Guardrails: migrate guardrails to the v1/checks endpoint

docs/filters/ai_guardrails.md

@@ -21,7 +21,7 @@ Calls an external AI guardrail provider to evaluate request and response bodies.
 filter: ai_guardrails
 provider:
   type: nemo
-  endpoint: "http://nemo:8000/v1/guardrail/checks"
+  endpoint: "http://nemo:8000/v1/checks"
   allow_private_endpoint: true
   timeout_ms: 5000
 phase:

examples/configs/nemo-guardrails-response.yaml

@@ -1,15 +1,17 @@
 # Guardrails (NeMo) - Response Phase Only
 #
 # Evaluates upstream responses against a NeMo Guardrails service.
+# Requires NeMo Guardrails >= 0.24 (the /v1/checks endpoint).
 # Request-side guardrails are disabled so the request always reaches
 # the upstream. This config is used by integration tests to verify
 # response-phase behavior in isolation.
 #
 #   Response phase:
-#     success  - response forwarded to the client unchanged
+#     passed   - response forwarded to the client unchanged
 #     blocked  - response body replaced with a JSON error payload
 #                (status remains 200 because headers are already committed)
-#     error    - response body replaced with a JSON error payload
+#     modified - forwarded unchanged until redaction support lands
+#     provider HTTP error - response body replaced with a JSON error payload
 #                (same as blocked; headers are already committed)
 
 listeners:
@@ -24,7 +26,7 @@ filter_chains:
       - filter: ai_guardrails
         provider:
           type: nemo
-          endpoint: "http://127.0.0.1:3001/v1/guardrail/checks"
+          endpoint: "http://127.0.0.1:3001/v1/checks"
           allow_private_endpoint: true
           timeout_ms: 5000
         phase:

examples/configs/nemo-guardrails.yaml

@@ -1,12 +1,18 @@
 # Guardrails (NeMo) - Request Phase
 #
 # Evaluates incoming requests against a NeMo Guardrails service.
+# Requires NeMo Guardrails >= 0.24 (the /v1/checks endpoint).
 # For response-phase guardrails, see nemo-guardrails-response.yaml.
 #
 #   Request phase:
-#     success  - request forwarded to the upstream unchanged
-#     blocked  - 403 returned to the client (triggered rail names in body)
-#     error    - 500 returned to the client
+#     passed   - request forwarded to the upstream unchanged
+#     blocked  - 403 returned to the client (triggered rail name in body)
+#     modified - forwarded unchanged until redaction support lands (#49)
+#     provider HTTP error - 500 returned to the client
+#
+# Each user message triggers one /v1/checks call (cumulative slices).
+# max_message_checks (default 32) caps the number of calls per request;
+# conversations exceeding the limit are rejected (fail-closed).
 #
 # Start a local NeMo Guardrails instance before running this config.
 #
@@ -36,7 +42,7 @@ filter_chains:
           type: nemo
           # NeMo callouts are anonymous; no downstream authentication
           # headers are forwarded to this endpoint.
-          endpoint: "http://127.0.0.1:3001/v1/guardrail/checks"
+          endpoint: "http://127.0.0.1:3001/v1/checks"
           allow_private_endpoint: true
           timeout_ms: 5000
         phase:

filters/src/guardrails/config.rs

@@ -11,7 +11,7 @@ use serde::Deserialize;
 /// filter: ai_guardrails
 /// provider:
 ///   type: nemo
-///   endpoint: "http://nemo:8000/v1/guardrail/checks"
+///   endpoint: "http://nemo:8000/v1/checks"
 ///   timeout_ms: 5000
 /// phase:
 ///   request: true
@@ -32,7 +32,7 @@ pub(super) struct AiGuardrailsConfig {
 #[derive(Debug, Clone, Deserialize, PartialEq, Eq)]
 #[serde(rename_all = "lowercase")]
 pub(super) enum ProviderType {
-    /// NVIDIA `NeMo` Guardrails via `/v1/guardrail/checks`.
+    /// NVIDIA `NeMo` Guardrails via `/v1/checks`.
     Nemo,
 }
 

filters/src/guardrails/filter.rs

@@ -32,7 +32,7 @@ const DEFAULT_MAX_BODY_BYTES: usize = 1_048_576;
 /// filter: ai_guardrails
 /// provider:
 ///   type: nemo
-///   endpoint: "http://nemo:8000/v1/guardrail/checks"
+///   endpoint: "http://nemo:8000/v1/checks"
 ///   allow_private_endpoint: true
 ///   timeout_ms: 5000
 /// phase:
@@ -49,7 +49,7 @@ const DEFAULT_MAX_BODY_BYTES: usize = 1_048_576;
 ///     r#"
 /// provider:
 ///   type: nemo
-///   endpoint: "http://nemo:8000/v1/guardrail/checks"
+///   endpoint: "http://nemo:8000/v1/checks"
 ///   allow_private_endpoint: true
 /// "#,
 /// )
@@ -270,7 +270,7 @@ fn record_verdict(
         },
         GuardResult::Block { reason } => Ok(enforce_block(body, reason, phase, phase_label, verdict)),
         GuardResult::Redact { reason, .. } => {
-            tracing::warn!(verdict, phase = phase_label, %reason, "ai_guardrails: verdict; forwarding unchanged until #579");
+            tracing::warn!(verdict, phase = phase_label, %reason, "ai_guardrails: verdict; forwarding unchanged until #49");
             Ok(FilterAction::Continue)
         },
     }

filters/src/guardrails/providers/mod.rs

@@ -55,10 +55,9 @@ pub enum GuardResult {
         reason: String,
     },
     /// Content contains sensitive data — forward with masked text.
-    #[cfg_attr(
-        not(test),
-        expect(dead_code, reason = "used once /v1/checks migration adds redaction support")
-    )]
+    ///
+    /// `modified_text` is populated from `NeMo` but not applied to the request
+    /// or response body until redaction support lands in `#49`.
     Redact {
         /// Provider-rewritten text with sensitive data masked.
         modified_text: String,