diff --git a/open-sse/executors/codex.ts b/open-sse/executors/codex.ts
index c645dbcb40..77074b0557 100644
--- a/open-sse/executors/codex.ts
+++ b/open-sse/executors/codex.ts
@@ -879,7 +879,7 @@ export class CodexExecutor extends BaseExecutor {
     // session_id header — enables prompt cache affinity on the Codex backend.
     // The official Codex client sets this to conversation_id (a stable UUID per session).
     // Ref: openai/codex codex-api/src/requests/headers.rs build_conversation_headers()
-    const cacheSessionId = this.getPromptCacheSessionId(credentials);
+    const cacheSessionId = this.getPromptCacheSessionId(credentials, null);
     if (cacheSessionId) {
       headers["session_id"] = cacheSessionId;
     }
@@ -889,12 +889,19 @@ export class CodexExecutor extends BaseExecutor {
 
   /**
    * Derive a stable session ID for prompt cache affinity.
-   * Uses workspaceId (chatgpt account ID) as the cache partition key.
-   * This mirrors the official Codex client's use of conversation_id for
-   * prompt_cache_key and session_id header.
+   * Priority: per-conversation session_id/conversation_id from request body → workspaceId.
+   * The official Codex client uses conversation_id (a unique UUID per session), NOT
+   * the account-wide workspaceId. Using workspaceId caps cache hit-rate at ~49%
+   * because all conversations share the same cache partition. (#1643)
    * Ref: openai/codex core/src/client.rs line 853
    */
-  private getPromptCacheSessionId(credentials): string | null {
+  private getPromptCacheSessionId(credentials, body: Record<string, unknown> | null): string | null {
+    // Prefer per-session identifiers from the client request body
+    const sessionId = body?.session_id ?? body?.conversation_id;
+    if (typeof sessionId === "string" && sessionId.length > 0) {
+      return sessionId;
+    }
+    // Fall back to workspaceId (account-wide) — better than nothing
     return credentials?.providerSpecificData?.workspaceId || null;
   }
 
@@ -1065,8 +1072,9 @@ export class CodexExecutor extends BaseExecutor {
     // The official Codex client sets this to conversation_id (a stable UUID per session).
     // Ref: openai/codex core/src/client.rs line 853:
     //   let prompt_cache_key = Some(self.client.state.conversation_id.to_string());
+    // IMPORTANT: Capture session/conversation IDs BEFORE deletion below (#1643).
     if (!body.prompt_cache_key) {
-      const cacheSessionId = this.getPromptCacheSessionId(credentials);
+      const cacheSessionId = this.getPromptCacheSessionId(credentials, body);
       if (cacheSessionId) {
         body.prompt_cache_key = cacheSessionId;
       }
diff --git a/tests/unit/codex-cache-key.test.ts b/tests/unit/codex-cache-key.test.ts
new file mode 100644
index 0000000000..a774b27395
--- /dev/null
+++ b/tests/unit/codex-cache-key.test.ts
@@ -0,0 +1,89 @@
+/**
+ * Tests for Codex prompt_cache_key per-session behavior (#1643).
+ *
+ * Validates that getPromptCacheSessionId prefers per-conversation
+ * session_id/conversation_id from request body over account-wide workspaceId.
+ */
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+
+// We test the logic inline since the method is private — replicate the priority logic
+function getPromptCacheSessionId(
+  credentials: { providerSpecificData?: { workspaceId?: string } } | null,
+  body: Record<string, unknown> | null
+): string | null {
+  const sessionId = body?.session_id ?? body?.conversation_id;
+  if (typeof sessionId === "string" && sessionId.length > 0) {
+    return sessionId;
+  }
+  return credentials?.providerSpecificData?.workspaceId || null;
+}
+
+describe("Codex prompt_cache_key (#1643)", () => {
+  it("should prefer body.session_id over workspaceId", () => {
+    const creds = { providerSpecificData: { workspaceId: "ws-account-001" } };
+    const body = { session_id: "sess-conv-unique-001" };
+    const result = getPromptCacheSessionId(creds, body);
+    assert.equal(result, "sess-conv-unique-001");
+  });
+
+  it("should prefer body.conversation_id over workspaceId", () => {
+    const creds = { providerSpecificData: { workspaceId: "ws-account-001" } };
+    const body = { conversation_id: "conv-unique-002" };
+    const result = getPromptCacheSessionId(creds, body);
+    assert.equal(result, "conv-unique-002");
+  });
+
+  it("should prefer session_id over conversation_id", () => {
+    const creds = { providerSpecificData: { workspaceId: "ws-account-001" } };
+    const body = { session_id: "sess-001", conversation_id: "conv-002" };
+    const result = getPromptCacheSessionId(creds, body);
+    assert.equal(result, "sess-001");
+  });
+
+  it("should fall back to workspaceId when body has no session identifiers", () => {
+    const creds = { providerSpecificData: { workspaceId: "ws-account-001" } };
+    const body = { model: "gpt-5.5" };
+    const result = getPromptCacheSessionId(creds, body);
+    assert.equal(result, "ws-account-001");
+  });
+
+  it("should fall back to workspaceId when body is null", () => {
+    const creds = { providerSpecificData: { workspaceId: "ws-account-001" } };
+    const result = getPromptCacheSessionId(creds, null);
+    assert.equal(result, "ws-account-001");
+  });
+
+  it("should return null when neither body nor credentials have IDs", () => {
+    const result = getPromptCacheSessionId({}, {});
+    assert.equal(result, null);
+  });
+
+  it("should return null for null credentials and empty body", () => {
+    const result = getPromptCacheSessionId(null, null);
+    assert.equal(result, null);
+  });
+
+  it("should ignore empty string session_id", () => {
+    const creds = { providerSpecificData: { workspaceId: "ws-001" } };
+    const body = { session_id: "" };
+    const result = getPromptCacheSessionId(creds, body);
+    assert.equal(result, "ws-001");
+  });
+
+  it("should ignore non-string session_id", () => {
+    const creds = { providerSpecificData: { workspaceId: "ws-001" } };
+    const body = { session_id: 12345 };
+    const result = getPromptCacheSessionId(creds, body);
+    assert.equal(result, "ws-001");
+  });
+
+  it("each conversation should get a unique cache key", () => {
+    const creds = { providerSpecificData: { workspaceId: "ws-shared" } };
+    const conv1 = { session_id: "conv-aaaa" };
+    const conv2 = { session_id: "conv-bbbb" };
+    const key1 = getPromptCacheSessionId(creds, conv1);
+    const key2 = getPromptCacheSessionId(creds, conv2);
+    assert.notEqual(key1, key2, "Different conversations should have different cache keys");
+  });
+});