@krynix/policy

Policy evaluation engine for Krynix — YAML policy parsing, rule matching, and trace evaluation with deterministic CI exit codes.

Install

npm install @krynix/policy

Key Exports

• evaluate(trace, policy) — evaluate a trace against a policy, returns verdict + violations
• matchRule(event, rule) — match a single event against a rule
• parsePolicy — parse a YAML policy file into a typed Policy object

Usage

import { evaluate, parsePolicy } from "@krynix/policy";
import { readTrace } from "@krynix/core";

const events = await readTrace("/path/to/trace.jsonl");
const policy = parsePolicy(`
  apiVersion: krynix.dev/v1
  kind: Policy
  metadata:
    name: no-shell-exec
    version: "1.0"
    description: Block shell command execution
  spec:
    scope:
      agents: ["*"]
      event_types: ["tool_call"]
    rules:
      - id: block-shell
        description: Deny shell tool calls
        match:
          event_type: tool_call
          payload:
            - field: tool_name
              operator: eq
              value: shell_exec
        action: deny
        severity: error
        message: "Shell execution is not permitted"
`);

const result = evaluate(events, policy);
// result.verdict: "pass" | "fail" | "require-approval"
// result.exitCode: 0 (pass), 1 (error), 2 (critical), 3 (require-approval)
// result.violations: array of matched deny/require-approval rules

Evaluation Semantics

• First-match-wins: for each event, the first matching rule determines the outcome. Order your rules from most specific to most general.
• Scope filtering: events outside scope.agents / scope.event_types are skipped entirely.
• Deterministic exit codes: same trace + same policy always produces the same result.

Operators

Exit Codes

Rule Types

• Per-event rules: Match individual events based on event_type and payload field conditions.
• Sequence rules: Match patterns across multiple events in a session (e.g., "tool_call followed by tool_call without an intervening llm_request").

License

Apache 2.0