Learning

Chanto can improve over time without quietly drifting:

• Signals: it records your own corrections locally in ~/.chanto/learned/ (outside the repo):

  • denials with feedback, and messages like "no, use X"
  • edits you make to files it just wrote
  • rewinds, !commands you run by hand, and done-gate failures

  Repo content and tool output never count, so a planted file can't "teach" it anything.

• Proposals: /learned review condenses signals into short lessons, each citing its evidence. Commands you keep running by hand (!cargo clippy three times) become proposed verification checks.

• Approval: nothing applies until you /learned accept <id>. Lessons about permissions, or ones that skip or weaken tests or checks, are refused outright.

• Visibility: accepted rules become feedback memories, indexed on the first turn within the memory budget. Accepted checks join verified mode. /learned lists everything; disable, reject, undo and add manage it.

• Measurability: --no-lessons runs without lessons, so a benchmark can compare with and without them before you trust them.

Turn it off with [learning] enabled = false.