v0.2.0: TOSEM-RCR Edition (2026-01-29)

TOSEM-RCR paper got accepted (2026-01-29)

Please note that pyproject.toml still shows the wrong version (either v0.0.1 or v0.0.2-wip).

This is a retcon as I am currently cleaning up the repository while preparing for a new release. After the TOSEM paper got accepted, I started to work on the TOSEM Replicated Computational Results Report (RCR), basically showing how to reproduce my setup.

The report can be found on the TOSEM website, it is actually quite nice as it showcases how you can also use all the analysis scripts and how to setup the testbed (GOAD).

For this, I mostly added support to run cochise within docker. Main problem for the RCR was mostly to document how to setup GOAD.

I was working on new features in a separate branch, those changes are now intermingled in this version, including:

• removal of the text-based logging (as I was only using the JSON-based logging together with cochise-replay)
• various fixes to improve gemini token accounting
• tried to introduce a separate analyzer agent between executor and planner
• added MITRE ATT&CK TTP classification (directly by the agent, so this is more or less a bit of LLM-as-as-Judge)
• added separate knowledge gathering phase