Skip to content

v0.3.6

Latest

Choose a tag to compare

@ante-github ante-github released this 17 Nov 21:13

v0.3.6 - 2026-11-17

Added

  • Gemini 4 Carbon and Hydrogen on the Gemini API, Vertex AI and OpenRouter. Hydrogen defaults to high effort on Gemini API and Vertex, and both routes support xhigh
  • OpenRouter offers Qwen4 Max, Tencent Hy4 Flash and Meta Muse Ember

Changed

  • The OpenAI API and subscription offer GPT-6.2 Astra in place of GPT-6 Astra, at low effort by default. The Astra subscription context window is corrected to 272,000 tokens
  • Gemini 4 output budgets on Gemini API and Vertex rise to 65,536 tokens. Gemini 3.8 Flash is removed from the pickers; Flash-Lite stays
  • Dark and light themes use the v2 palette: higher-contrast greys and blue, code blocks without a background, and explicit colors for notices and approval decisions
  • The offline engine is pinned to llama.cpp b12375
  • Dependency updates

Fixed

  • ante serve waits for MCP tools to register before the first turn. Previously the tool list could change between turns, which broke Claude Fable 5.2 sessions with an invalid thinking signature error and missed the prompt cache on other models
  • A response cut off at max_tokens still runs the tool calls that finished before the cut. Only a call cut mid-arguments is dropped
  • An Anthropic web search that pauses its turn is resumed, up to 3 times
  • An Agent call fails when the subagent's turn did not complete. The parent sees the reason, the child's last message and the transcript path, instead of a success carrying an [interrupted] marker
  • A cancellation that arrives after a response completed no longer marks that response as cancelled
  • Edit keeps the line endings of untouched lines in files that mix LF and CRLF

Crates

  • Public crates published as 0.3.6: ante-protocol-shape, ante-llm, ante-exec, ante-sdk