overllm

Version updated for https://github.com/theadamdanielsson/overllm to version v0.7.0.

  • This action is used across all versions by ? repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

What’s Changed

Recall improvements + one new rule, each validated on ~14k files of real code (framework-heavy and amateur) with no new false positives.

  • Matured call detection — follow a model through composition. LangChain chain.invoke(...) and bound-model .invoke(...) are now detected (an LCEL | pipe, .with_structured_output(), or an alias that composes a known model). A prompt | parser chain (no model) or a dict | dict merge is not tracked — precise. langchain detection 385→492.
  • kwargs-splat — create(**params) on the unambiguous chains.
  • Embeddings — embeddings.create in a per-item loop → llm-in-loop.
  • Module-constant prompts — a prompt/model held in a top-level constant used inside a function is resolved.
  • New rule json-mode-missing-json — response_format=json_object with a static prompt lacking “json” is a provable OpenAI 400 (fires only when the absence is provable).

155 tests. Precision held: amateur-corpus findings byte-identical to 0.6.2.