@magnitudedev/cli@0.2.1 @magnitudedev/cli@0.2.2
medium risk
cheaper-inference/OmniRouteSep 30, 2026 → Oct 1, 20260 days between releases
What's New
Fixes concurrency hangs, assessment stalls, and Codex tool call failures.
Breaking Changes
- The `count_tokens` function now requires the `max_tokens` parameter.
Improvements
- Model speed estimation now uses memory bandwidth instead of GPU kernels to prevent hangs.
- Repeated images are encoded once to reduce overhead.
- Large system prompts are cached to avoid full re-reads.
- Codex request continuation improves efficiency for follow-up calls.
- Request errors now properly terminate the request instead of leaving it waiting.
Bug Fixes
- Fixes app getting stuck on 'Assessing models' during hardware assessment.
- Fixes Codex hanging after the first tool call by improving request continuation.
- Fixes token usage counting for Anthropic and streaming responses.
- Fixes tool call IDs repeating across turns, which caused Claude Code to drop calls.
- Fixes tools with free-form object parameters failing on Gemma 4.
Upgrade Notes
If you use the `count_tokens` function, ensure you provide the `max_tokens` parameter. Tool call handling for free-form object parameters has been updated to support Gemma 4.
The Two Releases
@magnitudedev/cli@0.2.1
@magnitudedev/cli@0.2.1
Sep 30, 2026· 22 days ago
Patch release resolving concurrency issues where local models would hang indefinitely when handling multiple requests simultaneously.
Full release notes@magnitudedev/cli@0.2.2
@magnitudedev/cli@0.2.2
Oct 1, 2026· 21 days ago
Patch release addressing issues where the app would hang during model assessment, Codex tool calls would fail, and token usage was counted incorrectly.
Full release notes