DStudio 1.1.0

What's new since 1.0.0

DStudio 1.1.0 expands the original local Chat, Agent and Design workspace into a broader multi-model studio, while keeping inference, files and tools under the user's control.

Learn and knowledge work

• Added Learn, an interactive learning-path workspace with mandatory Deep Research, prerequisite-aware curricula, factual and curriculum quality gates, editable progress, checkpoints, high-resolution exports and dedicated Tutor rooms.
• Added Cowork, a folder-scoped knowledge-work agent for PDF, DOCX, PPTX, ODT, RTF, Markdown, XLSX, CSV and TSV workflows.
• Cowork now has Chat-style drag-and-drop attachments, clean file previews, live collapsed Thinking, a structured action timeline and direct paginated PDF creation through the safe write_pdf tool—without arbitrary shell access.
• Cowork respects the context selected in Settings instead of forcing the 393k true-Max profile.

Models and native multimodality

• Added GLM 5.3 Flash support on the unified upstream ds4/main runtime, replacing the former GLM 5.2 path.
• Added native image understanding for DeepSeek V4 Flash Vision-Exp and GLM 5.3 across Chat, Agent, Cowork and Design using the matching encoder of the currently loaded model.
• Removed the secondary Qwen visual-router path. Qwen3-Embedding-0.6B is used only for local multilingual text retrieval.
• Laguna S 2.1 remains strictly text-only: image tools are not exposed and scanned PDF pixels are explicitly reported as uninterpreted.
• Improved DSpark selection, model discovery, resumable downloads, visible partial transfers and persistent engine-checkout handling.

PDFs, media and creation

• PDFs now combine bounded local text extraction, multilingual semantic retrieval and up to four selected pages rendered for the active DeepSeek/GLM native vision encoder.
• Added and optimized local MiniMax H3 text-to-video and image-to-video generation with synchronized audio, including optional two-photo references.
• Hardened the direct Ideogram/Hunyuan image pipeline, model residency hand-off and interruption behavior.

Agent, Design and interface

• Agent, Cowork and Design now stream Thinking inside the disclosure from the first token and format reads/tools in a single timeline from the beginning of a turn.
• Fixed the Design interrupt hand-off race: new prompts wait for the previous SIGINT to settle and retry without creating failed phantom tasks.
• Preserved independent working folders and KV sessions per Agent, Cowork and Design conversation.
• Added Design chat beneath the brief, renewed loading/progress surfaces, stricter quality gates and more reliable canvas/refinement workflows.
• Added consistent dark-theme support to Learn and aligned user images/PDF previews beneath their message bubbles.
• Added Reverse Structure Analysis alongside the existing Plan and Guided Security Analysis workflows.

Platform and reliability

• Added and hardened the portable Windows build, WebView2/MSYS2 runtime handling, browser CDP launch and clean Agent interruption.
• Added the free macOS installer and DMG distribution path with clearer Gatekeeper guidance.
• Updated the reversible Agent/Cowork patch set for current ds4/main and Laguna sources; upstream source files are restored after every derived-runtime build.
• Expanded deterministic, browser, runtime, PDF, Office, media and release gates. The complete make -j4 check-fast suite passes for this release.

Distribution

This is a source-only release. No prebuilt .app is attached; use GitHub's source archives and follow the build instructions in the README.

Full changelog: v1.0.0...v1.1.0

Current source refresh

• Install missing now runs the generated GSA installer as a supervised background task, reports live task/log state, rechecks the catalog on completion and uses space-safe caches on macOS, Linux and Windows.
• Agent and Design prompts submitted while session maintenance is settling are queued and sent automatically instead of being dropped or producing phantom failures.
• Oversized DSpark configurations are no longer silently disabled: DStudio explains the estimated Metal-memory risk, asks for explicit confirmation and preserves DSpark when the user chooses to continue.
• Chat, Agent, Cowork and Design now pass external DSpark checkpoints with the current upstream --mtp-model interface; the former --mtp FILE mismatch that exited into CLI help is fixed.
• Added browser, launcher, loading and concurrency regressions. make -j4 check-fast passes on the refreshed source.

The release remains source-only; no .app or other binary asset is attached.

Frontend E2E and session lifecycle refresh

• Added an isolated real-runtime Playwright gate covering Chat, Agent, Cowork, Design, Learn and Settings with the saved DeepSeek Vision-Exp model; H3 and image generation/edit jobs remain excluded.
• Fresh Agent/Cowork/Design conversations no longer adopt stale or truncated bytes from the previous global pipe buffer, and internal /new/session bookkeeping is never rendered as assistant prose.
• Agent/Cowork/Design are now reported ready only after the authoritative WAITING marker, preventing prompts from landing during system-prompt prefill.
• Settings model discovery is bounded and falls back to exact installed DStudio model paths when macOS File Provider blocks directory enumeration.
• The reversible JSONL patch set now accepts the current upstream ds4 source shape for edit 063 without modifying the upstream checkout.

Validated with the complete deterministic suite, the signed macOS bundle smoke test and the live DeepSeek Vision frontend E2E.

Runtime prefill and startup refresh

• Design now honors the context selected in Settings at every Thinking level instead of silently forcing a 393,216-token allocation.
• Selected Design systems and Skills are loaded on demand, keeping their full packs and the 150+ item catalog out of the bootstrap prompt.
• Design stores an exact-text, model-validated system-prompt KV cache so repeated starts and fresh sessions can restore the prefix instead of recomputing it.
• Agent, Cowork and Design expose real system-prefill token counters; unchanged status polls are deduplicated and indeterminate work no longer shows a misleading linear ETA.
• Added a Playwright cold-start regression and revalidated the complete check-fast suite across Chat, Agent, Cowork, Design, Learn, Settings, GSA and RSA.

The release remains source-only; no .app or other binary asset is attached.

Visual element refinement

• Full-screen Design previews now include Select to edit: click the exact element, choose General, Layout, Text, Style, Image or Video, and leave a contextual instruction beside the selection.
• DStudio sends bounded selector, visible text, DOM attributes and geometry as private runtime evidence; only the human-readable comment appears in the conversation.
• The Design agent measures the selected target with inspect_layout, reviews the rendered composition with see_page, applies a focused revision and reruns its quality gates.
• Generated sites remain interactive inside an opaque sandbox. The annotation bridge is embedded in macOS and Windows builds, preserves relative assets and project CSP behavior, and serves image, video and font files with their correct MIME types.
• Browser, launcher, contract and packaging regressions cover the complete selection-to-Design hand-off.

Validated with the complete fast test suite and a freshly signed DStudio 1.1.0 macOS bundle. The release remains source-only; no app or binary asset is attached.

Conversation canvas recovery

• Reopening a completed Design conversation now reconstructs its canvas and selects the artifact owned by that chat.
• Existing 1.1.0 histories are migrated from their structured artifact event or referenced workspace HTML; new runs persist the association directly.
• Async snapshots are conversation-bound, so a background Design run cannot open or contaminate another chat's canvas.
• The contextual visual-edit form is hidden until its anchored position is calculated, eliminating a one-frame flash at the viewport origin.

GSA tool transparency and true Max context

• Every managed GSA command now shows its complete effective parameters directly in the live and completed Agent timeline; long invocations wrap and every value—including tokens, passwords, cookies, URL query values and Authorization headers—is shown verbatim.
• Thinking now defaults to high. Choosing true max below 384k explains the context and Metal-residency cost before restart, and leaving max restores the previous context.
• User prompts in Agent and Cowork have a copy control that remains reliable while the transcript updates.
• Settings migration preserves an explicitly selected video profile.
• Added catalog-wide contract coverage and Playwright regressions; the complete make check-fast suite passes.

The release remains source-only; no .app or other binary asset is attached.