Skip to content

v0.preview.84

Choose a tag to compare

View all tags
@ante-github ante-github released this 19 Aug 02:27
· 20 commits to main since this release
v0.preview.84
653f799

v0.preview.84 - 2026-08-18

Changed

  • Qwen 3.8 models send graded reasoning_effort levels instead of collapsing every level to thinking on/off, and the effort picker lists only the distinct supported levels
  • MCP servers are discovered in parallel (one slow server no longer delays the rest), and MCP tool calls time out after 10 minutes — overridable via ANTE_MCP_TOOL_TIMEOUT — instead of pinning a turn until interrupt

Fixed

  • Headless (ante -p) sessions wait for MCP warm-up before the first turn, so MCP tools are present in the model's schema instead of every MCP call failing silently
  • llama.cpp context overflows are recognized as context-full and trigger the one-time compaction retry instead of failing the turn
  • Writing to a closed or broken output pipe (consumer exits, macOS EIO) is handled leniently instead of panicking during terminal teardown
Assets 6
Loading