diff --git a/changelog.d/fixes/9436-claude-system-role-cache-boundary.md b/changelog.d/fixes/9436-claude-system-role-cache-boundary.md
new file mode 100644
index 0000000000..c2433f470a
--- /dev/null
+++ b/changelog.d/fixes/9436-claude-system-role-cache-boundary.md
@@ -0,0 +1 @@
+- **fix(sse):** hoisting a mid-conversation `system`/`developer` message into the top-level `system` field no longer carries its `cache_control` marker along, which left the conversation history without a cache breakpoint and forced a full re-read plus a rebuild on the next turn. The boundary is moved to the nearest preceding block that can carry one, and now survives the rewrites that turn `tool_result` and inlined file/document blocks into plain text; if the target block is already marked, both markers are kept unless Anthropic's TTL ordering forbids it. Both hoisting paths are fixed — `extractSystemRoleMessages` and `extractSystemMessagesToBody`. Regression guard: `tests/unit/claude-system-role-cache-boundary.test.ts`. ([#9436](https://github.com/diegosouzapw/OmniRoute/issues/9436))
diff --git a/open-sse/handlers/chatCore/claudeSystemRole.ts b/open-sse/handlers/chatCore/claudeSystemRole.ts
index 0c22180166..a661847f29 100644
--- a/open-sse/handlers/chatCore/claudeSystemRole.ts
+++ b/open-sse/handlers/chatCore/claudeSystemRole.ts
@@ -7,8 +7,96 @@
  * chat role, so they must be hoisted. `developer` is OpenAI's Responses-API rename of `system` and
  * is treated identically. Mutates the payload in place; behaviour is byte-identical to the previous
  * top-level definition (still re-exported from chatCore.ts for existing importers/tests).
+ *
+ * `relocateHoistedCacheBoundary` keeps that hoist from destroying the client's prompt-cache
+ * layout (#9436); both hoisting implementations share it.
  */
 
+export type HoistedCacheBoundary = "moved" | "kept" | "dropped";
+
+/** Effective cache TTL of a `cache_control` value; Anthropic defaults to 5m when `ttl` is absent. */
+function effectiveTtl(marker: unknown): string {
+  const ttl = (marker as Record<string, unknown> | null | undefined)?.ttl;
+  return typeof ttl === "string" ? ttl : "5m";
+}
+
+/**
+ * Whether a content block can carry a cache breakpoint. Excludes blocks Anthropic does not accept
+ * as one (thinking) and blocks the upstream normalisation discards or empties out anyway.
+ */
+function isCacheBreakpointTarget(block: unknown): block is Record<string, unknown> {
+  if (block === null || typeof block !== "object") return false;
+  const candidate = block as Record<string, unknown>;
+  switch (candidate.type) {
+    case "text":
+      // Empty text blocks are stripped before the payload goes upstream.
+      return typeof candidate.text === "string" && candidate.text.length > 0;
+    case "tool_use":
+    case "image":
+    case "image_url":
+    case "file":
+    case "file_url":
+    case "document":
+      return true;
+    case "tool_result": {
+      // A tool_result that yields no text collapses to nothing during normalisation.
+      const payload = candidate.content ?? candidate.text ?? candidate.output;
+      if (typeof payload === "string") return payload.length > 0;
+      if (Array.isArray(payload)) {
+        // Only the non-empty text parts of the array survive; images and unknown parts do not.
+        return payload.some((part) => {
+          const text = (part as Record<string, unknown> | null)?.text;
+          return (
+            (part as Record<string, unknown> | null)?.type === "text" &&
+            typeof text === "string" &&
+            text.length > 0
+          );
+        });
+      }
+      return payload != null;
+    }
+    default:
+      // thinking, redacted_thinking, and anything unrecognised.
+      return false;
+  }
+}
+
+/**
+ * Preserves a message-level cache boundary when a marked system/developer block is hoisted into
+ * top-level `system[]`.
+ *
+ * The marker is moved to the nearest preceding block that can carry a breakpoint. If that block is
+ * already marked, both are kept — except where the hoisted marker, which ends up ahead of the
+ * target in `system[]`, would put a 5m breakpoint before a 1h one; Anthropic requires the longer
+ * TTL first, so the hoisted marker is dropped instead.
+ *
+ * @returns `"moved"` or `"dropped"` — the caller must remove the marker from the hoisted block;
+ *          `"kept"` — the marker stays on it
+ */
+export function relocateHoistedCacheBoundary(
+  marker: unknown,
+  preceding: ReadonlyArray<{ content?: unknown }>
+): HoistedCacheBoundary {
+  for (let i = preceding.length - 1; i >= 0; i--) {
+    const content = preceding[i]?.content;
+    if (!Array.isArray(content)) continue;
+    for (let j = content.length - 1; j >= 0; j--) {
+      const block = content[j];
+      if (!isCacheBreakpointTarget(block)) continue;
+      if (block.cache_control == null) {
+        block.cache_control = marker;
+        return "moved";
+      }
+      // Occupied: overwriting would discard the client's own marker, and stepping further back
+      // would only shorten the prefix — so both stay, unless the TTL order forbids it.
+      return effectiveTtl(marker) === "5m" && effectiveTtl(block.cache_control) === "1h"
+        ? "dropped"
+        : "kept";
+    }
+  }
+  return "kept";
+}
+
 export function extractSystemRoleMessages(payload: Record<string, unknown>): void {
   if (!Array.isArray(payload.messages)) return;
   const messages = payload.messages as Array<{ role?: unknown; content?: unknown }>;
@@ -23,13 +111,27 @@ export function extractSystemRoleMessages(payload: Record<string, unknown>): voi
   if (systemMessages.length === 0) return;
 
   const extraBlocks: Array<Record<string, unknown>> = [];
-  for (const sm of systemMessages) {
+  // Walk in order rather than over the filtered list: re-anchoring a hoisted `cache_control`
+  // needs the messages that precede it and stay behind (#9436).
+  const preceding: Array<{ content?: unknown }> = [];
+  for (const sm of messages) {
+    if (!isSystemRole(sm.role)) {
+      preceding.push(sm);
+      continue;
+    }
     if (typeof sm.content === "string" && sm.content.length > 0) {
       extraBlocks.push({ type: "text", text: sm.content });
     } else if (Array.isArray(sm.content)) {
       for (const block of sm.content as Array<Record<string, unknown>>) {
         if (block?.type === "text" && typeof block.text === "string" && block.text.length > 0) {
-          extraBlocks.push({ ...block });
+          const hoisted = { ...block };
+          if (
+            hoisted.cache_control != null &&
+            relocateHoistedCacheBoundary(hoisted.cache_control, preceding) !== "kept"
+          ) {
+            delete hoisted.cache_control;
+          }
+          extraBlocks.push(hoisted);
         }
       }
     }
diff --git a/open-sse/handlers/chatCore/claudeUpstreamMessages.ts b/open-sse/handlers/chatCore/claudeUpstreamMessages.ts
index b34ff7b36d..3202ed6502 100644
--- a/open-sse/handlers/chatCore/claudeUpstreamMessages.ts
+++ b/open-sse/handlers/chatCore/claudeUpstreamMessages.ts
@@ -12,27 +12,58 @@
  */
 
 import type { ClaudeContentBlock, ClaudeMessage } from "./claudeMessageTypes.ts";
-import { extractSystemRoleMessages } from "./claudeSystemRole.ts";
+import { extractSystemRoleMessages, relocateHoistedCacheBoundary } from "./claudeSystemRole.ts";
 import { splitMisplacedToolResults } from "../../translator/helpers/claudeHelper.ts";
 
 type LoggerLike = { debug?: (...args: unknown[]) => void } | null | undefined;
 
+/**
+ * Carries a replaced block's `cache_control` onto its substitute. Rewriting a marked block into a
+ * plain text block would otherwise drop the breakpoint the client (or the #9436 hoist) put there.
+ */
+function withCacheControl(
+  replacement: ClaudeContentBlock,
+  original: ClaudeContentBlock
+): ClaudeContentBlock {
+  if (original.cache_control != null) replacement.cache_control = original.cache_control;
+  return replacement;
+}
+
 export function extractSystemMessagesToBody(payload: Record<string, unknown>) {
   if (!Array.isArray(payload.messages)) return;
   const messages = payload.messages as ClaudeMessage[];
-  const systemMessages = messages.filter((m) => {
-    const role = String(m.role || "").toLowerCase();
-    return role === "system" || role === "developer";
-  });
+  const isSystemRole = (role: unknown): boolean => {
+    const normalized = String(role || "").toLowerCase();
+    return normalized === "system" || normalized === "developer";
+  };
+  const systemMessages = messages.filter((m) => isSystemRole(m.role));
   if (systemMessages.length === 0) return;
   const extraBlocks: ClaudeContentBlock[] = [];
-  for (const sm of systemMessages) {
+  // Same in-order walk as extractSystemRoleMessages: re-anchoring a hoisted `cache_control`
+  // needs the messages that precede it and stay behind (#9436).
+  const preceding: ClaudeMessage[] = [];
+  for (const sm of messages) {
+    if (!isSystemRole(sm.role)) {
+      preceding.push(sm);
+      continue;
+    }
     if (typeof sm.content === "string" && sm.content.length > 0) {
       extraBlocks.push({ type: "text", text: sm.content });
     } else if (Array.isArray(sm.content)) {
       for (const block of sm.content as ClaudeContentBlock[]) {
         if (block?.type === "text" && typeof block.text === "string" && block.text.length > 0) {
-          extraBlocks.push(block);
+          // Blocks are pushed by reference here (the sibling implementation spreads them), so
+          // only a block whose marker actually moves is copied.
+          if (
+            block.cache_control != null &&
+            relocateHoistedCacheBoundary(block.cache_control, preceding) !== "kept"
+          ) {
+            const withoutMarker: ClaudeContentBlock = { ...block };
+            delete withoutMarker.cache_control;
+            extraBlocks.push(withoutMarker);
+          } else {
+            extraBlocks.push(block);
+          }
         }
       }
     }
@@ -47,10 +78,7 @@ export function extractSystemMessagesToBody(payload: Record<string, unknown>) {
       payload.system = extraBlocks;
     }
   }
-  payload.messages = messages.filter((m) => {
-    const role = String(m.role || "").toLowerCase();
-    return role !== "system" && role !== "developer";
-  });
+  payload.messages = messages.filter((m) => !isSystemRole(m.role));
 }
 
 export function normalizeClaudeUpstreamMessages(
@@ -104,7 +132,7 @@ export function normalizeClaudeUpstreamMessages(
           const fileName =
             (block.file as Record<string, unknown>)?.name ?? block.name ?? "attachment";
           if (typeof fileContent === "string" && fileContent.length > 0) {
-            return [{ type: "text", text: `[${fileName}]\n${fileContent}` }];
+            return [withCacheControl({ type: "text", text: `[${fileName}]\n${fileContent}` }, block)];
           }
         }
         return [block];
@@ -126,7 +154,9 @@ export function normalizeClaudeUpstreamMessages(
                   .join("\n")
               : JSON.stringify(resultContent);
         if (resultText.length > 0) {
-          return [{ type: "text", text: `[Tool Result: ${toolId}]\n${resultText}` }];
+          return [
+            withCacheControl({ type: "text", text: `[Tool Result: ${toolId}]\n${resultText}` }, block),
+          ];
         }
         return [];
       }
diff --git a/tests/unit/claude-system-role-cache-boundary.test.ts b/tests/unit/claude-system-role-cache-boundary.test.ts
new file mode 100644
index 0000000000..6d5a12d609
--- /dev/null
+++ b/tests/unit/claude-system-role-cache-boundary.test.ts
@@ -0,0 +1,409 @@
+// tests/unit/claude-system-role-cache-boundary.test.ts
+// Regression coverage for #9436.
+// Hoisting a marked system/developer block must not remove the effective cache boundary from
+// messages[]. Covers both hoisting implementations plus the native Claude path.
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { extractSystemRoleMessages } from "../../open-sse/handlers/chatCore/claudeSystemRole.ts";
+import {
+  extractSystemMessagesToBody,
+  normalizeClaudeUpstreamMessages,
+} from "../../open-sse/handlers/chatCore/claudeUpstreamMessages.ts";
+
+/** Breakpoint layout in the same terms the incident telemetry reports it: sys=N, msg=N. */
+function layout(payload: Record<string, unknown>): { sys: number; msg: number } {
+  const marked = (blocks: unknown): number =>
+    Array.isArray(blocks)
+      ? blocks.filter(
+          (b) => b !== null && typeof b === "object" && (b as Record<string, unknown>).cache_control != null
+        ).length
+      : 0;
+  const messages = (Array.isArray(payload.messages) ? payload.messages : []) as Array<{
+    content?: unknown;
+  }>;
+  return {
+    sys: marked(payload.system),
+    msg: messages.reduce((n, m) => n + marked(m.content), 0),
+  };
+}
+
+/**
+ * A Claude Code shaped turn: a cached system prompt, an accumulated mid-conversation system note,
+ * and the second breakpoint the client advances each round. `markerOn` selects where that second
+ * breakpoint sits — on the system note (the reported defect) or on the newest tool_result (what the
+ * healthy calls in the incident window show).
+ */
+function claudeCodeTurn(markerOn: "systemNote" | "toolResult"): Record<string, unknown> {
+  const note: Record<string, unknown> = { type: "text", text: "mid-conversation system note" };
+  const toolResult: Record<string, unknown> = {
+    type: "tool_result",
+    tool_use_id: "toolu_1",
+    content: "ok",
+  };
+  (markerOn === "systemNote" ? note : toolResult).cache_control = { type: "ephemeral" };
+
+  return {
+    system: [{ type: "text", text: "base system prompt", cache_control: { type: "ephemeral" } }],
+    messages: [
+      { role: "user", content: [{ type: "text", text: "turn 1" }] },
+      { role: "assistant", content: [{ type: "text", text: "reply 1" }] },
+      { role: "system", content: [note] },
+      { role: "user", content: [toolResult] },
+    ],
+  };
+}
+
+test("#9436 hoisting a marked system note keeps a cache boundary inside messages", () => {
+  const payload = claudeCodeTurn("systemNote");
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+
+  extractSystemRoleMessages(payload);
+
+  // Before the fix this was { sys: 2, msg: 0 } — the production signature behind 99.4 % of all
+  // uncached input tokens in the incident window.
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+});
+
+test("#9436 the boundary lands before the hoisted note, never after it", () => {
+  const payload = claudeCodeTurn("systemNote");
+  extractSystemRoleMessages(payload);
+
+  const messages = payload.messages as Array<{ content: Array<Record<string, unknown>> }>;
+  assert.equal(messages.length, 3);
+  // The assistant reply directly preceded the system note — same cut point in the reordered prefix.
+  assert.deepEqual(messages[1].content[0].cache_control, { type: "ephemeral" });
+  // Moving it onto the newer tool_result would cache content the client never marked.
+  assert.equal(messages[2].content[0].cache_control, undefined);
+});
+
+test("#9436 the marker does not ride along into the system array", () => {
+  const payload = claudeCodeTurn("systemNote");
+  extractSystemRoleMessages(payload);
+
+  const system = payload.system as Array<Record<string, unknown>>;
+  assert.equal(system.length, 2); // merged, not dropped — #7293
+  assert.equal(system[1].text, "mid-conversation system note");
+  assert.equal(system[1].cache_control, undefined);
+});
+
+test("hoisting still lifts system/developer content out of messages (#7293)", () => {
+  const payload = claudeCodeTurn("toolResult");
+  extractSystemRoleMessages(payload);
+
+  const messages = payload.messages as Array<{ role: string }>;
+  assert.ok(!messages.some((m) => m.role === "system"));
+  assert.equal((payload.system as unknown[]).length, 2);
+});
+
+test("a healthy cache-boundary layout remains unchanged", () => {
+  const payload = claudeCodeTurn("toolResult");
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+
+  extractSystemRoleMessages(payload);
+
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+  const messages = payload.messages as Array<{ content: Array<Record<string, unknown>> }>;
+  assert.deepEqual(messages[2].content[0].cache_control, { type: "ephemeral" }); // still the tool_result
+  assert.equal(messages[1].content[0].cache_control, undefined); // no marker added
+});
+
+/** A turn whose re-anchor target is already marked, so the two TTLs decide the outcome. */
+function occupiedTarget(targetTtl: string, hoistedTtl: string): Record<string, unknown> {
+  return {
+    messages: [
+      {
+        role: "user",
+        content: [
+          { type: "text", text: "turn 1", cache_control: { type: "ephemeral", ttl: targetTtl } },
+        ],
+      },
+      {
+        role: "system",
+        content: [
+          { type: "text", text: "note", cache_control: { type: "ephemeral", ttl: hoistedTtl } },
+        ],
+      },
+      { role: "user", content: [{ type: "text", text: "turn 2" }] },
+    ],
+  };
+}
+
+/** Reads back both markers of an `occupiedTarget` payload after hoisting. */
+function markers(payload: Record<string, unknown>): { target: unknown; hoisted: unknown } {
+  const messages = payload.messages as Array<{ content: Array<Record<string, unknown>> }>;
+  const system = payload.system as Array<Record<string, unknown>>;
+  return { target: messages[0].content[0].cache_control, hoisted: system[0].cache_control };
+}
+
+test("occupied 1h target, hoisted 5m marker: only the target marker survives", () => {
+  // Keeping both would leave a 5m breakpoint in system[] ahead of the 1h one in messages[];
+  // Anthropic requires the longer TTL first.
+  const payload = occupiedTarget("1h", "5m");
+  extractSystemRoleMessages(payload);
+
+  assert.deepEqual(markers(payload).target, { type: "ephemeral", ttl: "1h" });
+  assert.equal(markers(payload).hoisted, undefined);
+  assert.deepEqual(layout(payload), { sys: 0, msg: 1 });
+});
+
+test("occupied 5m target, hoisted 1h marker: both markers survive", () => {
+  const payload = occupiedTarget("5m", "1h");
+  extractSystemRoleMessages(payload);
+
+  assert.deepEqual(markers(payload).target, { type: "ephemeral", ttl: "5m" });
+  assert.deepEqual(markers(payload).hoisted, { type: "ephemeral", ttl: "1h" });
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+});
+
+test("occupied target, both markers 5m: both survive", () => {
+  const payload = occupiedTarget("5m", "5m");
+  extractSystemRoleMessages(payload);
+
+  assert.deepEqual(markers(payload).target, { type: "ephemeral", ttl: "5m" });
+  assert.deepEqual(markers(payload).hoisted, { type: "ephemeral", ttl: "5m" });
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+});
+
+test("occupied target, both markers 1h: both survive", () => {
+  const payload = occupiedTarget("1h", "1h");
+  extractSystemRoleMessages(payload);
+
+  assert.deepEqual(markers(payload).target, { type: "ephemeral", ttl: "1h" });
+  assert.deepEqual(markers(payload).hoisted, { type: "ephemeral", ttl: "1h" });
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+});
+
+test("two consecutive marked system messages compete for the same target", () => {
+  // No message between them, so both would re-anchor onto the same block. The first takes it;
+  // the second keeps its marker rather than overwriting the first.
+  const payload: Record<string, unknown> = {
+    messages: [
+      { role: "user", content: [{ type: "text", text: "turn 1" }] },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note a", cache_control: { type: "ephemeral", ttl: "5m" } }],
+      },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note b", cache_control: { type: "ephemeral", ttl: "1h" } }],
+      },
+      { role: "user", content: [{ type: "text", text: "turn 2" }] },
+    ],
+  };
+
+  extractSystemRoleMessages(payload);
+
+  const messages = payload.messages as Array<{ content: Array<Record<string, unknown>> }>;
+  const system = payload.system as Array<Record<string, unknown>>;
+  assert.deepEqual(messages[0].content[0].cache_control, { type: "ephemeral", ttl: "5m" });
+  assert.equal(system[0].cache_control, undefined); // note a — moved
+  assert.deepEqual(system[1].cache_control, { type: "ephemeral", ttl: "1h" }); // note b — kept
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 }); // two in, two out
+});
+
+test("several marked system messages keep one boundary each, without adding any", () => {
+  // The `sys=3,msg=0` shape from the incident window: more than one hoisted block carries a marker.
+  const payload: Record<string, unknown> = {
+    system: [{ type: "text", text: "base", cache_control: { type: "ephemeral" } }],
+    messages: [
+      { role: "user", content: [{ type: "text", text: "turn 1" }] },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note a", cache_control: { type: "ephemeral" } }],
+      },
+      { role: "assistant", content: [{ type: "text", text: "reply 1" }] },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note b", cache_control: { type: "ephemeral" } }],
+      },
+      { role: "user", content: [{ type: "text", text: "turn 2" }] },
+    ],
+  };
+
+  extractSystemRoleMessages(payload);
+
+  const after = layout(payload);
+  assert.equal(after.sys, 1); // only the client's own system prompt keeps its marker
+  assert.equal(after.msg, 2); // one boundary per hoisted note, each at its own position
+  // Anthropic caps a request at four breakpoints, so relocation must never inflate the count.
+  assert.equal(after.sys + after.msg, 3);
+});
+
+test("with no usable preceding message the marker stays on the hoisted block", () => {
+  // A shorter cached prefix is acceptable; a longer one would cache content the client did not
+  // mark. Nothing precedes the note here, so system[] is the same cut point.
+  const payload: Record<string, unknown> = {
+    messages: [
+      {
+        role: "system",
+        content: [{ type: "text", text: "note", cache_control: { type: "ephemeral" } }],
+      },
+      { role: "user", content: [{ type: "text", text: "turn 1" }] },
+    ],
+  };
+
+  extractSystemRoleMessages(payload);
+
+  assert.deepEqual(layout(payload), { sys: 1, msg: 0 });
+  const messages = payload.messages as Array<{ content: Array<Record<string, unknown>> }>;
+  assert.equal(messages[0].content[0].cache_control, undefined);
+});
+
+test("#9436 also holds for the second hoisting implementation", () => {
+  // extractSystemMessagesToBody is a near-duplicate of extractSystemRoleMessages; a fix that
+  // touches only one of them leaves the other path broken.
+  const payload = claudeCodeTurn("systemNote");
+  extractSystemMessagesToBody(payload);
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+});
+
+test("the second hoisting implementation resolves an occupied target identically", () => {
+  const dropped = occupiedTarget("1h", "5m");
+  extractSystemMessagesToBody(dropped);
+  assert.equal(markers(dropped).hoisted, undefined);
+  assert.deepEqual(layout(dropped), { sys: 0, msg: 1 });
+
+  const kept = occupiedTarget("5m", "1h");
+  extractSystemMessagesToBody(kept);
+  assert.deepEqual(markers(kept).hoisted, { type: "ephemeral", ttl: "1h" });
+  assert.deepEqual(layout(kept), { sys: 1, msg: 1 });
+});
+
+test("a thinking block is never chosen as the boundary", () => {
+  const payload: Record<string, unknown> = {
+    messages: [
+      {
+        role: "assistant",
+        content: [
+          { type: "text", text: "reply 1" },
+          { type: "thinking", thinking: "…", signature: "sig" },
+        ],
+      },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note", cache_control: { type: "ephemeral" } }],
+      },
+      { role: "user", content: [{ type: "text", text: "turn 2" }] },
+    ],
+  };
+
+  extractSystemRoleMessages(payload);
+
+  const messages = payload.messages as Array<{ content: Array<Record<string, unknown>> }>;
+  assert.deepEqual(messages[0].content[0].cache_control, { type: "ephemeral" }); // the text block
+  assert.equal(messages[0].content[1].cache_control, undefined); // the thinking block
+  assert.deepEqual(layout(payload), { sys: 0, msg: 1 });
+});
+
+test("with only thinking and empty text before it the marker stays on the hoisted block", () => {
+  const payload: Record<string, unknown> = {
+    messages: [
+      { role: "user", content: [{ type: "text", text: "" }] },
+      { role: "assistant", content: [{ type: "thinking", thinking: "…", signature: "sig" }] },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note", cache_control: { type: "ephemeral" } }],
+      },
+      { role: "user", content: [{ type: "text", text: "turn 2" }] },
+    ],
+  };
+
+  extractSystemRoleMessages(payload);
+
+  assert.deepEqual(layout(payload), { sys: 1, msg: 0 });
+  assert.deepEqual((payload.system as Array<Record<string, unknown>>)[0].cache_control, {
+    type: "ephemeral",
+  });
+});
+
+test("a tool_result whose array yields no text is skipped as a target", () => {
+  // Normalisation keeps only the non-empty text parts of a tool_result array, so this block is
+  // dropped entirely — anchoring the boundary on it would lose the marker again.
+  const payload: Record<string, unknown> = {
+    messages: [
+      {
+        role: "user",
+        content: [
+          {
+            type: "tool_result",
+            tool_use_id: "toolu_1",
+            content: [
+              { type: "image", source: { type: "base64", media_type: "image/png", data: "iVBOR" } },
+              { type: "text", text: "" },
+            ],
+          },
+        ],
+      },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note", cache_control: { type: "ephemeral" } }],
+      },
+      { role: "user", content: [{ type: "text", text: "turn 2" }] },
+    ],
+  };
+
+  extractSystemRoleMessages(payload);
+
+  assert.deepEqual(layout(payload), { sys: 1, msg: 0 });
+  assert.deepEqual((payload.system as Array<Record<string, unknown>>)[0].cache_control, {
+    type: "ephemeral",
+  });
+});
+
+test("a marker relocated onto a tool_result survives its collapse into text", () => {
+  const payload: Record<string, unknown> = {
+    messages: [
+      { role: "user", content: [{ type: "tool_result", tool_use_id: "toolu_1", content: "ok" }] },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note", cache_control: { type: "ephemeral" } }],
+      },
+      { role: "user", content: [{ type: "text", text: "turn 2" }] },
+    ],
+  };
+
+  // Without preserveToolResultBlocks the tool_result is rewritten into a plain text block.
+  normalizeClaudeUpstreamMessages(payload);
+
+  const messages = payload.messages as Array<{ content: Array<Record<string, unknown>> }>;
+  assert.equal(messages[0].content[0].type, "text");
+  assert.match(messages[0].content[0].text as string, /^\[Tool Result: toolu_1\]/);
+  assert.deepEqual(messages[0].content[0].cache_control, { type: "ephemeral" });
+  assert.deepEqual(layout(payload), { sys: 0, msg: 1 });
+});
+
+test("a marker relocated onto an inlined document survives its conversion to text", () => {
+  const payload: Record<string, unknown> = {
+    messages: [
+      {
+        role: "user",
+        content: [
+          { type: "document", name: "notes.txt", document: {}, content: "document body" },
+        ],
+      },
+      {
+        role: "system",
+        content: [{ type: "text", text: "note", cache_control: { type: "ephemeral" } }],
+      },
+    ],
+  };
+
+  normalizeClaudeUpstreamMessages(payload);
+
+  const messages = payload.messages as Array<{ content: Array<Record<string, unknown>> }>;
+  assert.equal(messages[0].content[0].type, "text");
+  assert.match(messages[0].content[0].text as string, /^\[notes\.txt\]\ndocument body$/);
+  assert.deepEqual(messages[0].content[0].cache_control, { type: "ephemeral" });
+  assert.deepEqual(layout(payload), { sys: 0, msg: 1 });
+});
+
+test("#9436 holds on the native Claude path through normalizeClaudeUpstreamMessages", () => {
+  // Path proof: the hoist is not confined to the semantic-passthrough branch — the function the
+  // native Claude passthrough calls routes through extractSystemRoleMessages.
+  const payload = claudeCodeTurn("systemNote");
+  normalizeClaudeUpstreamMessages(payload, { preserveToolResultBlocks: true });
+
+  const messages = payload.messages as Array<{ role: string }>;
+  assert.ok(!messages.some((m) => m.role === "system"));
+  assert.deepEqual(layout(payload), { sys: 1, msg: 1 });
+});