fix(translator): map non-streaming Claude stop reasons the way the streaming path does

changelog.d/fixes/16544-claude-nonstream-stop-reasons.md

@@ -0,0 +1 @@
+- **fix(translator):** a non-streaming Claude response now maps `refusal`, `model_context_window_exceeded`, `stop_sequence` and `pause_turn` the same way the streaming path does, so a Gemini client sees a refusal as `SAFETY` and a context-window stop as `MAX_TOKENS` instead of a clean `STOP`.

open-sse/handlers/responseTranslator.ts

@@ -647,13 +647,11 @@ export function translateNonStreamingResponse(
         message.content = "";
       }
 
-      let finishReason = toString(root.stop_reason, "stop");
-      if (finishReason === "end_turn") finishReason = "stop";
-      if (finishReason === "tool_use") finishReason = "tool_calls";
-      // Streaming claude-to-openai already maps this (convertStopReason). Leaving
-      // the raw spelling here makes a truncated completion look like an unknown
-      // finish to the chat empty-output check.
-      if (finishReason === "max_tokens") finishReason = "length";
+      // Same mapping as the streaming claude-to-openai path (convertStopReason):
+      // refusal -> content_filter, max_tokens and model_context_window_exceeded ->
+      // length, stop_sequence and pause_turn -> stop. A raw spelling here reached
+      // non-OpenAI clients unmapped (a Gemini client saw a refusal as STOP).
+      const finishReason = normalizeOpenAICompatibleFinishReasonString(root.stop_reason, "stop");
 
       const result: JsonRecord = {
         id: `chatcmpl-${toString(root.id, String(Date.now()))}`,

tests/unit/claude-nonstream-stop-reasons.test.ts

@@ -0,0 +1,61 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const { translateNonStreamingResponse } = await import("../../open-sse/handlers/responseTranslator.ts");
+const { FORMATS } = await import("../../open-sse/translator/formats.ts");
+
+function claudeBody(stopReason: string) {
+  return {
+    id: "msg_stop",
+    type: "message",
+    role: "assistant",
+    model: "claude-test",
+    content: [{ type: "text", text: "partial" }],
+    stop_reason: stopReason,
+    usage: { input_tokens: 3, output_tokens: 2 },
+  };
+}
+
+function openaiFinish(stopReason: string): unknown {
+  const out = translateNonStreamingResponse(claudeBody(stopReason), FORMATS.CLAUDE, FORMATS.OPENAI) as {
+    choices?: Array<{ finish_reason?: unknown }>;
+  };
+  return out.choices?.[0]?.finish_reason;
+}
+
+function geminiFinish(stopReason: string): unknown {
+  type GeminiBody = { candidates?: Array<{ finishReason?: unknown }> };
+  const out = translateNonStreamingResponse(claudeBody(stopReason), FORMATS.CLAUDE, FORMATS.GEMINI) as GeminiBody & {
+    response?: GeminiBody;
+  };
+  return (out.candidates ?? out.response?.candidates)?.[0]?.finishReason;
+}
+
+test("non-streaming Claude refusal reaches an OpenAI-shaped body as content_filter", () => {
+  assert.equal(openaiFinish("refusal"), "content_filter");
+});
+
+test("non-streaming Claude context-window stop reaches an OpenAI-shaped body as length", () => {
+  assert.equal(openaiFinish("model_context_window_exceeded"), "length");
+});
+
+test("non-streaming Claude stop_sequence and pause_turn reach an OpenAI-shaped body as stop", () => {
+  assert.equal(openaiFinish("stop_sequence"), "stop");
+  assert.equal(openaiFinish("pause_turn"), "stop");
+});
+
+test("a Gemini client sees a Claude refusal as SAFETY, not as a clean STOP", () => {
+  assert.equal(geminiFinish("refusal"), "SAFETY");
+});
+
+test("a Gemini client sees a Claude context-window stop as MAX_TOKENS", () => {
+  assert.equal(geminiFinish("model_context_window_exceeded"), "MAX_TOKENS");
+});
+
+test("existing non-streaming mappings are unchanged", () => {
+  assert.equal(openaiFinish("end_turn"), "stop");
+  assert.equal(openaiFinish("tool_use"), "tool_calls");
+  assert.equal(openaiFinish("max_tokens"), "length");
+  assert.equal(geminiFinish("end_turn"), "STOP");
+  assert.equal(geminiFinish("max_tokens"), "MAX_TOKENS");
+});