Additional

Release Notes 0.5.41

Synced from github.com/CoWork-OS/CoWork-OS/docs

This page summarizes the product and engineering changes included in 0.5.41, following v0.5.40.

Summary

Release 0.5.41 is the Everything Workbench release. It turns generated documents, spreadsheets, presentations, web pages, and live browser sessions into first-class in-app work surfaces with compact task cards, sidebar/fullscreen viewers, follow-up composer context, and safer save/preview paths. It also adds structured memory observations, Mission Control intelligence projections, format-aware file previews, visible browser automation, developer log capture, renderer bundle reductions, and broader test coverage across the new artifact surfaces.

New Features

  • Everything Workbench: generated documents, spreadsheets, decks, web pages, PDFs, and previews now share a unified artifact model: task-feed card, in-app open action, sidebar workspace, fullscreen mode, follow-up composer, and refresh after the agent finishes edits. Learn more
  • Document artifacts: Word-style outputs now render as compact artifact cards. DOCX opens in an editable document surface with heading/text controls, bold/italic/underline, list controls, save/copy, external app actions, fullscreen mode, model picker, voice input, attachments, and task-follow-up context. DOC, RTF, ODT, OTT, Pages, DOCM, DOTX, and DOTM get preview or native-app handling. Learn more
  • Spreadsheet artifacts: Excel workbooks and CSV/TSV files open in a spreadsheet workbench with sheet tabs, cell/range/row/column selection, formula bar, copy, zoom, add row/column, save, fullscreen mode, model picker, voice input, attachments, and follow-up context. Numbers, Google Sheets shortcuts, ODS, and XLSB are recognized with external-app/folder actions. Learn more
  • Presentation artifacts: PPTX outputs open in a resizable presentation viewer with thumbnails, slide navigation, zoom, speaker notes, text-first loading, cached rendered slide images, external actions, fullscreen mode, and follow-up context. Legacy PowerPoint formats are recognized with external-app/folder actions. Learn more
  • Web page artifacts: generated .html / .htm files and built React output entrypoints open in a sandboxed iframe workbench with browser/folder/copy actions and follow-up context. React-style source projects without build output now show a build-output-needed state instead of auto-starting a dev server. Learn more
  • Visible browser workbench: interactive browser-use tasks now open a live Electron webview in the task sidebar by default, with URL controls, persistent workspace browser partitions, fullscreen mode, visible cursor movement/action pulses, screenshot capture, screenshot annotation, and follow-up handoff from the same workbench. Browser tools can still fall back to forced headless Playwright, external Chrome profiles, or DevTools attach when explicitly requested. Learn more
  • Structured memory observations: local memory now has inspectable observation metadata, progressive index/timeline/detail recall tools, Memory Hub Inspector actions, deterministic rebuild/backfill behavior, workspace-scoped mutation boundaries, privacy controls, soft-delete suppression, and Supermemory mirroring exclusions. Learn more
  • Mission Control intelligence: Mission Control now has repository-backed projected items, evidence records, task/mention/activity/heartbeat refresh flows, intelligence UI surfaces, and stale terminal-task cleanup.
  • Format-aware file preview popup: clicked file links now open a format-aware preview modal with tailored sizing, metadata subtitles, and unified Copy path / Show in Finder / Open externally / Close actions. New branches include JSON tree/raw preview, CSV/TSV table rendering, audio playback, richer image metadata, and syntax highlighting for code and LaTeX.
  • Structured dev logging: npm run dev:log writes redacted text logs, JSONL logs, latest-run mirrors, retained-run manifests, and configurable cleanup controls for local debugging.

Enhancements

  • Cleaner task artifact display: document, spreadsheet, presentation, and web artifact cards now render only at the final output reference for a task instead of repeating throughout intermediate file and artifact events.
  • DOCX save fidelity: DOCX editing preserves unsupported package parts, updates parsed blocks in place, keeps inserted blocks in editor order, and avoids collapsing existing formatted runs during save.
  • Large CSV/TSV safety: delimited spreadsheet previews preserve rows outside the preview window, append newly added rows after hidden source rows, and prevent editing columns beyond the saveable preview cap.
  • PPTX generation and preview fidelity: presentation generation and preview extraction were expanded with better layouts, richer text handling, thumbnail/render caching, and test coverage.
  • HTML asset inlining: web previews now inline local relative and root-relative build assets so Vite/React-style dist/index.html outputs render in the sandboxed iframe.
  • Browser tool routing: browser_navigate, click/fill/type/press/scroll/wait/select/content/text/evaluate/screenshot, and related browser tools now prefer the visible in-app browser session for user-facing site testing, keep session ids across calls, emit cursor events for visible actions, and report profile-launch conflicts with a visible-workbench retry path.
  • Browser screenshots and annotations: browser workbench screenshots can be saved into the workspace, annotated directly in the app, and sent back to the agent as image attachments for follow-up inspection.
  • Renderer bundle reduction: the app shell lazy-loads secondary renderer views, Mermaid loads only when needed, and syntax highlighting imports a bounded highlight.js/lib/core language set.
  • LLM provider settings: OpenAI, Azure OpenAI, OpenAI-compatible, Pi AI, provider factory, OAuth, reasoning-effort typing, and model-selection behavior were tightened with expanded tests.
  • Image generation routing: ChatGPT subscription sign-in is now treated as a first-class image generation route, OpenAI Codex image generation prefers gpt-image-2, older OpenAI image model aliases are normalized, and setup/error language now points users to ChatGPT sign-in when applicable.
  • Vision routing: image analysis now uses the active image-capable provider path, supports OpenAI OAuth token refresh for vision requests, avoids treating Gemini as an inline-image provider, and gives clearer instructions when the current model cannot analyze images.
  • Follow-up permissions: follow-up messages can carry permission-mode and shell-access overrides so users can continue an existing task with broader or narrower runtime access without recreating it.
  • Presentation artifact planning: prompts that ask for presentations, slide decks, PowerPoint, or .pptx outputs now infer a required .pptx artifact, inject a presentation creation step when needed, and require real artifact evidence instead of accepting screenshots or text as a substitute.
  • Native setup reliability: native setup now routes better-sqlite3 Electron rebuilds through @electron/rebuild with isolated rebuild home directories, avoids unsupported npm job config keys, and keeps Windows ARM64 x64 fallback behavior explicit.
  • Timeline polish: active image-generation lanes show an inline frame preview, pending parallel lanes are treated as active, and follow-up placeholders now consistently invite users to add context or steer current work.
  • Control plane and Symphony runtime: added Symphony service plumbing, registration, protocol handling, and tests for richer control-plane orchestration.
  • Memory Hub controls: Memory Hub settings gained observation-facing controls and supporting styles for inspecting and managing structured memory metadata.
  • File Hub and IPC support: File Hub, preload, and IPC handlers now understand the new artifact preview/edit APIs and format-specific viewer payloads.
  • Documentation refresh: README, docs index, Features, Getting Started, Use Cases, Showcase, Architecture, Development, Troubleshooting, GTM, Project Status, and artifact docs now frame generated files as a unified local-first workbench.
  • Bundled skill metadata: frontend, frontend-design, Kami, Karpathy Guidelines, memory-kit, and routing eval metadata were refreshed.
  • App branding: macOS app/Dock icons and dark-mode renderer assets now use resources/branding/cowork-os-app-logo-dark.png; Windows and light-mode renderer assets continue to use resources/branding/cowork-os-app-logo.png.

Fixes

  • Repeated artifact cards: office/web artifact cards are no longer shown over and over in verbose task timelines; the final output remains visible.
  • Type-check blockers: delivery-mode timeline narrowing and delivery-row tests now type-check cleanly.
  • DOCX data loss: saving existing DOCX files no longer rebuilds the package from only editable text blocks and drop headers, images, comments, section settings, or other unsupported Word structures.
  • CSV data loss: editing and saving large CSV/TSV files no longer truncates rows after the preview window or overwrites the first hidden row when adding a row.
  • Root-relative web assets: sandboxed web previews now handle /assets/... references from built web apps.
  • Browser workbench reliability: visible browser sessions unregister/register through main-process IPC, keep status in sync with the renderer webview, and avoid silently launching separate profile-backed Chrome sessions when a visible session is already available.
  • Image attachment handling: tasks now stop before sending attachments to providers that cannot accept images and tell the user to switch to an image-capable model/provider instead of sending unsupported inline image payloads.
  • Bypass permissions semantics: bypass_permissions now consistently allows approval-gated actions unless a higher-precedence hard policy blocks them, while dont_ask still prompts for data export.
  • Follow-up runtime updates: queued and resumed follow-ups now update executor workspace and agent config state so permission and shell-access changes take effect immediately.
  • Mission Control stale rows: terminal tasks are removed from projected Mission Control task items so completed/cancelled work does not keep counting as active.
  • Provider and runtime reliability: provider parsing, runtime visibility, workspace preflight acknowledgements, verification runtime behavior, mailbox behavior, permission settings, and task pause banners received targeted reliability fixes and tests.
  • Email HTML rendering: email HTML layout handling now has dedicated utility coverage.

Upgrade Notes

  • This release prepares package metadata for 0.5.41; publish only from a clean checkout or release worktree.
  • Before npm publish, explicitly run npm ci --no-audit --no-fund, npm run build, and the tarball validation steps from AGENTS.md. Do not rely on prepack, because .npmrc sets ignore-scripts=true.
  • DOCX editing is intentionally conservative: DOCX can be edited in-app, while legacy/app-owned document formats remain preview/external-app oriented.
  • CSV/TSV and workbook editing is capped to the preview-safe row/column window; hidden source rows are preserved.
  • PPTX rendered slide images depend on local conversion tools. Text-first preview still works when render dependencies are unavailable.
  • Web page previews sandbox generated HTML and inline local assets; React/Vite source projects still need an explicit build output such as dist/index.html.

References

This page is the canonical high-level summary for the changes included in 0.5.41.

Previous

Release Notes 0.5.40

Next

Release Notes 0.5.42

Was this page helpful?Edit this page on GitHub