feat(core): explain why the provider blocked a response (#52518)

packages/ai/src/protocols/anthropic-messages.ts

@@ -451,6 +451,9 @@ const AnthropicStreamDelta = Schema.Struct({
   signature: Schema.optional(Schema.String),
   stop_reason: optionalNull(Schema.String),
   stop_sequence: optionalNull(Schema.String),
+  stop_details: optionalNull(
+    Schema.Struct({ category: optionalNull(Schema.String), explanation: optionalNull(Schema.String) }),
+  ),
 })
 type AnthropicStreamDelta = Schema.Schema.Type<typeof AnthropicStreamDelta>
 const decodeAnthropicStreamDelta = Schema.decodeUnknownOption(AnthropicStreamDelta)
@@ -1417,10 +1420,14 @@ const onMessageDelta = (
       stopSequence === null || stopSequence === undefined
         ? state.pendingFinish?.providerMetadata
         : providerMetadata(state.providerMetadataKey, { stopSequence })
+    const category = event.delta?.stop_details?.category
+    const explanation = event.delta?.stop_details?.explanation
     return {
       reason: {
         normalized: mapFinishReason(stopReason),
         raw: stopReason,
+        ...(category ? { category } : {}),
+        ...(explanation ? { explanation } : {}),
       },
       providerMetadata: finishMetadata,
     }

packages/ai/src/schema/events.ts

@@ -264,6 +264,10 @@ export type ToolError = Schema.Schema.Type<typeof ToolError>
 export const FinishReasonDetails = Schema.Struct({
   normalized: FinishReason,
   raw: Schema.optional(Schema.String),
+  /** The provider's policy area for a content-filter finish, such as `cyber`. */
+  category: Schema.optional(Schema.String),
+  /** The provider's human-readable reason for a content-filter finish. */
+  explanation: Schema.optional(Schema.String),
 }).annotate({ identifier: "LLM.FinishReasonDetails" })
 export type FinishReasonDetails = Schema.Schema.Type<typeof FinishReasonDetails>
 

packages/ai/test/provider/anthropic-messages.test.ts

@@ -1673,6 +1673,40 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("carries a refusal's category and explanation on the content-filter finish", () =>
+    Effect.gen(function* () {
+      const refusal = (stop_details: unknown) =>
+        LLMClient.generate(request).pipe(
+          Effect.provide(
+            fixedResponse(
+              sseEvents(
+                { type: "message_start", message: { usage: { input_tokens: 5 } } },
+                { type: "message_delta", delta: { stop_reason: "refusal", stop_details }, usage: { output_tokens: 0 } },
+                { type: "message_stop" },
+              ),
+            ),
+          ),
+        )
+
+      expect(
+        (yield* refusal({
+          type: "refusal",
+          category: "cyber",
+          explanation: "This request was declined because it could enable cyber harm.",
+        })).finishReason,
+      ).toEqual({
+        normalized: "content-filter",
+        raw: "refusal",
+        category: "cyber",
+        explanation: "This request was declined because it could enable cyber harm.",
+      })
+      expect((yield* refusal({ type: "refusal", category: null, explanation: null })).finishReason).toEqual({
+        normalized: "content-filter",
+        raw: "refusal",
+      })
+    }),
+  )
+
   it.effect("assembles streamed tool call input", () =>
     Effect.gen(function* () {
       const body = sseEvents(

packages/core/src/session/runner/publish-llm-event.ts

@@ -543,7 +543,15 @@ export const createLLMEventPublisher = (bus: Pick<Bus.Interface, "publish">, inp
         }
         if (event.reason.normalized === "content-filter") {
           providerFailed = true
-          yield* failAssistant({ type: "provider.content-filter", message: "Provider blocked the response" })
+          yield* failAssistant({
+            type: "provider.content-filter",
+            message: [
+              `Provider blocked the response${event.reason.category === undefined ? "" : ` (${event.reason.category})`}`,
+              event.reason.explanation,
+            ]
+              .filter(Boolean)
+              .join(": "),
+          })
           return
         }
         return

packages/core/test/session-runner-tool-events.test.ts

@@ -667,3 +667,29 @@ test("content-filter finish preserves partial streamed text and never ends the s
     error: { type: "provider.content-filter" },
   })
 })
+
+test("content-filter failure explains the refusal when the provider gives a reason", async () => {
+  const { published, publisher } = capture()
+  await Effect.runPromise(publisher.publish(LLMEvent.stepStart({ index: 0 })))
+  await Effect.runPromise(
+    publisher.publish(
+      LLMEvent.stepFinish({
+        index: 0,
+        reason: {
+          normalized: "content-filter",
+          raw: "refusal",
+          category: "cyber",
+          explanation: "This request was declined because it could enable cyber harm.",
+        },
+      }),
+    ),
+  )
+  await Effect.runPromise(publisher.publishStepFailure())
+
+  expect(published.at(-1)?.data).toMatchObject({
+    error: {
+      type: "provider.content-filter",
+      message: "Provider blocked the response (cyber): This request was declined because it could enable cyber harm.",
+    },
+  })
+})