release: 0.24.3

[yuanhao]

A patch release: reliability fixes from the October codebase review. No breaking changes.

Included:

• fix(providers): retry overloads; failed streams are errors; mark cancels #232 providers:
  • 503/529 and in-stream overloads are retried;
  • failed streams (OpenAI-compatible, Gemini, Vertex) are errors, not truncated answers, and Vertex shares Gemini's parser;
  • a cancel during tools is marked, and Batched { size: 0 } no longer panics.
• docs: onboarding fixes and a testing guide; anthropic() infers Claude options #233 onboarding:
  • ModelConfig::anthropic infers the Claude options from claude- ids;
  • the cli example's key handling is fixed;
  • the quick starts and the Ollama URL are corrected;
  • a new testing guide.
• fix(mcp): stdio id matching, real notifications, drained stderr, call timeout and cancel #234 MCP stdio:
  • responses are matched by id, and initialized is a real notification;
  • stderr is drained, with its tail reported on exit;
  • the call timeout and cancel work;
  • unknown content types are read leniently.
• fix(search): return matches with warnings when some files cannot be read #235 search: matches are kept when some files can't be read, with the errors returned as warnings.
• docs: explain what max_total_tokens counts #237 docs: what max_total_tokens counts, and that limits apply per run.
• feat(sub-agent): with_execution_limits — the token and duration caps are no longer hard-coded #238: SubAgentTool::with_execution_limits; sub-agent caps are no longer hard-coded.

Behaviour to note: a run cancelled while tools run now ends with an [Agent stopped: cancelled] user message. A bare anthropic("claude-…") now infers its thinking mode and native structured outputs.

Release checks:

• Price audit: pass.
• Live smoke (DeepSeek, run locally): 5/5.
• yoyo-evolve main against this code: 6277 tests pass. The only failure is its doc_version_claims guard, which asks for the version markers to be re-checked on any bump.
• integrations/* require ^0.24.0, which this version satisfies.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG