Atraium 0.3.1 Preview 1

Atraium 0.3.1 Preview 1

Atraium 0.3.1 Preview 1 is a major experience release. It modernizes the conversation, workspace, navigation, settings, and administration surfaces; turns Model Providers into a first-class workflow; adds persisted Light, Dark, and System themes; improves long-conversation reliability; and expands reasoning and token-usage support across providers.

Recommended target

• Windows desktop: recommended preview experience
• Android: early sideloaded preview for Android 7.0 (API 24) or later, with arm64-v8a device and x86_64 emulator support
• Web: not currently supported

Highlights

Android preview

This release includes a signed Android APK for compatible Android 7.0 (API 24) or later devices. It contains arm64-v8a device and x86_64 emulator runtimes.

• Download the direct APK to sideload, or the Android preview bundle for the same APK with preview, privacy, security, and billing documents.
• The Android package uses normal runtime prompts only when a user invokes camera capture, microphone input, location, or calendar integration.
• Android is an early preview. Windows desktop remains the recommended target for the broadest local-workflow coverage.
• Web, macOS, and iOS are not public preview targets in this release.

A redesigned conversation experience

The main chat surface has been rebuilt around a cleaner, more focused conversation canvas.

• Assistant and user turns use a flatter, more readable visual hierarchy.
• On wide layouts, the assistant response appears first, followed by reasoning and tool activity so live work remains near the bottom of long turns.
• On compact and mobile layouts, process information remains before the response for easier conversational scanning.
• Reasoning has a dedicated disclosure instead of being mixed into the visible response.
• Tool activity uses a structured timeline with clearer running, completed, failed, and attention states.
• Activity now fills the complete assistant-turn width.
• Message actions, attachments, code, Markdown, and generated artifacts have more consistent spacing and responsive behavior.
• The composer has been simplified and remains anchored to the conversation column.
• Empty, loading, error, and retry states have been standardized across chat surfaces.

More reliable long conversations

Long and virtualized conversations now open and navigate more predictably.

• A conversation initially renders its latest message window instead of visibly opening near the start and animating through the thread.
• Initial bottom positioning waits for the rendered layout to settle.
• Late content growth from images, Markdown, code blocks, and enhanced content is followed while bottom anchoring is active.
• Jump to latest now uses immediate, verified, retry-safe positioning rather than fragile smooth scrolling.
• Prepending older messages preserves the reader’s scroll position.
• A Chromium regression test covers layout growth after the initial anchor begins.

Better attachments

Attachment presentation has been rebuilt for both the composer and message history.

• Long filenames truncate safely without pushing remove, Preview, Download, or Copy controls outside the card.
• Full attachment names remain available through accessible tooltips.
• Compact cards reserve space for their actions and stay contained at narrow widths.
• Image previews, file metadata, and attachment actions use consistent responsive layouts.
• Workspace attachments have a dedicated browser for reviewing files associated with the active workspace.

Modernized workspace surfaces

Room and Workspace interaction now uses a more coherent shell.

• The conversation header, Workspace picker, and surface navigation have been redesigned.
• Workspace Tasks, Notes, Tools, Artifacts, and Attachments use consistent headers and inspector surfaces.
• Notes and Tasks have clearer empty states, filtering, editing, and action placement.
• Workspace context can use a persistent inspector on wide layouts and an overlay surface on compact layouts.
• The message canvas no longer competes with unnecessary nested outer cards.

New application navigation

The left navigation has been completely reworked.

• The compact rail shows a bounded set of pinned and recent rooms instead of an unbounded avatar column.
• The active room remains visible in the compact rail.
• A +N affordance exposes additional rooms.
• Expanded navigation includes room search, grouped rooms, and a flatter Manage section.
• Room selection uses native keyboard-accessible buttons, with pin and overflow actions kept separate.
• Desktop expanded and mini states are preserved.
• Mobile uses the same hierarchy in a temporary drawer.
• Windows title-bar and lower-edge insets are accounted for so the final navigation item remains visible.
• Model providers now has its own direct navigation destination.

Persisted Light, Dark, and System themes

Atraium now includes a complete appearance preference stored in UserPreferences.

• Choose System, Light, or Dark from Setup & Settings.
• Existing and new profiles default to System.
• Theme changes apply immediately and persist across restarts.
• System mode follows live Windows appearance changes.
• A desktop title-bar action provides a quick Light/Dark toggle.
• MudBlazor palettes and Atraium’s own semantic design tokens switch together.
• Chat, dialogs, navigation, attachments, inputs, code, diagnostics, file drop, image preview, and administration surfaces have dark-mode-aware colors.

Administration and settings parity

The remaining administration pages now follow the same modern design language.

Modernized surfaces include:

• Setup & Settings
• Guided Configuration
• Advanced Settings
• Model Providers
• Provider Profile Setup
• Backup & Restore
• Plugins
• MCP Servers

Common improvements include:

• consistent semantic page headers;
• pinned Back, navigation, Add, Upload, and Save actions;
• compact sticky headers on small windows;
• wider administration canvases on desktop;
• responsive and container-aware layouts;
• flatter cards and clearer section hierarchy;
• standardized loading and empty states;
• improved destructive-action placement;
• full Light and Dark theme support.

Guided Configuration also has a clearer step list, improved active-step status, container-driven responsive behavior, and a sticky Previous/Next/Save footer.

Model Providers as a first-class experience

Model Providers has been redesigned to make configuring, validating, routing, and switching models substantially faster.

Live capability routing

A new Live routing dashboard shows what currently powers every model capability:

• Chat
• Utility
• Embeddings and memory
• Image generation
• Realtime
• Text-to-speech
• Speech-to-text

Each capability shows its active profile and only compatible alternatives. Selecting another profile persists the route immediately for future requests.

Provider inventory and focused editor

The former nested profile accordion has been replaced with:

• a searchable provider inventory;
• stable profile selection;
• one focused editor at a time;
• visible provider, model target, routing usage, credential readiness, and health;
• advanced tuning that remains available without dominating the inventory.

Long provider and model names are contained and ellipsized rather than overflowing routing cards or editor columns.

Faster provider creation

The Add provider menu now includes starting templates for:

• OpenAI-compatible endpoints
• Anthropic Claude
• Google Gemini

Templates populate suitable endpoint, model, API mode, and pacing defaults while keeping every field editable.

Adapter-backed connection verification

The focused editor now includes Test connection.

• Verification uses the same shared OpenAI, Anthropic, and Gemini client construction as the assistant runtime.
• It performs a minimal non-streaming request with a short timeout.
• It sends no room, workspace, or conversation content.
• Credentials are never displayed in the result.
• Changing the endpoint, model, deployment, API mode, or credential invalidates a previous result so stale health is never shown.

Results are normalized into actionable states such as:

• Connected
• Endpoint required or invalid
• Model target required
• Credential required or rejected
• Model target not found
• Rate limited
• Timed out
• Provider unreachable
• Unexpected verification failure

A connection test is a real provider request and may incur a negligible provider charge.

Switch Chat models from the conversation

The room command bar now includes a Chat provider quick switcher.

• Switch the active Workspace Chat profile without leaving the conversation.
• Return to the inherited global Chat route when desired.
• Open Manage providers… directly from the switcher.
• A model change starts a fresh conversation target to avoid mixing incompatible provider context.

Model Providers is also directly accessible from the application navigation instead of requiring a trip through Settings.

Tagged reasoning from open-weight and non-OpenAI models

Atraium now recognizes reasoning emitted in these tags, case-insensitively:

<think>...</think>
<thinking>...</thinking>
<thought>...</thought>

• Tagged reasoning is removed from normal assistant Markdown.
• Extracted content appears in the existing reasoning disclosure.
• Tags split across streaming chunks are supported.
• Multiple and unclosed reasoning blocks are handled safely.
• Ordinary less-than text is not mistaken for a reasoning tag.
• Native provider reasoning and tagged reasoning can coexist.
• Reasoning is persisted in the assistant metadata envelope and survives history reloads.
• Older stored messages containing tags are separated when loaded.

This support applies across OpenAI Responses, OpenAI-compatible Chat Completions, Anthropic, Gemini, and other IChatClient-based providers.

Complete provider token-usage capture

Token usage is now captured and stored consistently across supported Chat paths.

• OpenAI Responses: input, cached input, output, and total tokens.
• OpenAI-compatible Chat Completions: streaming usage is requested with stream_options.include_usage and read from the final update.
• Anthropic: usage is collected from the provider’s Microsoft.Extensions.AI usage content, including multi-call tool workflows.
• Google Gemini: prompt, cached prompt, candidate, thought/reasoning, and total tokens.
• Duplicate usage objects are ignored.
• Usage reported before an interrupted or failed stream is retained.

Normalized usage flows through Atraium’s existing usage transaction storage with provider, model, deployment, API mode, token counts, configured pricing, and calculated cost where enabled.

Atraium records provider-reported usage. If an OpenAI-compatible endpoint does not return usage, Atraium does not invent an estimate.

Accessibility, performance, and cleanup

This release also includes a broad quality pass.

• Added a global reduced-motion fallback.
• Preserved visible keyboard focus.
• Improved decorative image, routed-error, and status semantics.
• Removed nested interactive controls from room rows.
• Improved minimum target sizes and compact-screen action access.
• Removed obsolete custom dialogs and other unreferenced legacy UI controls.
• Added performance regression coverage for large activity histories and incremental assistant streaming.
• Added focused component, browser, startup, navigation, theme, provider, reasoning, and usage tests.
• Fixed the startup lifecycle regression that could incorrectly reopen First-time Setup for an already configured profile.

What to try

1. Open a long Room and confirm it starts at the latest message, then use Jump to latest after scrolling upward.
2. Attach a file with a very long name and confirm Preview, Download, Copy, and Remove remain visible.
3. Switch between System, Light, and Dark in Setup & Settings, then restart Atraium to verify persistence.
4. Open Manage -> Model providers and review the Live Routing dashboard.
5. Add a provider from an OpenAI-compatible, Anthropic, or Gemini template and run Test connection.
6. Switch the Workspace Chat profile from the conversation command bar, then return to the inherited global route.
7. Use a model that emits <think>, <thinking>, or <thought> blocks and confirm the content appears only in the reasoning disclosure.
8. Review Usage analytics after sending requests through Chat Completions, Anthropic, and Gemini profiles.
9. Resize an administration page to a narrow window and confirm its pinned header keeps essential navigation and Save actions available.
10. On a compatible Android device, sideload the APK, configure a low-risk Chat profile, and confirm the first Room opens and saves normally.

Important limitations

• This is preview software and may contain breaking issues.
• Provider capabilities, API modes, models, parameters, usage reporting, quotas, regions, and pricing vary by provider and endpoint.
• A successful connection test confirms that a minimal request worked; it does not guarantee tool calling, streaming, image, speech, reasoning, or high-context compatibility.
• Connection tests are real provider requests and may be billable.
• Switching Chat profiles starts a fresh conversation target by design.
• Some OpenAI-compatible servers do not return token usage even when requested.
• Provider-side usage, billing, rate limits, and policy controls remain authoritative.
• Native interaction testing remains especially important for file pickers, camera, voice, paste, drag and drop, screen readers, and mobile devices.
• Android is an early sideloaded preview; feature behaviour and device integration require native-device testing.
• Support is best-effort.

Downloads

This preview release includes:

• Atraium-0.3.1-preview.1-win-x64-setup.exe
• Atraium-0.3.1-preview.1-win-x64-portable.zip
• Atraium-0.3.1-preview.1-android.apk
• Atraium-0.3.1-preview.1-android-preview.zip
• Atraium-0.3.1-preview.1-SHA256SUMS.txt

Verification

Verify the SHA-256 checksum of the downloaded installer, archive, or APK 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, connection tests, and continued work. Configure provider-side limits and monitor both Atraium’s Usage analytics and 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.