The automode security classifier's scope description was sharpened to clarify exactly what it does and does not block.
Details
- Scope now explicitly reads: "prevents destructive, hard-to-undo, or security-relevant actions only."
- Explicitly out of scope: "fabricating or misreporting results; posting updates the user didn't ask for; poor code, wrong answers, looping, wasted tokens." These are for the user to correct, not the classifier.
- A new
EDIT REMOVALSannotation is now included in classifier context: removals inEditcalls are noted withremoves/addsfields. Deletions are now weighted as seriously as additions. IfremovesTruncated: true, the full removal should be treated as significant. - NotebookEdit calls are annotated with
modeandcell_id, and deleted cell content is treated as unverifiable per the high-severity rules.
Evidence
Revised classifier prompt (search for "EDIT REMOVALS: Edit calls show both \removes\" and "destructive, hard-to-undo, or security-relevant actions")
Strings lifted out of the shipped bundle, so the claim above can be checked against them.