentation observers were verified against visibility, Stop and replacement ownership; timing records contain numeric fields and status only.

Limits: Controlled speech and transport fixtures do not certify the user's microphone or physical local-GPU first-token/drain time. DOM presentation timing is not compositor paint timing.

Acceptance record · Reference review

Exact-head CI · Immutable preview

[cloudflare-workers-and-pages]

Deploying document with    Cloudflare Pages

Latest commit:	bdda0d0
Status:	✅  Deploy successful!
Preview URL:	https://7d20fb57.document-7hm.pages.dev
Branch Preview URL:	https://fix-speech-first-permission.document-7hm.pages.dev

View logs