fix(provider): preserve Codex prompt cache affinity

nanobot/providers/openai_codex_provider.py

@@ -111,6 +111,7 @@ class OpenAICodexProvider(LLMProvider):
             model=_strip_model_prefix(model),
         )
         session_id = provider_context.session_id if provider_context is not None else None
+        session_routing_key = _prompt_cache_key(session_id) if session_id else None
 
         body: dict[str, Any] = {
             "model": _strip_model_prefix(model),
@@ -122,8 +123,8 @@ class OpenAICodexProvider(LLMProvider):
             "tool_choice": tool_choice or "auto",
             "parallel_tool_calls": True,
         }
-        if session_id:
-            body["prompt_cache_key"] = _prompt_cache_key(session_id)
+        if session_routing_key:
+            body["prompt_cache_key"] = session_routing_key
         body["include"] = ["reasoning.encrypted_content"]
         reasoning_options = _build_reasoning_options(reasoning_effort)
         if replayed and "gpt-5.6" in _strip_model_prefix(model).lower():
@@ -136,13 +137,20 @@ class OpenAICodexProvider(LLMProvider):
         if self._extra_body:
             # Apply explicit provider overrides last, matching other provider backends.
             body.update(self._extra_body)
+        effective_cache_key = body.get("prompt_cache_key")
 
         stage = "oauth_token"
         native_compaction_applied = False
         native_compaction_state: ProviderConversationState | None = None
         try:
             token = await asyncio.to_thread(get_codex_token, proxy=self.proxy)
-            headers = _build_headers(cast(str, token.account_id), token.access)
+            headers = _build_headers(
+                cast(str, token.account_id),
+                token.access,
+                session_routing_key=(
+                    effective_cache_key if isinstance(effective_cache_key, str) else None
+                ),
+            )
 
             async def _send(
                 request_body: dict[str, Any],
@@ -416,8 +424,13 @@ def _build_reasoning_options(reasoning_effort: str | None) -> dict[str, str] | N
     return options
 
 
-def _build_headers(account_id: str, token: str) -> dict[str, str]:
-    return {
+def _build_headers(
+    account_id: str,
+    token: str,
+    *,
+    session_routing_key: str | None = None,
+) -> dict[str, str]:
+    headers = {
         "Authorization": f"Bearer {token}",
         "chatgpt-account-id": account_id,
         "OpenAI-Beta": "responses=experimental",
@@ -426,6 +439,9 @@ def _build_headers(account_id: str, token: str) -> dict[str, str]:
         "accept": "text/event-stream",
         "content-type": "application/json",
     }
+    if session_routing_key:
+        headers["session-id"] = session_routing_key
+    return headers
 
 
 class _CodexHTTPError(RuntimeError):

tests/providers/test_openai_codex_provider.py

@@ -265,6 +265,7 @@ async def test_codex_request_uses_configured_proxy(monkeypatch) -> None:
 @pytest.mark.asyncio
 async def test_codex_omits_prompt_cache_key_without_session_id(monkeypatch) -> None:
     bodies: list[dict[str, Any]] = []
+    headers_seen: list[dict[str, str]] = []
 
     _mock_codex_token(monkeypatch)
 
@@ -280,6 +281,7 @@ async def test_codex_omits_prompt_cache_key_without_session_id(monkeypatch) -> N
     ):
         _ = proxy, on_thinking_delta, on_tool_call_delta
         bodies.append(body)
+        headers_seen.append(headers)
         return provider_base.LLMResponse(content="ok")
 
     monkeypatch.setattr("nanobot.providers.openai_codex_provider._request_codex", fake_request)
@@ -294,16 +296,19 @@ async def test_codex_omits_prompt_cache_key_without_session_id(monkeypatch) -> N
     )
 
     assert "prompt_cache_key" not in bodies[0]
+    assert "session-id" not in headers_seen[0]
     assert "service_tier" not in bodies[0]
 
 
 @pytest.mark.asyncio
 async def test_codex_prompt_cache_key_prefers_stable_session_id(monkeypatch) -> None:
     bodies: list[dict[str, Any]] = []
+    headers_seen: list[dict[str, str]] = []
     _mock_codex_token(monkeypatch)
 
-    async def fake_request(_url, _headers, body, **_kwargs):
+    async def fake_request(_url, headers, body, **_kwargs):
         bodies.append(body)
+        headers_seen.append(headers)
         return provider_base.LLMResponse(content="ok")
 
     monkeypatch.setattr("nanobot.providers.openai_codex_provider._request_codex", fake_request)
@@ -326,15 +331,22 @@ async def test_codex_prompt_cache_key_prefers_stable_session_id(monkeypatch) ->
 
     assert bodies[0]["prompt_cache_key"] == bodies[1]["prompt_cache_key"]
     assert bodies[0]["prompt_cache_key"] != bodies[2]["prompt_cache_key"]
+    assert headers_seen[0]["session-id"] != "session-a"
+    assert headers_seen[2]["session-id"] != "session-b"
+    assert headers_seen[0]["session-id"] == bodies[0]["prompt_cache_key"]
+    assert headers_seen[1]["session-id"] == bodies[1]["prompt_cache_key"]
+    assert headers_seen[2]["session-id"] == bodies[2]["prompt_cache_key"]
 
 
 @pytest.mark.asyncio
 async def test_codex_provider_applies_extra_body_from_config(monkeypatch) -> None:
     bodies: list[dict[str, Any]] = []
+    headers_seen: list[dict[str, str]] = []
     _mock_codex_token(monkeypatch)
 
-    async def fake_request(_url, _headers, body, **_kwargs):
+    async def fake_request(_url, headers, body, **_kwargs):
         bodies.append(body)
+        headers_seen.append(headers)
         return provider_base.LLMResponse(content="ok")
 
     monkeypatch.setattr("nanobot.providers.openai_codex_provider._request_codex", fake_request)
@@ -347,7 +359,10 @@ async def test_codex_provider_applies_extra_body_from_config(monkeypatch) -> Non
         },
         "providers": {
             "openaiCodex": {
-                "extraBody": {"service_tier": "priority"},
+                "extraBody": {
+                    "service_tier": "priority",
+                    "prompt_cache_key": "explicit-cache-key",
+                },
             },
         },
     })
@@ -357,6 +372,8 @@ async def test_codex_provider_applies_extra_body_from_config(monkeypatch) -> Non
 
     assert response.content == "ok"
     assert bodies[0]["service_tier"] == "priority"
+    assert bodies[0]["prompt_cache_key"] == "explicit-cache-key"
+    assert headers_seen[0]["session-id"] == "explicit-cache-key"
 
 
 @pytest.mark.asyncio