Appearance
Permission Decision
A permission profile describes risk facts. A permission decision records what actually happened for one invocation.
Behaviors
| Behavior | Meaning |
|---|---|
allow | The call may proceed. May include updated_input and feedback. |
ask | The runtime must ask a user, coordinator, policy tool, or external authority. |
deny | The call must not execute. Produce a terminal error result. |
passthrough | The tool delegates final permission resolution to another policy layer. |
Decision reasons
Initial reason.type values:
rulemodepolicyhookclassifierworking_dirsandbox_overridesafety_checkpermission_prompt_toolasync_agentsubcommand_resultsother
Rule sources
Permission rules SHOULD preserve origin:
user_settingsproject_settingslocal_settingsflag_settingspolicy_settingscli_argcommandsession
Suggested updates
An ask decision MAY include suggested updates such as adding allow/deny/ask rules, replacing rules, changing permission mode, or adding/removing scoped working directories. Suggested updates are not decisions until accepted by the authority that owns policy state.
Classifier state
Automated safety classifiers SHOULD be represented as decision inputs, not hidden side effects. Record pending_classifier_check, classifier_ref, confidence, classifier_reason, and whether the classifier can approve, deny, or only recommend.