even Opus struggles here

The field is probably too strict for a model-generated metadata field. It is useful but it is not part of the actual tool action. Failing the whole tool call because because of that is brittle, especially for weaker/non-native function-calling models. A solution could be 1.

security_risk

optional in the schema. 2. If it is missing, default to

UNKNOWN

. 3. Let

ConfirmRisky(confirm_unknown=True)

handle safety by requiring confirmation for unknown risk when confirmation mode is active. 4. Add a few more normalizations for common misspellings/casing. 5. Ideally only request

security_risk

when an analyzer/policy will actually use it. What do you think? @limited-student-10747

I think unknown seems like a good idea!

-> https://github.com/OpenHands/software-agent-sdk/pull/3126

@chilly-france-44439 I might request your review on this one since I think you've dealt with this the most

@modern-fireman-99755 lgtm and i've added a few more tests