feat(assistant): improve document workflows and request feedback

[chaxus]

Merge Verdict

[APPROVE]

35 files · +2367 / -90

The assistant now offers tasks for the current document, distinguishes reusable drafts from ordinary or interrupted answers, and makes review and waiting states easier to understand. Commits remain separated by function as requested.

Summary

• Stage editable tasks for Word, Excel, PowerPoint and PDF using the actual document or selection. An unchanged, current staged task skips the separate intent-classification call; explicit Send and Apply remain required.
• Clarify document-draft and incomplete-response identity, readable Markdown, compact controls and keyboard focus.
• Emphasize Apply/Cancel during review; keep Edit, Refine, Copy and Locate in More actions. Show truthful completion and native Undo guidance.
• Show actual request-understanding and generation phases immediately with localized elapsed waiting. Stop, first content, review, clear and disposal release the timer while preserving the next draft.

Risk Analysis

Risk	Level	Mitigation
A staged task outlives its selection or session	Medium	Bind text, target and session; validate again at submission; preserve the draft on expiration.
Presentation changes alter copied or applied text	Medium	Host metadata owns draft identity; source text and existing mutation guards remain authoritative; original-copy and native Undo regressions pass.
Waiting feedback survives a completed or stopped turn	Low	Monotonic clock and explicit lifecycle cleanup, including renderer-only clear and subsequent retry; live regions announce phases rather than elapsed ticks.

Design Decisions

• Reuse ranui components and tokens, existing bounded context capture, ChatView and ActionPreview.
• Keep provider defaults, persistence formats and editor mutation contracts unchanged.
• Add optional public activity/label fields and complete all seven host locale tables.
• Improve time to visible feedback without simulated reasoning, fake percentages or a claim of faster model computation.

Code Concerns

No outstanding actionable findings from specification and code-quality review. Verified findings were corrected before final integration.

Verification

Conclusion basis: Verified

• pnpm exec vitest run --maxWorkers=2: 5,448 tests passed in 187 files.
• pnpm run lint:ts, repository format check and production pnpm run build: passed. Final documentation formatting and git diff --check also passed.
• Combined Chromium run with one worker: 40 real-editor tests passed, covering task scope, classification bypass, review/application/Undo, Stop/resume, history, dictation, composer layout, context events, result identity and process feedback.
• Inspected light/dark, narrow/default/wide and Chinese narrow screenshots; checked shared alignment, compact controls, stable composer geometry, keyboard focus and reduced motion.
• Earlier screenshot writes exhausted disk space; after scoped generated-output cleanup, the four affected reruns and the final complete 40-case run passed.

Conclusion basis: Inferred from code, not measured on the user's hardware
Ordinary browser-local questions still classify intent before generating an answer, and context capture can attach up to 8,000 document characters. Reducing repeated inference and selecting context are separate follow-up work.

Not covered
Physical microphone behavior, manual screen-reader/high-zoom acceptance, actual WebGPU/CPU first-token latency and model answer quality. Provider responses and speech events are controlled in browser tests.

See experience review and acceptance.

[cloudflare-workers-and-pages]

Deploying document with    Cloudflare Pages

Latest commit:	2b3689e
Status:	✅  Deploy successful!
Preview URL:	https://f8f41bfc.document-7hm.pages.dev
Branch Preview URL:	https://feat-assistant-experience.document-7hm.pages.dev

View logs