Changelog

What's new in each release of Saient.

2.0.0-dev.6 3 October 2026 · Linux beta

Native node studio: build and save workflows for prompts, image generation, masks, inpainting and previews, with undo/redo and JSON view. No ComfyUI dependency.
Inpainting: paint a mask to edit part of an image through supported image models.
Agent continuity: retain successful actions and their results between proposals so the agent has a record of its completed work. Command output now preserves meaningful whitespace.
Long Project completion: consistent instructions for requesting the configured acceptance check, with the existing permissions and verification gates.
Confirmations and recovery: repaired asynchronous confirmations, checkpoint handling, interruption and cancellation paths.
Creative tools: repaired image-model selection for LoRA/Merge, removed unsafe checkpoint loading, improved setup dependency checks and implemented WAV saving.
Usability: clearer resident-model controls, save filename handling and matching navigation shortcuts.
Release scope: a manually installed Linux x86_64 Debian-package beta for glibc 2.39+. Previous 1.0.24 Linux/Windows downloads and stable updater feeds remain available. Requirements, SHA-256 and verification details.
Known limits: the local 7B correction and 14B Long Project recovery tests pass, but harder local coding tasks still fail. Model loading can be slow. This beta has no new Windows or AppImage package.

1.0.24 7 September 2026

Evidence-based terminal reports: file read-back, exact paths and the scope of checks are recorded separately from message delivery. Empty folders and unrelated successful commands no longer establish task completion.
Workspace and binding: workspace input is committed only after the backend accepts it; stale folder results are isolated. Rejected bindings require an explicit retry instead of hidden repeat profiling.
Cancellation: interrupted work remains unverified and child-process cleanup is checked.
Windows portability: native PowerShell execution, Python quoting and exit codes, UTF-8 file tools, redirected Desktop paths and cross-process file locking are covered by native Windows CI.
Release identity: Linux and Windows builds use one application commit and a pinned engine commit. Final bundle hashes and all 48 runtime file hashes accompany the downloads.
Verification limits: native Windows runtime, frontend and Rust tests pass; an installed Windows desktop/GPU session has not been verified for this release. Windows engines require Microsoft's Visual C++ v14 Redistributable (x64). Authenticode signing remains unavailable; updater signatures are separate.

1.0.20 13 August 2026

Installed model selected automatically: after downloading a suggested base model, Image Gen switches away from any stale checkpoint selection and selects the complete Diffusers folder
Legacy SDXL repair: loading the old sd_xl_base_1.0.safetensors download now uses an installed complete SDXL Base folder or opens the repair picker when that folder is missing
Correct LoRA selection: newly downloaded LoRAs are selected in the LoRA control instead of altering the base-model control

1.0.19 13 August 2026

Working suggested Image Gen models: SDXL Base, SDXL Turbo, DreamShaper XL, SD 1.5, and SD 2.1 now install as complete Diffusers repositories instead of incompatible bare checkpoint files
Correct model location: full image repositories install into the managed models/image directory that Image Gen actually scans
Checkpoint choice retained: normal search still lets advanced users download tuned single-file checkpoints

1.0.18 12 August 2026

No persistent Hugging Face runtime cache: model libraries are forced offline even when general Internet access is enabled, and run-only library state is cleared at startup
Visible managed assets: Full Setup downloads pinned Kokoro and Moondream files directly into runtime-assets, with disk preflight, byte-size validation, and failed-part cleanup
Storage cleanup: Settings reports managed assets, legacy Hugging Face cache bytes, and incomplete downloads, with a scoped two-step legacy-cache removal action
Windows Image Gen repair: Windows installers now bundle every creative Python helper script, including generate_sdxl.py
Actionable setup errors: setup drains output safely and surfaces the last real subprocess error instead of only an exit code

1.0.17 12 August 2026

Manual updates while offline: Check now and Install can use temporary access limited to Saient's signed update endpoints without enabling Internet for the agent
Accurate update errors: the update dialog now shows the backend failure instead of reporting every problem as an unreachable update server
Fail-closed authority: temporary updater access is revoked after each check or installation attempt and on dialog teardown

1.0.16 12 August 2026

Setup recovery: Settings now has a Setup & repair page with a Run setup again button that preserves existing models, downloads, settings, and environment files
Clear Image Gen diagnosis: missing PyTorch/diffusers packages now point directly to Settings → Setup → Full setup instead of exposing a raw Python import error
Safer skipping: skipping the managed creative environment requires a second confirmation and explains which studios will remain unavailable
Windows Python recovery: runtime discovery now supports the normal Windows Python launchers when a user intentionally supplies their own environment

1.0.15 12 August 2026

Dismissible autonomy controls: the autonomy dialog now has an accessible close button, so an accidental click on Agent no longer traps the user in the dialog or changes the saved project level
Formal local-model binding: Saient checks the selected local model before chat is unlocked and preserves identity and record-authority boundaries
Network controls: normal model traffic stays on numeric loopback endpoints and setup/download authorization is explicit and temporary
Cleaner interface checks: the strict Svelte gate passes with no errors or warnings, including corrected image-lightbox click handling
Direct Pi hosting: Windows x64 and Debian/Ubuntu amd64 packages and signed update metadata are served from saient.co.uk, not Vercel

1.0.9 9 August 2026

Multiline terminal tasks: pasted prompts are coalesced into one agent turn instead of being submitted one line at a time
Correct project context: the terminal agent reports and works from the accepted workspace; the release test used /home/tiny/projects/saient/desktop
Isolated diagnostics: new env and tempdir tools let the agent inspect its environment and create a managed @temp workspace without writing diagnostic files into the project
Loop protection: repeated identical tool calls are bounded and interrupted generations return to the prompt without a Python traceback
Release evidence: the exact diagnostic task completed end to end on the Linux release host; the Windows x64 package passed cross-build, content and updater-signature checks but still awaits a real Windows runtime test

1.0.8 9 August 2026

Workspace switch fix: changing the accepted project now restarts the Agent terminal in that project instead of leaving it in the previous workspace
Windows x64 test installer: the first published free no-key current-user NSIS package is hosted directly on saient.co.uk with the CPU GGUF engine
Signed future Windows updates: the Windows app uses Tauri's signed updater and direct Saient-hosted release endpoint; installation from older builds is manual once
Honest Windows boundary: the cross-built package passed compile, package-content and updater-signature checks, but installation/runtime and GPU/Python-studio behaviour still require a real Windows machine
Clean platform packages: the Linux package contains only Linux CPU/CUDA engines, while the Windows package contains only the Windows CPU engine

1.0.7 9 August 2026

One-click Linux updates: after the signed package is downloaded and verified, Saient installs it and restarts the desktop app
Existing administrator policy: machines already configured for passwordless administrator commands update without another prompt
Normal Linux fallback: when passwordless approval is unavailable, Saient asks through the operating system's standard administrator dialog
No GitHub detour: release metadata and Debian packages continue to come directly from saient.co.uk

1.0.6 9 August 2026

Signed in-app Linux updates: Saient can download the Debian package, verify its Ed25519 signature, size and SHA-256, request normal administrator approval, install it and restart
Direct Saient hosting: update metadata and packages are served from saient.co.uk; the update button no longer detours through GitHub
Safe package rename: the Debian package declares its replacement of the historical ai-workshop package instead of forcing file overwrites
One-time bootstrap: v1.0.3 users still need to install this package manually once because v1.0.3 only opened a download page
Local inference remains free and requires no Saient licence key or hosted-inference key

1.0.4 9 August 2026

Free, no-key release: removed the former trial and product-key gate; no Saient account or hosted-inference key is required for local inference
Persistent agent runtime: project modes, checkpoints, memory/provenance and the Saient loop are bundled with the desktop application
Separate safety controls: project write permission and Internet permission are distinct
Distribution at release: the audited Debian/Ubuntu amd64 package shipped first; the replacement Windows test build was first published with v1.0.8 and AppImage remains pending
The historical v1.0.3 release remains available only as an explicitly labelled obsolete licensed build

1.0.3 Image Gen fixes

Fixed: video models no longer show up in the Image Gen model list (they have their own tab) — clears a confusing load error
Clearer message if a model can't be used for image generation
NSFW safety checker fully disabled — generations are never blacked out

1.0.2 Windows download fix

Fixed: models downloaded in Image Gen / Merge now appear in the dropdowns on Windows (home-folder path resolution)

1.0.1 Polish & Windows fixes

Image & model downloads: one-click base-model shortcuts (SDXL, SD 1.5/2.1, DreamShaper) plus search, in Image Gen & Merge
Analyze images right in chat — attach a picture and ask about it
Model knows your OS — suggests the right (Windows/Linux) commands
Smoother on CPU — no false "connection dropped", longer warm-up before timeout
Windows polish: no flashing console windows, fixed temp paths, terminal greeting

1.0.0 First public release

Chat with any local GGUF model, with live HTML artifacts
Saient coding agent — terminal + plan→act→verify loop, sandboxed
Image generation (SDXL + LoRA, optional face-detail pass)
Video generation (Wan & CogVideoX, image-to-video, refine & upscale)
Vision analyzer — ask questions about any image (Moondream2)
Text-to-speech and LoRA training + checkpoint merge
One-click model downloads from any Hugging Face GGUF repo, straight into your models folder
Launch password protection; this historical release still used the former trial/product-key gate and is now obsolete
Runs 100% offline on Linux & Windows — bundled CUDA engine, no setup