v0.5.18

alibaba/MNNv0.5.18Jul 3, 2026by decolua

AI Summary

This release focuses on improving usage tracking accuracy by calculating cached token costs correctly and introduces new model support for providers like NVIDIA and Kiro. It also addresses several critical bugs, including stream pipe crashes and cross-IdP account overwrites.

Key Highlights

Added accurate tracking for cached tokens and corrected cost calculations.
Introduced new models and capabilities for NVIDIA and Kiro (Claude Sonnet 5).
Implemented regional routing and key validation for Xiaomi-tokenplan.
Fixed critical crashes related to non-SSE streams and cross-IdP account conflicts.

New Features

Track cached tokens and correct input/output/cache cost (#2209)
Show reset credit expiry details in Codex (#2290)
Add new models and capabilities for NVIDIA
Add provider support for ClinePass

Full Release Notes

Features:
- Usage: track cached tokens + correct input/output/cache cost (#2209)
- Codex: show reset credit expiry details (#2290)
- NVIDIA: add new models and capabilities
- ClinePass: add provider support

Fixes:
- Usage: dedupe streaming request-details log entries
- Claude: drop foreign thinking signatures in passthrough
- Prevent non-SSE stream pipe crash and cross-IdP account overwrites (#2244)
- Kiro: route IdC auth to regional CodeWhisperer surface (#2297)
- Kiro: add Claude Sonnet 5 model support (#2264)
- Xiaomi-tokenplan: region selector, key validation, multi-connection (#2251)
- Translator: strict Anthropic content block compliance (#2225)
- Kimchi: strip reasoning_content echo to bound multi-turn input tokens
- Kimchi: bump User-Agent to kimchi/0.1.40 (#2256)
- Codebuddy-cn: strip empty tool_calls arrays to preserve reasoning
- Antigravity: preserve Claude tool delta index (#2223)
- MITM: generate root CA on server startup (#2228)