feat(assistant): improve document writing and inline task feedback

[chaxus]

Merge Verdict

[APPROVE] — Editor-scoped tasks now wait for real execution feedback, retain compact inline review, and preserve formatted document content. No unresolved actionable findings remain after independent review and repairs.

66 files · +4,051 / -252 · assistant host, editor adapters, agent-core, chat-ui, translations, regression tests and product help

Summary

• Configured assistants reopen in chat; cached model preparation is explicit and preserves drafts. Speech languages open in one ranui select, settings share disclosure styling, and waiting/review states retain keyboard and history behavior.
• Native and schema providers share the existing guarded tool loop. Confirmed, cancelled, expired and failed reviews return actual outcomes; compound tasks continue from refreshed frozen targets. PDF body changes cannot silently become annotations, and repeated writes receive honest partial-result feedback.
• Word drafts preserve safe headings, lists and emphasis; long presentation drafts fit measured native shapes or continue across pages with one Undo. Read receipts identify actual source scope; spreadsheet reviews distinguish formulas from calculated values. Seven product guides explain these capabilities and limits.

Risk Analysis

Risk	Level	Mitigation
Native content changes	Medium	Frozen editor/worksheet ownership, explicit review, native readback and owned rollback/Undo; export/reopen and Unicode pagination verified.
Small-model judgment	Medium	Qwen3 4B stays experimental; validated arguments, effect boundaries and repeat/budget guards remain. Host tests do not certify model task selection.
Limited readable scope	Low	Selection/body/ranges/slides/current PDF page are reported separately; unavailable and clipped results are explicit.
Tool-result context size	Low	Original calls retain arguments; unchanged mutation results avoid duplication, while user-revised arguments remain explicit. The context budget is unchanged.

Design Decisions

• Keep the current ranui controls, theme tokens, inline review and native history; do not add an Agent framework or floating approval cards.
• Share task execution and read receipts across provider transports. Editor capabilities and frozen targets remain in the host; generic agent-core receives typed stop reasons and a scope callback.
• Default Word insertion remains literal. Explicit formatting serializes a bounded safe subset; formatted tables and structured insertion under Track Changes are refused before writing.
• Measure slides with the bundled native layout engine, preserve every character and append at most 20 fitted pages as one owned operation.
• Exclude only generated isolated browser builds from TypeScript discovery, preserving all source checks and avoiding duplicate bundled editor declarations.

Code Concerns

No unresolved actionable findings. Independent review identified input-specific reads ending compound tasks, readonly navigation being invalidated after success, and formatted whitespace/table verification gaps; all were repaired and rechecked. Native long-slide verification also exposed duplicated result arguments exhausting the existing context budget; a failing regression now covers compact results and retained user revisions.

Verification

Check	Result
Shared, agent-core and chat-ui builds	PASS
Root TypeScript	PASS; zero generated isolated-build declarations included
Strict repository lint	PASS
Repository formatting	PASS
Full unit suite	PASS: 5,039 tests across 171 files
Coverage thresholds	PASS: 88.15% statements, 83.69% branches, 90.54% functions, 91.41% lines; unchanged thresholds
Focused checks after type declaration repair	PASS: 20 tests
Fresh serial production assistant regression	PASS: 38 cases in the combined run plus the corrected formula-label case in a fresh production rerun; 39 cases total, no skipped legacy fixtures
Reverse-disable verification	PASS: native Word formatting test fails when formatted insertion is disabled; restored implementation passes
Privacy and diff hygiene	PASS: added/untracked text contains no personal paths, credentials, private research identifiers or debug probes
Actual browser-model acceptance	Partial evidence only: earlier factual PDF reading and slide creation worked, but 4B repeated a completed write; long formatted writing is not independently qualified
Real microphone recognition and screen-reader use	Not covered; keyboard, focus and accessibility snapshots are covered

Conclusion basis

• Verified: workspace package builds; root tsc --noEmit --listFiles; strict lint and formatting; full Vitest suite and coverage; the seven changed assistant Playwright suites against a fresh production build, including public legacy fixtures; native text/shape/history checks, DOCX export/reopen and reverse-disable failure.
• Inferred from code, not tested: arbitrary external provider quality and unsupported document content outside the exercised capabilities.
• Not covered: all site-wide browser suites, container/Pages/deployed-site CI environments, real microphone accuracy in all seven languages, actual screen-reader sessions, or a production qualification of the browser model.

[cloudflare-workers-and-pages]

Deploying document with    Cloudflare Pages

Latest commit:	5851754
Status:	✅  Deploy successful!
Preview URL:	https://779f95f0.document-7hm.pages.dev
Branch Preview URL:	https://codex-assistant-content-revi.document-7hm.pages.dev

View logs