Off Grid AI Desktop 0.0.41-beta.76

Highlights

New

• Clean, awaitable unload - free the model port without a force-quit
• Use the model's trained context window as the ceiling, not a fixed cap
• Inline STT model picker (view + change) in meetings
• Expose active STT engine/model provenance (core)

Fixes

• Explicit null check for the in-flight init in unload (SonarCloud reliability)
• Wrap composer chips instead of overflowing the input box
• Idle (inter-token) stream timeout, not total-duration - long answers survive
• Don't yank the view down while streaming when the user scrolls up
• Stop capping the agentic answer at 1024 tokens - inherit the max-output setting
• Default max output to auto (until EOS / window fills), not a 2048 cap
• Reattach restores the image-gen progress panel on remount

What's Changed

• fix(llm): explicit null check for the in-flight init in unload (SonarCloud reliability) (0eacf94)
• docs(gaps): log 4 deferred edges from desktop-pro #32 review (Gitar) (a0a3972)
• docs(gaps): log 2 deferred items from PR #60 review (bench resilience, provenance fallback) (4134368)
• review: address CodeRabbit findings on the release PR (805280c)
• chore(test): set the coverage floor to a uniform 85% across all metrics (486a84f)
• fix(chat): wrap composer chips instead of overflowing the input box (3123a8b)
• test(settings): label the ctx/max-output selects + update the response-limit test (932ff51)
• fix(llm): idle (inter-token) stream timeout, not total-duration — long answers survive (e7aef2f)
• fix(chat): don't yank the view down while streaming when the user scrolls up (3f163f6)
• fix(tools): stop capping the agentic answer at 1024 tokens — inherit the max-output setting (f9ead97)
• fix(llm): default max output to auto (until EOS / window fills), not a 2048 cap (63cc87d)
• feat(engine): clean, awaitable unload — free the model port without a force-quit (e2ce09c)
• feat(llm): use the model's trained context window as the ceiling, not a fixed cap (ab9af57)
• fix(chat): reattach restores the image-gen progress panel on remount (78b2b96)
• chore(bench): vision-only capture benchmark with batch sweep + quality judge (3889417)
• feat(transcription): inline STT model picker (view + change) in meetings (ef05f07)
• feat(transcription): expose active STT engine/model provenance (core) (a5918a5)

Full Changelog: v0.0.41-beta.75...v0.0.41-beta.76