Promptfoo 0.123.0
Promptfoo 0.123.0 introduces a breaking change defaulting GPT-5.6+ models to the Responses API, exposes MCP tool calls in response metadata, supports native audio grading with llm-rubric, and adds broad model support including GPT-6 Astra, Grok 4.6, and Claude Fable.
原始内容为英文；当前页面提供中文导航与来源说明，具体事实请以原文为准。
为什么值得关注
Promptfoo 0.123.0 brings a significant breaking change by defaulting GPT-5.6+ models to the Responses API. The release significantly expands provider coverage, adding support for Claude Fable, Mythos 5.1, GPT-6 Astra, Grok-4.6, grok-imagine models, Gemini 3.8 Flash, Muse Spark 1.3, GPT Image 2.5 generation, and GPT Transcribe file uploads. Additionally, MCP tool calls are now exposed in response metadata, and zero-argument MCP commands are allowed. In assertions, native audio can now be graded with llm-rubric, tool-call-f1 recognizes OpenAI Responses function call items, and multiline tool calls are properly parsed. The release also completes the deprecation and retirement of GitHub Models inference, alongside numerous provider fixes for Bedrock, SageMaker, Vercel AI SDK 6, and local model parameters.
可执行摘要
Promptfoo 0.123.0 changes the default provider interface for GPT-5.6+ to Responses API and exposes MCP tool calls within response metadata. It updates assertions with multiline tool parsing, tool-call-f1 support for Responses function calls, and native audio grading via llm-rubric. It also retires GitHub Models inference and incorporates broad dependency and provider fixes.
- Agent 实用度
- 82/100
- 可信度
- 96%
- 机器格式
- JSON + Markdown
开发者应核对什么
- Review and update configurations targeting GPT-5.6+ models to accommodate the new default Responses API behavior.
- Migrate away from GitHub Models inference, which has been fully retired in this release.
- Verify tool call evaluation workflows to take advantage of MCP tool calls exposed in response metadata and multiline parsing fixes.