Aura

This release separates Planner and Worker provider selection, refines vision preprocessing pipeline for DeepSeek(or non multi modal), simplifies prompts for faster dispatch, and includes a large batch of fixes and quality-of-life improvements.

Independent Planner/Worker Backends

The Planner and Worker can now use different AI providers and models, persisted per role.

• AppSettings includes planner_provider and worker_provider fields.
• The Settings dialog offers dedicated provider and model selectors for each role, with dynamic model lists.
• The sidebar model selectors reflect the correct provider per role.
• Conversation files store and restore the role-specific providers.
• Existing settings and conversations default to the global DeepSeek provider for backward compatibility.

Vision Preprocessing for DeepSeek

Aura now supports image inputs by routing them through a preprocessing workflow.

• When an image is attached, a lightweight vision model describes it.
• The text description is injected into the user message for the main chat model.
• Prevents raw image payloads from being sent to text-only models.

Prompt Simplification

System prompts have been stripped down to reduce over-prompting.

• Planner and Worker prompts are leaner, cutting unnecessary instruction bloat.
• The result is faster dispatch, lower token usage, and a snappier workflow.

Workflow & UI Additions

• Retry last message button: Resend the last user message without retyping.
• Improved file editing animations: Smoother transitions when code cards update, making changes feel more responsive.
• Tool control model overhaul: Category caps (read/dispatch/research/terminal/write limits) have been removed. Loop detection now drives Planner recovery from repeated failures, with structured repeated-failure detection added to loop_detection.py.

Fixed

• Added startup diagnostics so launch failures produce a useful log instead of failing silently.
• Hardened startup error handling around the main window initialization path.
• Settings loading now safely falls back on invalid or stale provider/model values.
• Fixed Settings changes not propagating to the send handler, toolbar, worker handler, and conversation persistence.
• Fixed Settings accidentally resetting max_tool_rounds.
• Single-mode provider/model behavior now syncs from the main provider/model when Planner/Worker mode is off.
• Fixed unsupported image sends: Aura no longer sends raw image payloads to non-vision models; the user’s prompt and attachments are preserved after the error.
• Fixed onboarding’s “Open Settings” button so it actually opens Settings and does not trap users in the first-launch loop.
• Fixed SpecCard status updates so worker cards no longer stay stuck on “Worker running...” after completion or cancellation.
• Improved packaged update validation for newer Python runtime markers.
• Added Anthropic model catalog/pricing and native Anthropic Messages streaming support.

Improved

• More defensive runtime behavior around providers, models, settings, and image handling.
• More consistent Settings → runtime state synchronization.
• Better user feedback when selected models cannot process image input.

Full Changelog: compare v1.3.1...v1.3.2