Sirius 0.6.2 Beta 1

Sirius 0.6.2 Beta 1

This release concentrates on predictable local work: Sirius keeps the fast
Direct Qwen path for ordinary requests, gives fx only the capabilities a turn
needs, and prevents local inference clients from competing for the same model.

What changed

• Direct Qwen, fx, BrowserCode, and model prewarm now share one serialized
  inference lane. This prevents overlapping prefill, duplicate model pressure,
  and another native inference request is already active failures.
• Compiled workspace context now follows the served model's actual context
  window and reserves space for output and tool results.
• Long fx conversations are bounded deterministically while preserving the
  newest complete user and tool turn.
• Capability-based fx selection advertises only the workspace, terminal,
  browser, app-control, specialist, or extended tools required by the request.
• Direct Qwen retains bounded native workspace tools for common file work, so
  fx is not started when the model can safely complete the task itself.
• Recoverability contracts and an inspectable log now record model swaps,
  memory-scope changes, context truncation, and interrupted specialist handoffs.
• KITKAT cross-project context remains read-only while reviewed proposals can
  be written only to the governed Sirius Memory channel.
• The fx diagnostic log is privacy-safe, line-bounded, and rotates at 1 MiB.
• Smart Inbox requests remain isolated, hash-checked, and review-only.

Verification

• Full Swift test suite
• Optimized release build
• Developer ID signature verification
• Apple notarization and ticket stapling
• Gatekeeper assessment
• SHA-256 release checksum

Package

• Version: 0.6.2 Beta 1
• Build: 29
• File: Sirius-0.6.2-beta.1-arm64.dmg