fix(provider): count only wire-surviving ephemeral turns for cache suffix

System-role ephemeral turns are hoisted into the head system block and never reach wire messages, so counting them in the suffix undercounts the split and moves the breakpoint onto q1 instead of q2. Filter role!=system and add a q1/a1/q2 + system-ephemeral regression.

Author: TheGreatAxios

src/provider/anthropic-cache-breakpoint.test.ts

@@ -41,6 +41,14 @@ function assistantTurn(text: string): ConversationTurn {
   };
 }
 
+function systemTurn(text: string): ConversationTurn {
+  return {
+    role: "system",
+    timestamp: 0,
+    content: [{ type: "text", text }],
+  };
+}
+
 type WireBlock = {
   cache_control?: unknown;
   text?: string;
@@ -121,6 +129,30 @@ describe("anthropic cache breakpoint with ephemeral turns", () => {
     ).toEqual([]);
   });
 
+  test("system-role ephemeral is hoisted, so the breakpoint stays on q2 not q1", () => {
+    const persisted = [userTurn("q1"), assistantTurn("a1"), userTurn("q2")];
+    const sysEphemeral = systemTurn("sys-nudge");
+    const request = adapters
+      .resolve(sourceFor("anthropic"))
+      .buildRequest([...persisted, sysEphemeral], "test-model", {
+        ephemeralTurns: [sysEphemeral],
+      } as InferenceOptions);
+    const body = wireBody(request.body);
+
+    // Hoisted system turns never reach wire messages, so only q1/a1/q2
+    // remain and the breakpoint must sit on the last persisted user turn.
+    expect(body.messages).toHaveLength(3);
+    expect(
+      body.messages[0]?.content.filter(
+        (block) => block.cache_control !== undefined,
+      ),
+    ).toEqual([]);
+    expect(
+      body.messages[2]?.content[body.messages[2].content.length - 1]
+        ?.cache_control,
+    ).toEqual({ type: "ephemeral" });
+  });
+
   test("without ephemeral turns the request is byte-identical to the base adapter", () => {
     const turns = [userTurn("q1"), assistantTurn("a1"), userTurn("q2")];
     const base = inner

src/provider/anthropic-cache-breakpoint.ts

@@ -23,12 +23,15 @@ type WireMessage = {
 };
 
 function ephemeralSuffixLength(options: ExtendedInferenceOptions): number {
-  // Ephemeral turns are appended as the request suffix (see reactor.ts), so
-  // every entry counts — including system-role turns, which the Anthropic
-  // wire adapter hoists out of the message list rather than leaving in place.
+  // Ephemeral turns are appended as the request suffix (see reactor.ts), but
+  // system-role turns never survive into wire messages — the Anthropic wire
+  // adapter hoists them into the head `system` block (or drops them when
+  // `systemPrompt` is set). Only turns that survive to wire messages count
+  // toward the suffix, otherwise the split undercounts and the breakpoint
+  // lands on q1 instead of the last persisted user turn q2.
   const turns = options.ephemeralTurns;
   if (turns === undefined) return 0;
-  return turns.length;
+  return turns.filter((turn) => turn.role !== "system").length;
 }
 
 function isWireBlock(block: unknown): block is WireBlock {