fix(assistant): refine sidebar resizing and copy feedback

[chaxus]

Summary

Sidebar resizing reflows the editable draft and caps long input with internal scrolling. Layout changes follow the newest response for readers already at the end, preserving history position otherwise. Message arrivals now consume the existing reader attachment state instead of overwriting it from temporary reflow geometry, including arrivals before a scheduled layout frame.

Copy feedback uses a fixed-size icon, remains retryable after failure and announces results separately from running activity. Message spacing, lists, code and tables retain the existing theme; overflowing code and tables support keyboard scrolling. Resize observation defers dimension writes to the next animation frame. Cleanup cancels pending work and ignores late clipboard completions while preserving back/forward cache restoration. All implementation is independently authored using the current public components and theme.

Word reading tool metadata now explains that it supplies source text for factual answers even without a selection. Schemas, permissions, execution and refusal handling are unchanged. No request-specific routing is added.

Verification

• Latest full unit suite: 167 files, 4,961 tests passed. Two asynchronous Promise warnings were traced to existing converter rejection tests.
• Four reflow/content-arrival regression cases failed before the repair and passed afterwards; a fifth verifies intentional historical scrolling before the pending frame.
• Latest production-build composer UX regression: all four DOCX/XLSX/PPTX/PDF cases passed, covering resizing, editable drafts, historical position, focus, fixed-width copy feedback and keyboard scrolling.
• Current-code cross-format production acceptance: 14 cases passed (four composer surfaces, eight imported-format cases and two PDF contract cases). Legacy fixture hashes were verified. CSV was covered by the separate ten-case native write suite.
• For the tool-description change, the full 4,961-test suite, strict lint and TypeScript checks passed; all ten production native editor write cases passed with controlled providers.
• Formatting, strict lint, TypeScript and workspace chat package build passed.
• Added content checked for local paths, personal identifiers and credential patterns. No private source, research artifacts or generated screenshots are included. Browser error checks and acceptance assertions remain intact.

Remaining validation

Interaction tests use controlled responses and do not qualify the local model. Earlier exploratory Qwen3 4B trials found Portuguese failures. A follow-up development session used equal counts (three fresh conversations per language): English, Chinese, Japanese, Korean, German and Portuguese returned the exact budget in all three trials; Spanish did so once and returned operation-selection errors twice. This contradicts a Portuguese-only failure and does not qualify broader multilingual use. Subsequent provider diagnostics located two Spanish failures in operation choice. Clarifying the reading tool purpose yielded five correct source facts in fresh final-code requests, with one requested-format failure. This is a small development probe, not production or multilingual qualification. Questions, results and limitations are recorded in the regression notes. Actual screen-reader acceptance and native readback stability under competing heavy workloads also remain open. The browser cases do not independently reproduce the exact frame interleaving covered by the unit regressions.

Actual 4B development probes additionally verified Excel A1:A100 writes, factual reads, independent CSV readback and Undo/Redo; presentation text insertion/readback/Undo; and PDF factual reading, reviewed annotation, exported annotation contents/page and Undo. Two actual requests failed acceptance: a new slide with title/body became an empty slide, and a PDF body edit became a pending comment. The latter was cancelled without changing the body. These are unresolved task-planning/capability-boundary limitations, not passing model acceptance. A temporary unit timing probe also found history detachment fails when upward wheel input coincides with a viewport-height change; real-browser reproduction and repair remain open. Settings/language redesign and actual screen-reader acceptance remain open.

[cloudflare-workers-and-pages]

Deploying document with    Cloudflare Pages

Latest commit:	2404961
Status:	✅  Deploy successful!
Preview URL:	https://8f384347.document-7hm.pages.dev
Branch Preview URL:	https://codex-assistant-sidebar-deta.document-7hm.pages.dev

View logs