fix(sse): widen max_tokens/tool_use exemption to lone empty text blocks (#12968)

changelog.d/fixes/12968-anthropic-shim-tiny-probe-empty.md

@@ -0,0 +1 @@
+- **fix(sse):** stop misclassifying a truncated Anthropic-compatible `max_tokens` probe response (`content:[{type:"text",text:""}]`) as an empty upstream response (#12968) — thanks @pranay-gpt

open-sse/utils/diagnostics.ts

@@ -255,7 +255,9 @@ export function detectMalformedNonStream(resp: unknown): MalformedReason | null
     //  1) A block IS present but invalid (e.g. text:"", a lone "(empty response)"
     //     sentinel, or only null entries) — the model genuinely produced no
     //     usable output. That is a MALFORMED-200 empty_choices regardless of
-    //     stop_reason (parity with the OpenAI content:"" path).
+    //     stop_reason (parity with the OpenAI content:"" path) — UNLESS the
+    //     terminal stop_reason is one of the legitimate truncated-completion
+    //     exemptions below (#12968).
     //  2) `content: []` — no block at all. #9971: a truncated / non-terminal
     //     body (no stop_reason) must not become empty_choices. A terminal
     //     stop_reason with no output usually is empty_choices — except the
@@ -265,12 +267,16 @@ export function detectMalformedNonStream(resp: unknown): MalformedReason | null
     //     return content:[] + stop_reason max_tokens. Treating that as
     //     empty_choices turns a valid 200 into MALFORMED-200 → 502 even
     //     though errorClassifier would have let it through.
-    if (content.length === 0) {
-      const stopReason = typeof body.stop_reason === "string" ? body.stop_reason : "";
-      if (stopReason.length === 0) return null;
-      if (stopReason === "max_tokens" || stopReason === "tool_use") return null;
-      return "empty_choices";
-    }
+    const stopReason = typeof body.stop_reason === "string" ? body.stop_reason : "";
+    // #12968: the #9971 exemption above only fired when `content` was a
+    // completely empty array. A tiny `max_tokens` probe against an
+    // Anthropic-compatible shim can instead return content:[{type:"text",
+    // text:""}] — one block, just with no visible text — which is the exact
+    // same legitimate truncated-completion shape, so the exemption must apply
+    // whenever there is no visible output, not only when content is [].
+    if (stopReason === "max_tokens" || stopReason === "tool_use") return null;
+    // content:[] with no stop_reason at all is non-terminal, not empty (#9971).
+    if (content.length === 0 && stopReason.length === 0) return null;
     return "empty_choices";
   }
 
@@ -323,7 +329,8 @@ export function describeMalformedNonStream(
 ): { message: string; code: string; type: string } {
   const body = resp && typeof resp === "object" ? (resp as Record<string, unknown>) : null;
   if (body?.object === "response" && body.status === "failed") {
-    const err = body.error && typeof body.error === "object" ? (body.error as Record<string, unknown>) : null;
+    const err =
+      body.error && typeof body.error === "object" ? (body.error as Record<string, unknown>) : null;
     const rawMessage =
       typeof err?.message === "string" && err.message.trim().length > 0 ? err.message.trim() : null;
     return {

tests/unit/issue-12968-anthropic-shim-empty-text-block.test.ts

@@ -0,0 +1,39 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { detectMalformedNonStream } from "../../open-sse/utils/diagnostics.ts";
+
+// Exact upstream body captured in the issue (freeaiapikey.com probe response).
+const upstreamBody = {
+  content: [{ text: "", type: "text" }],
+  id: "msg_4d5e123dda0d4eda8055cd21",
+  model: "anthropic/claude-sonnet-5",
+  role: "assistant",
+  stop_reason: "max_tokens",
+  stop_sequence: null,
+  type: "message",
+  usage: {
+    cache_creation_input_tokens: 2623,
+    cache_read_input_tokens: 0,
+    input_tokens: 2234,
+    output_tokens: 1,
+  },
+};
+
+test("#12968 max_tokens probe with content:[{text:''}] must NOT be flagged empty_choices", () => {
+  const reason = detectMalformedNonStream(upstreamBody);
+  assert.equal(
+    reason,
+    null,
+    `expected legitimate truncated-probe response to pass through, got reason=${reason}`
+  );
+});
+
+test("#12968 control — content:[] + max_tokens already exempted (#9971)", () => {
+  const reason = detectMalformedNonStream({ ...upstreamBody, content: [] });
+  assert.equal(reason, null);
+});
+
+test("#12968 control — empty text block with end_turn stop_reason stays flagged", () => {
+  const reason = detectMalformedNonStream({ ...upstreamBody, stop_reason: "end_turn" });
+  assert.equal(reason, "empty_choices");
+});