agents: drop Anthropic thinking blocks on replay

jmcte · jmcte · commit 18985e31b00a · 2026-03-12T18:02:43.000-05:00
diff --git a/src/agents/pi-embedded-runner.sanitize-session-history.test.ts b/src/agents/pi-embedded-runner.sanitize-session-history.test.ts
@@ -52,6 +52,20 @@ describe("sanitizeSessionHistory", () => {
       sessionId: TEST_SESSION_ID,
     });
 
+  const sanitizeAnthropicHistory = async (params: {
+    messages: AgentMessage[];
+    provider?: string;
+    modelId?: string;
+  }) =>
+    sanitizeSessionHistory({
+      messages: params.messages,
+      modelApi: "anthropic-messages",
+      provider: params.provider ?? "anthropic",
+      modelId: params.modelId ?? "claude-opus-4-6",
+      sessionManager: makeMockSessionManager(),
+      sessionId: TEST_SESSION_ID,
+    });
+
   const getAssistantMessage = (messages: AgentMessage[]) => {
     expect(messages[1]?.role).toBe("assistant");
     return messages[1] as Extract<AgentMessage, { role: "assistant" }>;
@@ -776,22 +790,15 @@ describe("sanitizeSessionHistory", () => {
     expect(types).not.toContain("thinking");
   });
 
-  it("does not drop thinking blocks for non-copilot providers", async () => {
+  it("drops assistant thinking blocks for anthropic replay", async () => {
     setNonGoogleModelApi();
 
     const messages = makeThinkingAndTextAssistantMessages();
 
-    const result = await sanitizeSessionHistory({
-      messages,
-      modelApi: "anthropic-messages",
-      provider: "anthropic",
-      modelId: "claude-opus-4-6",
-      sessionManager: makeMockSessionManager(),
-      sessionId: TEST_SESSION_ID,
-    });
+    const result = await sanitizeAnthropicHistory({ messages });
 
-    const types = getAssistantContentTypes(result);
-    expect(types).toContain("thinking");
+    const assistant = getAssistantMessage(result);
+    expect(assistant.content).toEqual([{ type: "text", text: "hi" }]);
   });
 
   it("does not drop thinking blocks for non-claude copilot models", async () => {
diff --git a/src/agents/pi-embedded-runner/run/attempt.ts b/src/agents/pi-embedded-runner/run/attempt.ts
@@ -1944,9 +1944,10 @@ export async function runEmbeddedAttempt(
         activeSession.agent.streamFn = cacheTrace.wrapStreamFn(activeSession.agent.streamFn);
       }
 
-      // Copilot/Claude can reject persisted `thinking` blocks (e.g. thinkingSignature:"reasoning_text")
-      // on *any* follow-up provider call (including tool continuations). Wrap the stream function
-      // so every outbound request sees sanitized messages.
+      // Anthropic Claude endpoints can reject replayed `thinking` blocks
+      // (e.g. thinkingSignature:"reasoning_text") on any follow-up provider
+      // call, including tool continuations. Wrap the stream function so every
+      // outbound request sees sanitized messages.
       if (transcriptPolicy.dropThinkingBlocks) {
         const inner = activeSession.agent.streamFn;
         activeSession.agent.streamFn = (model, context, options) => {
diff --git a/src/agents/provider-capabilities.ts b/src/agents/provider-capabilities.ts
@@ -29,9 +29,11 @@ const DEFAULT_PROVIDER_CAPABILITIES: ProviderCapabilities = {
 const PROVIDER_CAPABILITIES: Record<string, Partial<ProviderCapabilities>> = {
   anthropic: {
     providerFamily: "anthropic",
+    dropThinkingBlockModelHints: ["claude"],
   },
   "amazon-bedrock": {
     providerFamily: "anthropic",
+    dropThinkingBlockModelHints: ["claude"],
   },
   // kimi-coding natively supports Anthropic tool framing (input_schema);
   // converting to OpenAI format causes XML text fallback instead of tool_use blocks.
diff --git a/src/agents/transcript-policy.ts b/src/agents/transcript-policy.ts
@@ -80,9 +80,9 @@ export function resolveTranscriptPolicy(params: {
     });
   const requiresOpenAiCompatibleToolIdSanitization = params.modelApi === "openai-completions";
 
-  // GitHub Copilot's Claude endpoints can reject persisted `thinking` blocks with
-  // non-binary/non-base64 signatures (e.g. thinkingSignature: "reasoning_text").
-  // Drop these blocks at send-time to keep sessions usable.
+  // Anthropic Claude endpoints can reject replayed `thinking` blocks unless the
+  // original signatures are preserved byte-for-byte. Drop them at send-time to
+  // keep persisted sessions usable across follow-up turns.
   const dropThinkingBlocks = shouldDropThinkingBlocksForModel({ provider, modelId });
 
   const needsNonImageSanitize =
