Honcho 3.2 adds include_evidence to trace how answers were built
Honcho 3.2 lets chat calls pass include_evidence=True to return the conclusions the agent read, the messages it referenced and the tools it called, off by default and at no extra model token cost.
New in Honcho 3.2 - Now you can check what an answer about a user was built from before you act on it. Pass include_evidence=True on a chat call and the response comes back with every conclusion the agent read, references to the messages it pulled, and the tools it called. This means you can catch a bad answer before it reaches a user, debug a wrong personalization from one request instead of reproducing it, and tell someone exactly what informed a decision about them. It's collated from what the agent actually accessed rather than self-reported by the model, so it's deterministic and costs zero extra model tokens. It does over-report: a conclusion is listed because it was read, which isn't proof the answer leaned on it. Off by default, so nothing changes for calls you already have. Read more in the Changelog - https://go.honcho.dev/Gr2rMqV