fix(agents): classify Anthropic orphaned tool-use replay errors

masatohoshino · masatohoshino · commit 641a96cf347a · 2026-06-30T22:07:22.000+09:00
diff --git a/src/auto-reply/reply/provider-request-error-classifier.test.ts b/src/auto-reply/reply/provider-request-error-classifier.test.ts
@@ -113,6 +113,10 @@ describe("provider request error classifier", () => {
       "local replay invariant guard",
       "invalid_replay_transcript: OpenAI Responses replay contains dangling_tool_call toolCallId=call_1 at message index 4",
     ],
+    [
+      "Anthropic orphaned tool_use replay",
+      "messages.1: `tool_use` ids were found without `tool_result` blocks immediately after: toolu_01A09q90qw90lq917835lq9. Each `tool_use` block must have a corresponding `tool_result` block in the next message.",
+    ],
   ])("classifies %s as provider conversation-state errors", (_label, message) => {
     expect(classifyProviderRequestError(new Error(message))).toEqual({
       code: "provider_conversation_state_error",
@@ -125,6 +129,16 @@ describe("provider request error classifier", () => {
     expect(classifyProviderRequestError(new Error("400 status code (no body)"))).toBeUndefined();
   });
 
+  it("does not classify generic tool_use/tool_result mentions without the orphan signal", () => {
+    // Both block names appear but there is no "without" orphan signal, so this
+    // generic guidance text must not trip the conversation-state classifier.
+    expect(
+      classifyProviderRequestError(
+        new Error("Each tool_use block must have a corresponding tool_result block."),
+      ),
+    ).toBeUndefined();
+  });
+
   it("leaves explicit HTTP 429 rate-limit failures on the existing rate-limit path", () => {
     expect(classifyProviderRequestError(new Error("429: rate limit exceeded"))).toBeUndefined();
   });
diff --git a/src/auto-reply/reply/provider-request-error-classifier.ts b/src/auto-reply/reply/provider-request-error-classifier.ts
@@ -105,6 +105,10 @@ export function isProviderConversationStateErrorMessage(message: string): boolea
       lower.includes("tooluse") &&
       lower.includes("exceeds the number") &&
       lower.includes("previous turn")) ||
+    // Anthropic/Bedrock orphaned tool-call replay: "`tool_use` ids were found
+    // without `tool_result` blocks immediately after: ...". Same broken-turn
+    // shape as the toolResult/toolUse count mismatch above, just snake_case.
+    (lower.includes("tool_use") && lower.includes("tool_result") && lower.includes("without")) ||
     lower.includes("function call turn comes immediately after") ||
     lower.includes("incorrect role information") ||
     lower.includes("roles must alternate") ||
