Every release makes the harness harder to fool: LLMKube 0.9.19

LLMKube 0.9.19 was released this morning, revealing a surprising twist behind its creation. Initially, about half of the code was developed due to the detection of lies from the agent pipeline, occurring in four different ways. The other half was written after fixing the pipeline. This release demonstrates that every update since 0.8.0 has made the system harder to deceive, with this version showing the loop closing on itself.

The release consists of nine files with full test coverage and zero production callers, initially appearing as dead code. However, the fix to identify and remove these dead files was straightforward - simply run the linter a second time without test files in the analysis graph, which successfully exposed and removed them. The release also addresses several issues that went unnoticed previously.

These include a linter not warning about unused package-level functions, a reviewer agent with an empty system prompt and missing tool list leading to incorrect review decisions, and a bug class where an issue mentioning "add tests for X" resulted in a poorly-written, unwinnable test suite. The release introduces several improvements to address these problems, such as the introduction of a "receipts layer" where rails record why they could not run, an execution-first reviewer rubric that instructs reviewers to run new tests and probe near-miss cases, and a bounded verification that accurately claims exhaustiveness.