@avrldotdev Rser, and truncates repeated line of stack traces etc. We could also look into a lightweight semantic summarization before saving to the steps table. Also instead of files we can store git diffs to store the changes.
@avrldotdev It would store hashes of files, so if manual edits it doesnt overwrite the code made by the developer blindly.
Obv not a raw dump would happen,
I would first structure the data with typed tool schemas and Pydantic for agent thoughts. Before commiting, we run a deterministic pa
@luisvelasco I honestly would love to see agentic evals, as well as the harness utilisation of Codex on premise, and how you guys deploy and what considerations you guys make while doing so