diff --git a/agent/transports/codex.py b/agent/transports/codex.py
index 6738ed3220cf..ccc36005be9b 100644
--- a/agent/transports/codex.py
+++ b/agent/transports/codex.py
@@ -5,12 +5,31 @@
 streaming, or the _run_codex_stream() call path.
 """
 
+import hashlib
 from typing import Any, Dict, List, Optional
 
 from agent.transports.base import ProviderTransport
 from agent.transports.types import NormalizedResponse, ToolCall
 
 
+_PROMPT_CACHE_KEY_MAX_LENGTH = 64
+
+
+def _safe_prompt_cache_key(value: str) -> str:
+    """Return a Codex-compatible prompt_cache_key.
+
+    The Codex backend rejects prompt_cache_key values longer than 64
+    characters. Hermes session IDs can exceed that when gateway/platform
+    routing embeds long external identifiers. Preserve short IDs for
+    debuggability and hash only oversized values into a stable 64-character
+    scope key.
+    """
+    raw = str(value or "").strip()
+    if len(raw) <= _PROMPT_CACHE_KEY_MAX_LENGTH:
+        return raw
+    return hashlib.sha256(raw.encode("utf-8")).hexdigest()
+
+
 class ResponsesApiTransport(ProviderTransport):
     """Transport for api_mode='codex_responses'.
 
@@ -101,7 +120,7 @@ def build_kwargs(
 
         session_id = params.get("session_id")
         if not is_github_responses and session_id:
-            kwargs["prompt_cache_key"] = session_id
+            kwargs["prompt_cache_key"] = _safe_prompt_cache_key(str(session_id))
 
         if reasoning_enabled and is_xai_responses:
             from agent.model_metadata import grok_supports_reasoning_effort
diff --git a/tests/agent/transports/test_codex_transport.py b/tests/agent/transports/test_codex_transport.py
index 6a4cda173ad4..25a98196a59b 100644
--- a/tests/agent/transports/test_codex_transport.py
+++ b/tests/agent/transports/test_codex_transport.py
@@ -91,6 +91,16 @@ def test_session_id_sets_cache_key(self, transport):
         )
         assert kw.get("prompt_cache_key") == "test-session-123"
 
+    def test_long_session_id_cache_key_is_hashed_to_backend_limit(self, transport):
+        messages = [{"role": "user", "content": "Hi"}]
+        long_session_id = "atlas:4CGVtvUxCQrizW2tuRbqNa25nLnClXPm:494fd2fb-ca04-431b-a26d-f6c255d0804e"
+        kw = transport.build_kwargs(
+            model="gpt-5.4", messages=messages, tools=[],
+            session_id=long_session_id,
+        )
+        assert len(kw.get("prompt_cache_key")) == 64
+        assert kw.get("prompt_cache_key") != long_session_id
+
     def test_github_responses_no_cache_key(self, transport):
         messages = [{"role": "user", "content": "Hi"}]
         kw = transport.build_kwargs(