llm-echo 0.5a0

The more telling detail here is that llm-echo and datasette-llm are shipping alpha releases on the same day, suggesting Willison is coordinating updates across his plugin suite to track LLM's alpha builds in lockstep rather than letting individual tools drift out of sync with the core library.

This pairs directly with the datasette-llm 0.1a7 release from the same day (covered here), where model-level configuration defaults landed. Taken together, the two releases sketch a coherent strategy: standardize how models are configured, then ensure mock tooling faithfully replicates their output signatures, including reasoning traces. That combination matters for teams building CI/CD pipelines around LLM-powered data workflows, because you need both layers to test reliably without burning API credits. The procedural execution research from arXiv in early May is also relevant context: if models are unreliable across multi-step tasks, having mock infrastructure that accurately simulates their output structure becomes more valuable, not less, since it lets developers isolate failures in their own logic rather than attributing them to model behavior.

Watch whether LLM's stable release formalizes extended reasoning output as a first-class interface contract. If it does, llm-echo's mock implementation will have effectively set the testing baseline before the spec was finalized, which would confirm Willison's plugins are tracking internal roadmap rather than reacting to public releases.