fix codex app server to launch with a default model

src/ucode/agents/codex.py

@@ -29,8 +29,10 @@
 from ucode.managed_files import OS, current_os, write_managed_file
 from ucode.smart_routing.codex_hooks import (
     remove_smart_routing_hooks,
+    routing_models,
     sync_smart_routing_hooks,
 )
+from ucode.smart_routing.codex_routing import codex_model_id
 from ucode.state import mark_tool_managed, save_state
 from ucode.telemetry import agent_version, ucode_version
 from ucode.ui import print_warning_err
@@ -49,6 +51,7 @@
 # Shared across agents: one opt-in enables smart routing for every routing-capable
 # tool (codex, claude), so a workspace turns it on once.
 SMART_ROUTING_STATE_KEY = "smart_routing_enabled"
+APP_SERVER_SMART_ROUTING_STARTING_MODEL = "gpt-5.6-luna"
 
 SPEC: ToolSpec = {
     "binary": "codex",
@@ -456,11 +459,20 @@ def launch(state: dict, tool_args: list[str]) -> None:
         # path so flag-off launches retain their existing dependencies and behavior.
         from ucode.smart_routing import v2 as smart_routing_v2
 
+        def _app_server_start_model() -> str:
+            managed_model = default_model(state)
+            if managed_model:
+                return managed_model
+            models = routing_models(state)
+            if models:
+                return codex_model_id(models[0])
+            return APP_SERVER_SMART_ROUTING_STARTING_MODEL
+
         smart_routing_v2.launch_codex(
             state,
             tool_args,
             binary=binary,
-            start_model=default_model(state),
+            start_model=_app_server_start_model(),
             render_overlay=render_overlay,
         )
     if workspace:

src/ucode/smart_routing/codex_interposer.py

@@ -7,6 +7,7 @@
 import time
 import uuid
 from collections.abc import Callable
+from dataclasses import replace
 from pathlib import Path
 
 from websockets.asyncio.client import connect
@@ -90,6 +91,7 @@ def on_tui_frame(self, raw: str) -> str:
                 if decision is None:
                     self.log(f"[ROUTE] selection failed; keeping current model: {reason}")
                     return raw
+                decision = replace(decision, model=codex_routing.codex_model_id(decision.model))
                 self.target = decision.model
                 if self.switch_message_fn is not None:
                     self.switch_message = self.switch_message_fn(decision.model, decision.rationale)

src/ucode/smart_routing/codex_routing.py

@@ -159,7 +159,7 @@ def record(payload, task, decision, requested):
             workspace, token, task, available_models, timeout=timeout
         ),
         default_task_label="Codex subagent task",
-        model_id_mapper=_codex_model_id,
+        model_id_mapper=codex_model_id,
         record_decision=record,
     )
 
@@ -203,7 +203,13 @@ def _model_strength(model: str) -> tuple[int, int, int, int]:
     return major, minor, patch, 1 if not suffix else 0
 
 
-def _codex_model_id(model: str) -> str:
+def codex_model_id(model: str) -> str:
+    """Map a UC GPT service ID to Codex's bundled catalog slug.
+
+    Codex's bundled GPT catalog owns the model metadata for these aliases,
+    while the AI Gateway resolves them back to the matching ``system.ai`` service.
+    Leave non-GPT models unchanged because their metadata comes from the gateway catalog.
+    """
     tail = model.rsplit("/", 1)[-1]
     if tail in {"databricks-gpt-5-2-codex", "databricks-gpt-5-4-nano"}:
         return tail

src/ucode/smart_routing/v2.py

@@ -468,9 +468,9 @@ def launch_codex(
     os.environ[OAUTH_TOKEN_ENV_VAR] = get_databricks_token(workspace, profile)
     available_models = _cached_routing_models(state)
     if not available_models:
-        raise RuntimeError(
-            "Smart routing v2 has no cached Unity Catalog model services; "
-            "run `ucode configure codex` to refresh them."
+        print_note(
+            "Smart routing model metadata is unavailable; starting Codex on gpt-5.6-luna "
+            "without automatic model switching. Run `ucode configure codex` to enable routing."
         )
     overlay = render_overlay(
         workspace,

tests/test_codex_routing.py

@@ -204,6 +204,21 @@ def test_spawn_rewrite_uses_codex_model_id_for_uc_endpoint(monkeypatch):
     assert output["hookSpecificOutput"]["updatedInput"]["model"] == "gpt-5.6-luna"
 
 
+def test_codex_model_id_maps_uc_gpt_models_to_codex_slugs():
+    expected = {
+        "system.ai.gpt-5-2": "gpt-5.2",
+        "system.ai.gpt-5-4": "gpt-5.4",
+        "system.ai.gpt-5-4-mini": "gpt-5.4-mini",
+        "system.ai.gpt-5-5": "gpt-5.5",
+        "system.ai.gpt-5-6-luna": "gpt-5.6-luna",
+        "system.ai.gpt-5-6-sol": "gpt-5.6-sol",
+        "system.ai.gpt-5-6-terra": "gpt-5.6-terra",
+    }
+    assert {model: codex_routing.codex_model_id(model) for model in expected} == expected
+    assert codex_routing.codex_model_id("system.ai.gpt-5-6-experimental") == "gpt-5.6-experimental"
+    assert codex_routing.codex_model_id("system.ai.glm-5-2") == "system.ai.glm-5-2"
+
+
 def test_spawn_glm_decision_applies_glm_model(monkeypatch):
     # GLM is no longer skipped for Codex subagents: a GLM routing decision is
     # applied like any other arm.

tests/test_codex_smart_routing_v2.py

@@ -53,6 +53,7 @@ def test_codex_launch_dispatches_when_flag_enabled(self, monkeypatch):
         calls = []
         monkeypatch.setenv(v2.ENV_VAR, "1")
         monkeypatch.setattr(codex, "default_model", lambda state: "gpt-start")
+        monkeypatch.setattr(codex, "clear_model_preferences", lambda state: False)
 
         def launch_v2(state, tool_args, **kwargs):
             calls.append((state, tool_args, kwargs))
@@ -77,6 +78,26 @@ def launch_v2(state, tool_args, **kwargs):
             )
         ]
 
+    def test_codex_launch_normalizes_cached_bootstrap_model(self, monkeypatch):
+        calls = []
+        monkeypatch.setenv(v2.ENV_VAR, "1")
+        monkeypatch.setattr(codex, "clear_model_preferences", lambda state: False)
+        monkeypatch.setattr(codex, "default_model", lambda state: None)
+
+        def launch_v2(state, tool_args, **kwargs):
+            calls.append(kwargs)
+            raise SystemExit(0)
+
+        monkeypatch.setattr(v2, "launch_codex", launch_v2)
+
+        with pytest.raises(SystemExit):
+            codex.launch(
+                {"workspace": WS, "codex_models": ["system.ai.gpt-5-6-luna"]},
+                [],
+            )
+
+        assert calls[0]["start_model"] == "gpt-5.6-luna"
+
     def test_owns_app_server_interposer_and_tui_lifecycle(self, monkeypatch):
         processes = []
         interposer_args = {}
@@ -233,18 +254,41 @@ def test_v2_pre_tool_hook_replaces_existing_ucode_hook(self, tmp_path, monkeypat
         assert "--model system.ai.gpt-5-6-sol" in routing_commands[0]
         assert "--model old" not in routing_commands[0]
 
-    def test_missing_cached_models_blocks_launch(self, monkeypatch):
+    def test_missing_cached_models_starts_with_bootstrap_model(self, monkeypatch):
         monkeypatch.setattr(v2, "get_databricks_token", lambda workspace, profile: "token")
+        monkeypatch.setattr(codex, "agent_version", lambda binary: "unknown")
+        monkeypatch.setattr(v2, "_free_port", lambda: 41001)
+        monkeypatch.setattr(v2, "_wait_for_app_server", lambda port, timeout: True)
+        monkeypatch.setattr(
+            v2.subprocess,
+            "Popen",
+            lambda *args, **kwargs: type(
+                "Process",
+                (),
+                {
+                    "wait": lambda self, timeout=None: 0,
+                    "terminate": lambda self: None,
+                    "kill": lambda self: None,
+                },
+            )(),
+        )
+        monkeypatch.setattr(
+            codex_interposer,
+            "start_interposer_thread",
+            lambda *args, **kwargs: (41002, lambda: None),
+        )
 
-        with pytest.raises(RuntimeError, match="ucode configure codex"):
+        with pytest.raises(SystemExit) as exc:
             v2.launch_codex(
                 {"workspace": WS},
                 [],
                 binary="codex",
-                start_model="gpt-start",
+                start_model="gpt-5.6-luna",
                 render_overlay=codex.render_overlay,
             )
 
+        assert exc.value.code == 0
+
 
 def test_interposer_startup_failure_is_propagated(monkeypatch):
     async def fail_to_serve(*args, **kwargs):
@@ -368,7 +412,7 @@ def select(prompt):
         assert json.loads(output)["params"]["model"] == "claude-opus-4-8"
         assert "Task classified as bugfix." in sess.switch_message
 
-    def test_shows_routing_notice_when_selected_model_is_already_active(self):
+    def test_maps_selected_uc_gpt_model_and_shows_routing_notice(self):
         def select(_prompt):
             return (
                 codex_interposer.routing.RoutingDecision(
@@ -387,14 +431,16 @@ def select(_prompt):
         )
         frame = self._turn_start("system.ai.gpt-5-6-luna")
 
-        assert sess.on_tui_frame(frame) == frame
+        output = sess.on_tui_frame(frame)
+        assert json.loads(output)["params"]["model"] == "gpt-5.6-luna"
         injected = sess.on_engine_frame(self._turn_started("turn-1"))
 
         assert [message["method"] for message in injected] == [
+            codex_interposer.SETTINGS_UPDATED,
             codex_interposer.ITEM_STARTED,
             codex_interposer.ITEM_COMPLETED,
         ]
-        assert "Selected Model : system.ai.gpt-5-6-luna" in (injected[0]["params"]["item"]["text"])
+        assert "Selected Model : gpt-5.6-luna" in (injected[1]["params"]["item"]["text"])
 
     def test_routes_first_prompt_to_oss_model(self):
         def select(_prompt):