Problem

You need a decision that isn't expressible in YAML rules or the default actions matrix. Examples:

Different severity thresholds per tenant.
Different block conditions during business hours vs off-hours (though we recommend against time-based policy, sometimes you need it).
Additional safety checks based on data.json operator overlays.

Solution

Write a Rego module, drop it in ~/.defenseclaw/policy/rego/, validate it, then ask the running gateway to reload policy.

Example: per-tenant strict mode

# ~/.defenseclaw/policy/rego/tenant_overlay.rego
package guardrail.overlay

import future.keywords.if
import future.keywords.in

# Tenants that get the strict treatment
strict_tenants := {"finance", "legal", "hr"}

# If the session is a strict tenant and severity >= MEDIUM, upgrade to block
action_override[action] if {
  input.context.tenant in strict_tenants
  input.verdict.severity in {"MEDIUM", "HIGH", "CRITICAL"}
  action := "block"
}

Example: admission with data.json overlay

# ~/.defenseclaw/policy/rego/admission_overlay.rego
package admission.overlay

import data.trust.allowlist

# Allow anything in the operator trust list
allow if {
  input.kind == "skill"
  input.fingerprint in allowlist.skills
}

With data.json:

{
  "trust": {
    "allowlist": {
      "skills": ["sha256:abc...", "sha256:def..."]
    }
  }
}

Reload and test

defenseclaw policy validate
defenseclaw policy test
defenseclaw-gateway policy reload

defenseclaw policy validate compiles the Rego tree and validates data.json; defenseclaw policy test runs Rego tests; defenseclaw-gateway policy reload tells the running sidecar to swap in the updated policy snapshot.

Testing before reload

Write a Rego test next to the module:

# ~/.defenseclaw/policy/rego/tenant_overlay_test.rego
package guardrail.overlay_test

import data.guardrail.overlay

test_finance_medium_blocks if {
  overlay.action_override["block"] with input as {
    "context": {"tenant": "finance"},
    "verdict": {"severity": "MEDIUM"}
  }
}

test_engineering_medium_not_blocked if {
  count(overlay.action_override) == 0 with input as {
    "context": {"tenant": "engineering"},
    "verdict": {"severity": "MEDIUM"}
  }
}

Run the tests:

defenseclaw policy test
# PASS  ~/.defenseclaw/policy/rego/tenant_overlay.rego
# PASS  tenant_overlay_test.rego

Where overlays are composed

Operator overlays are merged into the compiled snapshot after the shipped defaults. Package naming matters: for the overlay to participate in a decision point, either:

Add to an existing package (package guardrail) to augment, or
Define in package <domain>.overlay — the shipped policies import data.<domain>.overlay and merge its outputs.

See Writing Rego for the full list of overlay hooks.

Caveats

Operator overlays can tighten decisions (turn allow into block) but cannot loosen them (turn block into allow) for security-critical packages. Admission allow-lists are the exception — they're specifically a loosening tool.
Every overlay module must have an associated _test.rego file. policy test fails if an overlay has no tests.
Don't rely on wall-clock time in policy. If you must, centralize in one helper and test with injected time.

Add a custom Rego policy — DefenseClaw