This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Summary
AI summaryRomanize now correctly handles mixed-language dictation without disabling streaming cleanup.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Low |
Dictated issue reports, feedback, tasks, and requirements can become numbered lists without an explicit list command. Dictated issue reports, feedback, tasks, and requirements can become numbered lists without an explicit list command. Source: llm_adapter@2026-07-19 Confidence: high |
— |
| Feature | Low |
Literal list markers are preserved rather than rewritten by a hidden post‑processing heuristic. Literal list markers are preserved rather than rewritten by a hidden post‑processing heuristic. Source: llm_adapter@2026-07-19 Confidence: high |
— |
| Performance | Medium |
Server stop‑to‑final processing time improved from 4.519 s to 0.952 s on a 70‑second dictation. Server stop‑to‑final processing time improved from 4.519 s to 0.952 s on a 70‑second dictation. Source: llm_adapter@2026-07-19 Confidence: high |
— |
| Bugfix | Medium |
Romanize no longer disables streaming cleanup for Latin-script dictation, keeping long English dictations on the fast path. Romanize no longer disables streaming cleanup for Latin-script dictation, keeping long English dictations on the fast path. Source: llm_adapter@2026-07-19 Confidence: high |
— |
| Bugfix | Medium |
Fallback to full-text romanization when a Latin opening is followed by a Hindi-heavy final segment, preventing native-script leakage. Fallback to full-text romanization when a Latin opening is followed by a Hindi-heavy final segment, preventing native-script leakage. Source: llm_adapter@2026-07-19 Confidence: high |
— |
| Bugfix | Low |
Streaming list items remain on separate lines with continuous numbering from 1; invalid segment-level numbering falls back to safe whole‑text cleanup. Streaming list items remain on separate lines with continuous numbering from 1; invalid segment-level numbering falls back to safe whole‑text cleanup. Source: llm_adapter@2026-07-19 Confidence: high |
— |
Full changelog
This supersedes 0.10.5 with one important mixed-language correctness fix.
What changed:
- Romanize no longer disables streaming cleanup for Latin-script dictation, so long English dictations stay on the fast during-recording path.
- A Latin opening followed by a Hindi-heavy final segment now falls back to full-text romanization instead of leaking native-script text through the English streaming prompt.
- Dictated issue reports, feedback, tasks, and requirements can become numbered lists without an explicit list command.
- Streaming list items stay on separate lines, numbering must remain continuous from 1, and invalid segment-level numbering falls back to safe whole-text cleanup.
- Literal list markers are preserved rather than rewritten by a hidden post-processing heuristic.
Measured on an archived 70-second dictation, server stop-to-final improved from 4.519s to 0.952s.
Verified with 334 engine tests, 581 Swift self-checks, stricter exact-model formatting benchmarks, Apple notarization, stapling, and Gatekeeper assessment.
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Velora
All releases →Beta — feedback welcome: [email protected]