OptionalallowedPositive allowlist for the pre-execution action gate (#57). When set, AgentArmor.checkAction admits only requests matching one of these rules; everything else fails closed. An empty array denies all actions.
OptionalbehaviouralPer-detector toggles within the Behavioural Control category. All default true.
OptionalcognitivePer-detector toggles within the Cognitive State category. All default true.
OptionalcontentPer-detector toggles within the Content Injection category. All default true.
OptionalcustomCustom detectors to add to the pipeline
OptionalmaxLargest input, in characters, that a scan will take on. Longer input is
not scanned: the result is not clean, carries one congestion-trap
threat from detector input-limit, and has an empty sanitized, so an
oversized payload can neither stall the scan nor pass as safe. Set to
Infinity to turn the limit off. Default: 1,000,000.
OptionalmlML classifier configuration (requires @stylusnexus/agentarmor-ml)
OptionalnormalizeApply Unicode normalization (NFKC + confusable folding + invisible-char stripping) before semantic detectors run, so homoglyph-obfuscated payloads are matched. Evidence and offsets still report against the original text. Structural detectors (content-injection) always scan the raw input. Default: true.
OptionalonDiagnostics callbacks — route internal warnings/errors to your own logging (#24).
OptionalsemanticPer-detector toggles within the Semantic Manipulation category. All default true.
OptionalsessionMulti-turn / session scanning configuration (#35)
OptionalstrictnessConfidence-threshold preset. Default: 'balanced'.
OptionaltransportPer-detector toggles within the Transport Integrity category. All default true.
Configuration passed to AgentArmor.regexOnly or AgentArmor.create.