Skip to content

The agent that grows with you

fix: clamp a messaging chat's displayed effort for its route's api_mode

Teknium79f2c7cfb67f
diff --git a/gateway/session_messaging_picks.py b/gateway/session_messaging_picks.py
index 9e5f73a4e9a..52c4fda274a 100644
--- a/gateway/session_messaging_picks.py
+++ b/gateway/session_messaging_picks.py
@@ -116,6 +116,9 @@ def messaging_route_info(runner, live):
override = runner._session_model_override(live.route) or runner.session_store.get_model_override(live.route) or {}
model = override.get('model') or model
provider, base_url = override.get('provider') or provider, override.get('base_url') or base_url
+ # The wire level clamps per api_mode (a Codex app-server route keeps ``ultra``); config.yaml's
+ # api_mode belongs to its own provider only.
+ api_mode = override.get('api_mode') or (block.get('api_mode') if provider == block.get('provider') else None)
reasoning = runner._resolve_session_reasoning_config(source=live.source, session_key=live.route, model=model)
tier = runner._resolve_session_service_tier(source=live.source, session_key=live.route)
- return {'model': model, **effort_fields(reasoning, tier, model, provider, base_url)}
+ return {'model': model, **effort_fields(reasoning, tier, model, provider, base_url, api_mode)}
diff --git a/tests/gateway/test_messaging_session_picks.py b/tests/gateway/test_messaging_session_picks.py
index aa7df937ad6..be47d34d3a7 100644
--- a/tests/gateway/test_messaging_session_picks.py
+++ b/tests/gateway/test_messaging_session_picks.py
@@ -147,3 +147,24 @@ async def test_a_cold_model_pick_resolves_against_the_chats_saved_provider(tmp_p
finally:
await viewer.close()
runner.session_store.close_all_db_handles()
+
+
+async def test_a_messaging_chats_effort_wire_is_clamped_for_its_api_mode(tmp_path, monkeypatch):
+ """A chat on the Codex app-server route displays ``ultra`` on the wire, as that route sends it;
+ clamping without the route's api_mode showed ``max``."""
+ import gateway.run as gateway_run
+ monkeypatch.setattr(gateway_run, '_hermes_home', tmp_path)
+ (tmp_path / 'config.yaml').write_text(
+ 'model:\n default: gpt-5.6\n provider: openai-codex\n api_mode: codex_app_server\n')
+ runner, _source, _route, viewer, root, fences = await _compressed_chat(tmp_path)
+ try:
+ picked = await viewer.dispatch(_mutate(root, fences, 'reasoning', {'effort': 'ultra'}, 'pick-ultra'))
+ assert picked.get('result', {}).get('effort') == 'ultra', picked
+
+ info = await viewer.dispatch({'id': 3, 'method': 'session.info', 'params': {'session_id': root}})
+
+ assert (info['result']['model'], info['result']['reasoning_effort_wire']) == ('gpt-5.6', 'ultra'), info
+ finally:
+ await viewer.close()
+ runner.session_store.close_all_db_handles()