AI Runner Files

Linux v1 release-prep batch

A batch of hardening, correctness fixes, and release-planning documentation, produced against the Linux v1 release tracker and independently code-reviewed before merge.

Fixes and hardening

• S01/S02: Authenticate the LLM WebSocket before accept; bound message size and rate-limit per principal (fixed: unbounded max_tokens when omitted, missing extra="forbid", malformed-JSON handling).
• D01/D02: Honor pinned model revisions and verify digests before load; resume only the same model artifact and promote it safely (fixed: a fail-open identity check on partial-file resume, and an unverified "already complete" promotion shortcut).
• D03: Return every image from a requested art batch, not just the first.
• D04: Validate art request resource limits before job creation (fixed: cfg_scale rejected the documented 0.0 value for Z-Image Turbo).
• O01/O02: Define and enforce an explicit offline egress policy, applied to model downloads and online tools (fixed: an error message pointing at a nonexistent Preferences control).
• O03: Make diagnostics private and user-controlled (fixed: the export could leak arbitrary exception-message content a denylist sanitizer couldn't catch; now an allowlisted failure-code shape).
• P01: Stop compiling Qt UI resources at app startup (fixed: a packaging regression that broke pip install from the sdist).

Design and documentation

• B01: Desktop companion contracts and port-mapping design (fixed: a missing cancellation method, an incorrect assumption about Desktop's conversation model, and a non-persisting worked example).
• R01/R02: Desktop feature inventory matrix and hardware/model validation manifest (fixed: several citation errors in R01).
• R03: Corrected stale tooling paths in Copilot instructions.

All fixes were re-verified against their full test suites (186+ tests across the touched areas) before merge.

Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com