fix(translator): prevent doubled tool args in OpenAI-to-Claude (#5828)

Integrated into release/v3.8.43

CHANGELOG.md

@@ -110,6 +110,8 @@
 
 - **fix(kiro):** stop injecting a placeholder user turn on trailing tool-result turns so agentic loops aren't disrupted. (thanks @jetmiky)
 
+- **fix(translator):** prevent doubled tool arguments in OpenAI-to-Claude responses (duplicate finish_reason guard + string tool-input passthrough). (thanks @vishalrajv)
+
 ### 📝 Maintenance
 
 ---

open-sse/translator/request/claude-to-openai.ts

@@ -410,7 +410,10 @@ function convertClaudeMessage(msg, preserveCacheControl = false) {
             type: "function",
             function: {
               name: block.name,
-              arguments: JSON.stringify(block.input || {}),
+              arguments:
+                typeof block.input === "string"
+                  ? block.input
+                  : JSON.stringify(block.input || {}),
             },
           });
           break;

open-sse/translator/response/openai-to-claude.ts

@@ -225,8 +225,8 @@ export function openaiToClaudeResponse(chunk, state) {
     }
   }
 
-  // Finish
-  if (choice.finish_reason) {
+  // Finish — guard against duplicate finish_reason chunks (common with OpenAI-compatible models)
+  if (choice.finish_reason && !state.finishReason) {
     stopThinkingBlock(state, results);
     stopTextBlock(state, results);
 

tests/unit/translator-claude-to-openai.test.ts

@@ -259,6 +259,42 @@ test("Claude -> OpenAI turns thinking and tool_use blocks into assistant tool_ca
   });
 });
 
+test("Claude -> OpenAI passes pre-stringified tool_use input through verbatim (no double-encoding)", () => {
+  // #2279: some upstream Claude sources emit tool_use.input already JSON-encoded
+  // as a string. Re-running JSON.stringify on it would double-encode the
+  // payload (wrapping it in an extra pair of quotes with escaped internals).
+  const result = claudeToOpenAIRequest(
+    "gpt-4o",
+    {
+      messages: [
+        {
+          role: "assistant",
+          content: [
+            {
+              type: "tool_use",
+              id: "tu_2",
+              name: "read_file",
+              input: '{"path":"/tmp/demo"}',
+            },
+          ],
+        },
+      ],
+    },
+    false
+  );
+
+  assert.deepEqual(result.messages[0].tool_calls, [
+    {
+      id: "tu_2",
+      type: "function",
+      function: {
+        name: "read_file",
+        arguments: '{"path":"/tmp/demo"}',
+      },
+    },
+  ]);
+});
+
 test("Claude -> OpenAI converts tool_result blocks into tool messages and preserves trailing user text", () => {
   const result = claudeToOpenAIRequest(
     "gpt-4o",

tests/unit/translator-resp-openai-to-claude.test.ts

@@ -148,6 +148,53 @@ test("OpenAI stream: tool calls strip Claude OAuth prefix and keep cache usage",
   assert.equal(result[5].usage.cache_creation_input_tokens, 1);
 });
 
+test("OpenAI stream: two finish_reason chunks emit finish events exactly once", () => {
+  // #2279: some OpenAI-compatible upstreams resend a terminal chunk that still
+  // carries finish_reason (e.g. an empty-delta echo of the finish chunk). Without
+  // a guard, the whole finish block (content_block_stop / message_delta /
+  // message_stop) re-runs and duplicates those events downstream.
+  const state = createState();
+  const first = openaiToClaudeResponse(
+    {
+      id: "chatcmpl-4",
+      model: "gpt-4.1",
+      choices: [
+        {
+          index: 0,
+          delta: {
+            tool_calls: [
+              {
+                index: 0,
+                id: "call_1",
+                type: "function",
+                function: { name: "read_file", arguments: '{"path":"/tmp/a"}' },
+              },
+            ],
+          },
+          finish_reason: "tool_calls",
+        },
+      ],
+      usage: { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 },
+    },
+    state
+  );
+  // Duplicate terminal chunk: empty delta, same finish_reason.
+  const second = openaiToClaudeResponse(
+    {
+      id: "chatcmpl-4",
+      model: "gpt-4.1",
+      choices: [{ index: 0, delta: {}, finish_reason: "tool_calls" }],
+      usage: { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 },
+    },
+    state
+  );
+  const result = flatten([first, second]);
+
+  assert.equal(result.filter((e) => e.type === "content_block_stop").length, 1);
+  assert.equal(result.filter((e) => e.type === "message_delta").length, 1);
+  assert.equal(result.filter((e) => e.type === "message_stop").length, 1);
+});
+
 test("OpenAI non-stream: chat completion becomes Claude message with thinking and tool_use", () => {
   const result = translateNonStreamingResponse(
     {