Atraium 0.2.3 Preview 1

Atraium 0.2.3 Preview 1

Atraium 0.2.3 Preview 1 introduces reusable Model Provider Profiles, granular Room and Workspace tool routing, safer Chat Completions context handling, and more capable Workspace actions for the assistant.

Recommended target

• Windows desktop: recommended preview experience
• Android: not included in this release unless listed in the attached assets
• Web: not currently supported

Highlights

Reusable Model Provider Profiles

Provider configuration is now centralized in reusable Model Provider Profiles. Configure a provider target once, then assign it wherever it is needed.

• Create, edit, filter, and delete reusable provider profiles from Settings -> Model providers.
• Store provider, endpoint, API mode, model, deployment, token limits, pacing, pricing, image settings, and credentials in one profile.
• API keys remain in secure platform storage and are not shown again after saving.
• Assign compatible profiles independently to Chat, Utility, Embeddings, Image generation, Realtime, Text-to-speech, and Speech-to-text.
• Advanced LLM settings now choose capability profiles instead of duplicating endpoints, keys, and model settings.
• Guided configuration now begins with provider-profile setup, followed by capability assignment.
• Existing per-capability settings remain supported as a compatibility fallback while you move to profiles.
• Open-weight models can be connected through a compatible OpenAI-style endpoint, including Qwen 3.5, Qwen 3.6, and KAT-Coder V2.5 variants where the serving stack supports the required API behavior.

Room and Workspace Chat profile overrides

Room and Workspace overrides now select a full Chat provider profile instead of overriding only a model or deployment name.

• A Room can override the global Chat profile.
• A Workspace can override its Room profile.
• Effective precedence is:

Workspace Chat profile -> Room Chat profile -> Global Chat profile

• A selected profile applies the complete Chat target: provider, endpoint, credentials, API mode, model, deployment, token limits, reasoning effort, sampling, pacing, and pricing.
• Atraium safely refreshes the affected Chat runtime at the next request boundary.

Flexible reasoning and sampling controls

Provider profiles better support modern reasoning models.

• Temperature and Top P are optional and clearable.
• Clearing either field omits that parameter from provider requests instead of sending a default value.
• This avoids unsupported-parameter failures from models that reject sampling options.
• Reasoning effort is now an autocomplete field with standard suggestions:

minimal
low
medium
high
xhigh

• Custom provider-specific reasoning values remain supported.

Fine-grained direct and gateway tool routing

Room and Workspace tool controls now separate availability from route.

• Configure sources and individual tools with inherited, enabled, or disabled availability.
• Route enabled tools through the Gateway or expose them Directly to the active model.
• Room policy establishes the baseline; Workspace policy can refine it.
• Direct routes are supported for Responses, Anthropic, and Semantic Kernel Chat Completions paths.
• Direct tools are registered correctly in the Semantic Kernel, including wrapped local tools such as the Text tools.
• Up to 24 catalog tools can be exposed directly; any overflow remains gateway-accessible.

Coding-agent compatible tool surface

Atraium now exposes a compact Coding tool source for models trained on common coding-agent conventions.

• Direct core tools are:

Read
Write
Edit
Glob
Grep
Bash
apply_patch

• The surface follows familiar Claude Code, Codex, OpenClaw, Cline, and Roo Code concepts while preserving Atraium's specialist tools.
• Common gateway aliases are accepted, including read_file, read_files, write_to_file, replace_in_file, search_files, execute_command, exec, shell_command, and apply_diff.
• Bash accepts one normal command. Its optional cwd is no longer required for ordinary project work: Atraium infers a safe working directory from accessible file or project paths and prefers a detected repository or solution root.
• apply_patch accepts the standard *** Begin Patch envelope with add, update, delete, and move operations.
• Read, Glob, and Grep provide lightweight discovery; Edit is preferred for exact targeted changes; Write is reserved for complete file replacement.
• The Coding source is direct by default when enabled, but Room and Workspace policy can disable it or route it through the gateway.
• Existing Text_*, structured configuration, Roslyn, Visual Studio, and document tools remain available for advanced or specialized work.

Assistant direct-tool access requests

The assistant can request direct access to an eligible catalog tool, but it cannot grant access to itself.

• Requests are stored as durable pending proposals scoped to the Room or active Workspace.
• Review proposals in Workspace -> Tools & access.
• Approve a request to apply the scoped Direct route, or dismiss it without changing access.
• Requests include the tool, source, scope, and assistant rationale.

Faster, organized Workspace Tools

The Workspace Tools surface is more responsive for large catalogs.

• The tool catalog loads lazily when Tools is opened and remains cached while moving between Workspace surfaces.
• Tool groups stay organized by source or category.
• Source-level policy controls remain immediately available.
• Individual tool controls are rendered only after expanding Fine-tune this category.
• Expanded category lists use virtualization to avoid building hundreds of routing controls at once.
• Search and an explicit catalog refresh action are available from the Tools surface.

More ergonomic Workspace actions

Assistant Workspace actions now avoid unnecessary round trips.

• Create tasks that are already complete with isComplete: true.
• Reuse IDs returned by Workspace create, update, and list operations directly.
• Update a task's title, description, priority, due date, and completion state in one operation.
• Delete tasks and notes when the user explicitly asks to remove them.
• Room and Workspace writes now use short-lived database contexts and bounded concurrency retries, reducing conflicts with simultaneous assistant or chat updates.

Chat Completions context caching control

Chat Completions profiles now include an advanced Chat Completions context caching option.

• Default: off. Atraium keeps the root system message stable and inserts current-turn context only into a cloned provider request immediately before the latest user message. This preserves prompt-cache reuse.
• Optional: on. Enable Update system message with current-turn context to re-render the single pinned root system message in place each turn. Enable this for Chat Completions models that accept a system message only in the first position, including affected open-weight or quantized server builds.
• The root system message remains protected from history compaction.
• The Responses API route retains its existing request-local context behavior.

Settings and navigation refinements

• Settings, Model Providers, and guided setup components now use corrected scoped-CSS isolation patterns for MudBlazor content.
• Model Providers Back navigation now returns through Atraium's shared navigation stack and no longer loops between Settings and Model Providers.
• Room and Workspace configuration use the new Chat profile selector rather than legacy model-name fields.
• Chat composer, message stream, and responsive Room layouts include additional mobile quality-of-life refinements.

What to try

1. Open Settings -> Model providers and create a reusable Chat profile.
2. Assign it as the global Chat profile, then optionally select a different profile from Room or Workspace configuration.
3. Open Workspace -> Tools & access, confirm the Coding source is enabled, then ask the assistant to inspect a project with Read, Glob, or Grep and make a small exact change with Edit.
4. Ask the assistant to run a project command through Bash without specifying cwd when the command includes an accessible project path.
5. Set a small, suitable non-Coding category to Direct, or ask the assistant to request the direct tool it needs and approve or dismiss the proposal from the Tools surface.
6. Create a Chat Completions profile and leave context caching off for stable prompt caching. Enable it for an open-weight model or server that requires the system message to remain first, such as affected KAT-Coder V2.5 GGUF configurations.
7. Ask the assistant to record completed work using a task created with isComplete: true.

Important limitations

• This is preview software and may contain breaking issues.
• Direct model-tool exposure is limited to 24 catalog tools. Additional enabled tools remain available through the gateway.
• Model availability, feature support, quotas, regions, and parameter support vary by provider, account, endpoint, and model.
• OpenAI-compatible providers can differ in their support for API modes, tool calling, reasoning parameters, prompt caching, system-message ordering, and chat templates.
• Tool behavior depends heavily on the selected model and its tool-calling quality.
• Provider-side usage, billing, and policy controls remain authoritative.
• Support is best-effort.

Downloads

This preview release includes:

• Atraium-0.2.3-preview.1-win-x64-setup.exe
• Atraium-0.2.3-preview.1-win-x64-portable.zip
• Atraium-0.2.3-preview.1-SHA256SUMS.txt

Verification

Verify the SHA-256 checksum of the downloaded installer or archive against the attached checksum file.

Token usage, costs, and credentials

You are solely responsible for all usage and charges from the providers, endpoints, plugins, MCP servers, and infrastructure you configure.

Atraium does not include provider credentials, model credits, or token allowance. It does not enforce your provider budget, control provider pricing, or issue refunds for provider charges.

Tool-rich workflows can involve multiple model calls, long context, searches, embeddings, images, speech, retries, and continued work. Configure provider-side limits and monitor your provider dashboard.

Read Token usage and provider billing before using paid credentials.

Feedback

Use GitHub Issues for reproducible bugs and structured feedback. Use Discussions for broader questions and workflow ideas.

Do not publish API keys, tokens, private endpoint URLs, confidential prompts, source code, or sensitive files.