Prevent xAI replay from exposing persisted Anthropic reasoning

The xAI Responses path reuses convertToResponsesApiInput() when replaying
persisted conversation history. Anthropic thinking blocks were being
flattened into ordinary assistant-visible output_text entries, which
changes conversation semantics and leaks hidden reasoning into later
prompt context.

This keeps the fix intentionally narrow: persisted Anthropic thinking
blocks are now skipped for Responses replay, and a regression test
covers the behavior.

Constraint: Keep the change scoped to the replay transform without changing provider selection or task persistence
Rejected: Preserve thinking by converting it to [Thinking] output_text | turns hidden reasoning into visible assistant content
Confidence: medium
Scope-risk: narrow
Reversibility: clean
Directive: If a provider-specific hidden reasoning replay format is added later, prefer that over flattening Anthropic thinking into visible text
Tested: pnpm --dir src exec vitest run api/transform/__tests__/responses-api-input.spec.ts api/providers/__tests__/xai.spec.ts
Not-tested: Full end-to-end provider-switch replay with live xAI credentials

src/api/transform/__tests__/responses-api-input.spec.ts

@@ -280,6 +280,28 @@ describe("convertToResponsesApiInput", () => {
 				}),
 			)
 		})
+
+		it("should not replay Anthropic thinking blocks as assistant-visible text", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				{
+					role: "assistant",
+					content: [
+						{ type: "thinking", thinking: "SECRET_CHAIN_OF_THOUGHT", signature: "sig-1" } as any,
+						{ type: "text", text: "visible answer" },
+					],
+				},
+			]
+
+			const result = convertToResponsesApiInput(messages)
+
+			expect(result).toEqual([
+				{
+					type: "message",
+					role: "assistant",
+					content: [{ type: "output_text", text: "visible answer" }],
+				},
+			])
+		})
 	})
 
 	describe("multi-turn conversations", () => {

src/api/transform/responses-api-input.ts

@@ -53,14 +53,11 @@ export function convertToResponsesApiInput(messages: Anthropic.Messages.MessageP
 						})
 						break
 					case "thinking":
-						// Include reasoning if it has content
-						if ((part as any).thinking && (part as any).thinking.trim().length > 0) {
-							input.push({
-								type: "message",
-								role: "assistant",
-								content: [{ type: "output_text", text: `[Thinking] ${(part as any).thinking}` }],
-							})
-						}
+						// Anthropic thinking blocks represent hidden reasoning. The
+						// Responses API input format does not have a compatible hidden
+						// reasoning representation for these persisted blocks, so replaying
+						// them as normal assistant-visible text changes conversation
+						// semantics. Skip them instead of flattening them into output_text.
 						break
 				}
 			}