* finalized wording on the messages to make agent targeting clear
* pushing message frequency for evals, want to see behavior on small n
* adjusted the stop hook to only check the last turn of interactions, not the whole transcript
* emitting a block decision
* fixing some missing names
* missed another spot
* shifted message to tell agent not to discuss feedback with user
* shifting back to additionalContext
* retargeting feedback hooks to preserve user replies
Co-authored-by: Cursor <cursoragent@cursor.com>
Committed-By-Agent: cursor
* dropped tool batch, reporting on every use, leaving sample high for one last test
* lowering submission bar by a smidge
* added debug log
* added a few more hooks
* forgot posttoolbatch
* strengthening the hook which fires after failure
* tried including a comprehensive command example right in the steer
* prepare feedback hooks for production
Use release sampling rates and keep skill usage reporting silent.
Co-authored-by: Cursor <cursoragent@cursor.com>
Committed-By-Agent: cursor
* Update Claude plugin feedback steering
Co-authored-by: Cursor <cursoragent@cursor.com>
Committed-By-Agent: cursor
* minor version bump, as these hooks are a pretty material change but nonbreaking
* moved claude plugin script tests into a symmetric directory under a new top-level , that way test files don't ship to our users
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
John O'Sullivan