feat: report content-filter stops as a safety rating

README.md

@@ -114,6 +114,10 @@ Every field is optional. An absent field keeps the protocol default. Unknown key
 
 Interchange's neutral `effort` (`low | medium | high | max`) is sent as `reasoning.effort`, even when the source has no `reasoning` quirk: `low`, `medium` and `high` as given, `max` as `reasoning.maxEffort` or `high` when that is unset. A non-empty `providerOptions[reasoning.effortOption]` wins over it. `includeReasoningEffort` filters either.
 
+### Incomplete responses
+
+An incomplete response (`response.incomplete`, or a JSON body with `status: "incomplete"`) keeps its partial output and usage. When `incomplete_details.reason` is `content_filter`, an `inference.safety_rating` event with `blockReason: "content_filter"` comes before the usage.
+
 ## Using with Interchange
 
 Interchange loads custom adapters from an operator-configured `AdapterManifest`. Add one entry per provider id you serve:

src/protocol/iterator.ts

@@ -82,6 +82,21 @@ const IncompleteEvent = type({
     "usage?": ResponsesUsage,
   },
 });
+
+// A content-filter stop is a safety signal the caller must tell apart from a
+// complete or truncated turn, so it surfaces as `inference.safety_rating`.
+// Any other reason (e.g. max_output_tokens) reports usage alone.
+function incompleteReasonEvent(
+  reason: string | undefined,
+  seq: number,
+): InferenceEvent | undefined {
+  if (reason !== "content_filter") return undefined;
+  return {
+    type: "inference.safety_rating",
+    seq,
+    data: { safetyRating: { type: "safety_rating", blockReason: reason } },
+  };
+}
 const ErrorEvent = type({ "message?": "string" });
 
 // Maps the Responses API's usage object onto the internal TokenUsage,
@@ -573,6 +588,11 @@ export function parseResponse(
           parsed,
         );
       }
+      const reasonEvent = incompleteReasonEvent(
+        validated.response.incomplete_details?.reason,
+        seq,
+      );
+      if (reasonEvent !== undefined) events.push(reasonEvent);
       if (validated.response.usage !== undefined) {
         events.push({
           type: "inference.usage",
@@ -844,6 +864,11 @@ export function parseJSONResponse(
     // has nothing to do with; ignored, mirroring the SSE default case.
   }
 
+  const reasonEvent = incompleteReasonEvent(
+    response.incomplete_details?.reason,
+    seq,
+  );
+  if (reasonEvent !== undefined) events.push(reasonEvent);
   events.push({
     type: "inference.usage",
     seq,

src/responses.test.ts

@@ -412,6 +412,29 @@ describe("Responses parser — event schema validation", () => {
     expect(events.map((e) => e.type)).toEqual(["inference.usage"]);
   });
 
+  test("response.incomplete on content_filter reports a safety rating before usage", () => {
+    const adapter = createOpenAIResponsesAdapter(source, {});
+    adapter.buildRequest(turns, "model", {});
+    const incomplete = JSON.stringify({
+      type: "response.incomplete",
+      response: {
+        status: "incomplete",
+        incomplete_details: { reason: "content_filter" },
+        usage: { input_tokens: 10, output_tokens: 3 },
+      },
+    });
+    const events = adapter.parseResponse(incomplete);
+    expect(events.map((e) => e.type)).toEqual([
+      "inference.safety_rating",
+      "inference.usage",
+    ]);
+    expect(events[0]).toMatchObject({
+      data: {
+        safetyRating: { type: "safety_rating", blockReason: "content_filter" },
+      },
+    });
+  });
+
   // A function_call_arguments.delta is only routable to the tool_call.start
   // the harness already saw; one for an item_id that never arrived via
   // output_item.added is an orphan fragment, not a fresh block to
@@ -637,6 +660,27 @@ describe("Responses parser — non-streaming failure states", () => {
     expect(events.some((e) => e.type === "inference.usage")).toBe(true);
   });
 
+  test("reports a safety rating before usage on status:incomplete content_filter", () => {
+    const adapter = createOpenAIResponsesAdapter(source, {});
+    const incomplete = JSON.stringify({
+      output: [],
+      status: "incomplete",
+      error: null,
+      incomplete_details: { reason: "content_filter" },
+      usage: { input_tokens: 10, output_tokens: 0 },
+    });
+    const events = adapter.parseJSONResponse(incomplete);
+    expect(events.map((e) => e.type)).toEqual([
+      "inference.safety_rating",
+      "inference.usage",
+    ]);
+    expect(events[0]).toMatchObject({
+      data: {
+        safetyRating: { type: "safety_rating", blockReason: "content_filter" },
+      },
+    });
+  });
+
   // OpenAI sends both fields as null on every completed body.
   test("decodes a completed body with null error and incomplete_details", () => {
     const adapter = createOpenAIResponsesAdapter(source, {});