Skip to content

v0.7.10

Choose a tag to compare

View all tags
@sherifkozman sherifkozman released this 02 Apr 15:00
· 11 commits to main since this release
v0.7.10
2f2b209

Fixes

  • Anthropic thinking deprecation: Switch from deprecated thinking.type=enabled to thinking.type=adaptive for claude-4.x models
  • Anthropic schema rejection: Promote all schema properties to required during Anthropic normalization, staying under the 24-optional-parameter grammar compilation limit
  • Timeout enforcement: asyncio.wait_for now uses the provider-specific bounded timeout instead of the global 120s config timeout — bounded caps are actually enforced
  • Lock budget accounting: Provider call-slot lock wait time is subtracted from the asyncio timeout budget so lock contention doesn't silently extend request duration
  • OpenAI client timeout: Default raised from 15s to 60s to prevent SDK-level timeouts before the orchestrator's own timeout fires
  • Error classification: Server error patterns (500, 502, 503, overloaded) now classified as retryable instead of unknown
  • Unknown error retry: Unknown errors get one retry before continue/skip, so transient CLI and API errors are not silently dropped

Verified

All fixes tested against live providers (anthropic, openai, claude-code CLI) with non-trivial tasks. 319 unit tests passing.

Assets 2
Loading