Skip to content

How it works

What it does

Every prompt, tool call, and response that flows through an agent harness session is captured by the plugin, scanned against your rule packs, and routed to an outcome — before the content reaches the model (prompts, tool inputs) or after it responds (tool outputs) — then logged.

Traffic flow
Prompts Tool calls Responses File reads AKA policy engine Monitor Warn Redact Block Exception

How rules are applied

Every event is scanned against your installed rule packs.

  • Most rules match with regex — a pattern shaped like an AWS access key, an email address, a bank routing number — since that covers the majority of secrets and PII, which have a predictable shape.
  • A smaller set of rules match on keyword presence (contextual phrases like "[REDACTED:PII] is"), and some regex matches pass through a post-match validator — a Luhn checksum for card numbers, a Shannon-entropy check for high-entropy secrets — to cut down false positives.
  • Either way, a match becomes a Finding: the rule id, category, severity, and matched span, before policy decides what happens to it.

Tip

See Built-in Detections for the full catalog, or Writing Rules to add your own.

Policy outcomes

Every finding resolves to one of five outcomes:

Outcome What happens
Monitor Logged only — every rule ships active under this default, so nothing is enforced until you promote it
Warn Surfaces a warning in the session; the content still goes through unchanged
Redact The matched value is replaced in place before it reaches the model — only available on tool inputs/outputs, not prompt text
Block The prompt or tool call is stopped entirely, with a message explaining what fired
Exception A manually-granted, exact-value override that lets one specific match through despite its rule's policy

Promote any detection to Warn, Redact, or Block from the dashboard, per rule or per category.

Granting exceptions

Sometimes a blocked or redacted value is legitimate such as a sandbox credential, a documented test fixture. An exception is the sanctioned way through: an explicit, audited, exact-value grant made from a terminal, never from inside the session it would exempt. See Adding Exceptions for the full walkthrough.

Key concepts

Term Description
Event A prompt, response, or code change captured from an agent harness session
Finding A rule match produced by the detection engine against an event
Rule A JSON file describing what to detect: keyword list, regex pattern, or validator
Rule Pack A directory of rules + mandatory fixtures with a manifest.json
Policy A decision: what action to take when a rule or category fires
Plugin The harness extension (hooks) that intercepts sessions — one package, used by both Claude Code and Claude Desktop