feat(assistant): refine reviewed edits and add PDF assistance

[chaxus]

Merge Verdict

[NEEDS_CHANGE] Submitted for review. Editor execution, review and grounded read-answer continuation are validated. Repeated multilingual model qualification and real screen-reader acceptance remain open.

Summary

The assistant previously displayed separate, bulky approval panels and lost the meaning of pending edits during follow-up requests. Review now stays with the relevant conversation turn, shows text or cell changes, supports refining an unexecuted proposal, and collapses completed operations. Failed, unchanged or stopped refinements preserve the previous suggestion.

Content questions now continue from native read receipts to a tool-free answer using the original question and captured scope. The host discards late answers after Stop or a target change and uses the existing reasoning presentation filter. Empty or unavailable sources remain explicit; navigation and mixed read/write sequences retain their existing flow.

During a response, the next draft stays editable without sending or queuing it. Actual reading and response phases use a polite status, pending review remains inline, and completed changes stay compact. New proposals preserve historical scroll position, and response completion does not steal focus from another control. Seven-language product guides describe the behavior.

PDFs now expose selectable text on the current loaded page and reviewed page notes. Native writes validate the captured target, verify actual editor results, and support rollback and Undo. Imported document regressions cover ODT, ODS, ODP, RTF, TXT and public legacy DOC, XLS and PPT files alongside native Word, spreadsheet, presentation and PDF cases.

Browser GPU presets now start at Qwen3 4B. Retired small presets cannot be selected or reused, saved choices migrate to 4B, and their cached files remain removable. Custom sources and the separate CPU path retain their existing behavior. Product guides, translated help, README files, changelog and sitemap dates are updated.

Risk Analysis

• High — model interpretation: 4B is still experimental. Changing the default does not establish production readiness. Native execution tests use controlled responses; actual 4B trials are partial and separate.
• Medium — stale or destructive edits: captured editor/page/range identity, review for clearing or overwriting, native readback, cancellation and rollback prevent applying an obsolete suggestion. Existing document Undo remains available.
• Medium — PDF scope: only selectable text on the loaded page is read; scanned text recognition and PDF body rewriting are unavailable. Page notes require review and are reversible.

Design Decisions

• Keep pending review inline and settled records compact, with explicit action labels and keyboard-accessible controls.
• Preserve validated operation parameters during refinement rather than interpreting conversational prose as successful execution.
• Use native editor tools and verified receipts across document types. Do not treat model claims as execution evidence.
• Keep public legacy fixtures outside Git; fetch a pinned upstream revision and verify signature, size and SHA-256.

Code Concerns

• [Needs Decision] Repeated seven-language model qualification and real screen-reader acceptance are still open. Do not infer these from unit or editor contract tests.

Verification

Conclusion basis: Verified execution contracts with automated tests; partial real model trials are documented separately and do not qualify the default.

• Latest full unit suite: 166 files, 4,949 tests passed.
• Final proposal/review regression rerun: 3 files, 13 tests passed; independent core review: 4 files, 160 tests passed.
• Native editor and PDF browser regression: 12 tests passed, including the production build.
• Imported-format browser regression: 8 tests passed with the pinned public legacy corpus enabled; no skipped corpus cases.
• Latest combined browser regression: 24 tests passed, including pending drafts and focus in DOCX/XLSX/PPTX/PDF, native writes and Undo, scoped PDF behavior, and eight imported formats with all public legacy cases enabled.
• Composer, panel and inline-review unit regression: 4 files, 177 tests passed.
• Initial combined checks had a bounded-file unit timeout and two imported-file readback failures. Isolated rechecks and the final full unit and sequential browser runs passed without relaxing assertions or timeouts; ordinary sequential requests now await the true running state. A dedicated three-worker repeat passed all six TXT/ODP cases. The same repeat alongside the heavy type check had three failures; competing-workload stability remains an open observation, not a claimed repair.
• Formatting, strict lint and TypeScript checks passed.
• Reverse verification reproduced the late-answer target bug and visible reasoning tags; both regressions pass after repair.
• One final-code actual Qwen3 4B browser trial answered the synthetic document budget question with exactly 120, with no visible reasoning tags. This is a single trial, not model qualification.
• All hosted checks passed for the preceding commit e21ce40. CI for this interaction follow-up must finish separately.
• Reverse verification for default, preset, saved-model migration and retired cache behavior reproduced failures before the repair; native rollback and unchanged-refinement evidence is recorded in the regression documents.
• Three read-only review partitions checked orchestration, editor adapters, UI and documentation. Fixed the discovered README placement issue and strict lint failure before submission.
• Added content was checked for local filesystem paths, personal identifiers, credential patterns, oversized files and generated artifacts. No private corpus or local evaluation outputs are included.

[cloudflare-workers-and-pages]

Deploying document with    Cloudflare Pages

Latest commit:	e60bb13
Status:	✅  Deploy successful!
Preview URL:	https://83752fe6.document-7hm.pages.dev
Branch Preview URL:	https://codex-assistant-contextual-r.document-7hm.pages.dev

View logs