fix: setup wizard overwrites custom endpoint config (#4172)

_model_flow_custom() saved model.provider and model.base_url to disk
via its own load_config/save_config cycle, but never updated the
setup wizard's in-memory config dict.  The wizard's final
save_config(config) then overwrote the custom settings with the
stale default string model value.

Fix: after saving to disk, also mutate the caller's config dict so
the wizard's final save preserves model.provider='custom' and the
base_url.  Both the model_name and no-model_name branches are
covered.

Added regression tests that simulate the full wizard flow including
the final save_config(config) call — the step that was previously
untested.

hermes_cli/main.py

@@ -1278,10 +1278,25 @@ def _model_flow_custom(config):
         save_config(cfg)
         deactivate_provider()
 
+        # Sync the caller's config dict so the setup wizard's final
+        # save_config(config) preserves our model settings.  Without
+        # this, the wizard overwrites model.provider/base_url with
+        # the stale values from its own config dict (#4172).
+        config["model"] = dict(model)
+
         print(f"Default model set to: {model_name} (via {effective_url})")
     else:
         if base_url or api_key:
             deactivate_provider()
+        # Even without a model name, persist the custom endpoint on the
+        # caller's config dict so the setup wizard doesn't lose it.
+        _caller_model = config.get("model")
+        if not isinstance(_caller_model, dict):
+            _caller_model = {"default": _caller_model} if _caller_model else {}
+        _caller_model["provider"] = "custom"
+        _caller_model["base_url"] = effective_url
+        _caller_model.pop("api_mode", None)
+        config["model"] = _caller_model
         print("Endpoint saved. Use `/model` in chat or `hermes model` to set a model.")
 
     # Auto-save to custom_providers so it appears in the menu next time

tests/hermes_cli/test_setup.py

@@ -118,11 +118,125 @@ def test_custom_setup_clears_active_oauth_provider(tmp_path, monkeypatch):
     # Core assertion: switching to custom endpoint clears OAuth provider
     assert get_active_provider() is None
 
-    # _model_flow_custom writes config via its own load/save cycle
+    # Simulate what the real setup wizard does: save_config(config) AFTER
+    # setup_model_provider returns.  This is the step that previously
+    # overwrote model.provider/base_url (#4172).
+    save_config(config)
+
     reloaded = load_config()
-    if isinstance(reloaded.get("model"), dict):
-        assert reloaded["model"].get("provider") == "custom"
-        assert reloaded["model"].get("default") == "custom/model"
+    assert isinstance(reloaded.get("model"), dict), (
+        "model should be a dict after custom setup, not "
+        + repr(type(reloaded.get("model")))
+    )
+    assert reloaded["model"].get("provider") == "custom"
+    assert reloaded["model"].get("default") == "custom/model"
+    assert "custom.example" in reloaded["model"].get("base_url", "")
+
+
+def test_custom_setup_preserves_provider_after_wizard_save_config(
+    tmp_path, monkeypatch
+):
+    """Regression test for #4172: the setup wizard's final save_config(config)
+    must not overwrite model.provider/base_url that _model_flow_custom set.
+
+    Simulates the full flow:
+      1. load config (fresh install — model is a string)
+      2. setup_model_provider picks custom
+      3. wizard calls save_config(config) afterward
+      4. verify resolve_requested_provider returns "custom"
+    """
+    monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+    _clear_provider_env(monkeypatch)
+
+    config = load_config()
+    # Sanity: fresh install has model as a string
+    assert isinstance(config.get("model"), str) or config.get("model") is None
+
+    def fake_prompt_choice(question, choices, default=0):
+        if question == "Select your inference provider:":
+            return 3  # Custom endpoint
+        tts_idx = _maybe_keep_current_tts(question, choices)
+        if tts_idx is not None:
+            return tts_idx
+        raise AssertionError(f"Unexpected prompt_choice call: {question}")
+
+    monkeypatch.setattr("hermes_cli.setup.prompt_choice", fake_prompt_choice)
+
+    input_values = iter([
+        "http://localhost:11434/v1",  # Ollama URL
+        "",                           # no API key (local)
+        "qwen3.5:32b",               # model name
+        "",                           # context length (auto-detect)
+    ])
+    monkeypatch.setattr("builtins.input", lambda _prompt="": next(input_values))
+    monkeypatch.setattr("hermes_cli.setup.prompt_yes_no", lambda *a, **kw: False)
+    monkeypatch.setattr("hermes_cli.auth.detect_external_credentials", lambda: [])
+    monkeypatch.setattr("hermes_cli.main._save_custom_provider", lambda *a, **kw: None)
+    monkeypatch.setattr(
+        "hermes_cli.models.probe_api_models",
+        lambda api_key, base_url: {"models": ["qwen3.5:32b"], "probed_url": base_url + "/models"},
+    )
+
+    # Full wizard cycle
+    setup_model_provider(config)
+    save_config(config)  # ← this is what the real wizard does
+
+    # Verify config on disk
+    reloaded = load_config()
+    assert isinstance(reloaded["model"], dict)
+    assert reloaded["model"]["provider"] == "custom"
+    assert reloaded["model"]["base_url"] == "http://localhost:11434/v1"
+    assert reloaded["model"]["default"] == "qwen3.5:32b"
+    assert "api_mode" not in reloaded["model"]
+
+    # Verify the runtime resolver sees "custom", not "auto"
+    from hermes_cli.runtime_provider import resolve_requested_provider
+    assert resolve_requested_provider() == "custom"
+
+
+def test_custom_setup_no_model_name_still_preserves_endpoint(
+    tmp_path, monkeypatch
+):
+    """When the user enters a URL and key but skips the model name,
+    model.provider and model.base_url must still survive the wizard's
+    final save_config(config)."""
+    monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+    _clear_provider_env(monkeypatch)
+
+    config = load_config()
+
+    def fake_prompt_choice(question, choices, default=0):
+        if question == "Select your inference provider:":
+            return 3
+        tts_idx = _maybe_keep_current_tts(question, choices)
+        if tts_idx is not None:
+            return tts_idx
+        raise AssertionError(f"Unexpected prompt_choice call: {question}")
+
+    monkeypatch.setattr("hermes_cli.setup.prompt_choice", fake_prompt_choice)
+
+    input_values = iter([
+        "http://192.168.1.50:8080/v1",  # URL
+        "my-key",                        # API key
+        "",                              # no model name
+        "",                              # context length
+    ])
+    monkeypatch.setattr("builtins.input", lambda _prompt="": next(input_values))
+    monkeypatch.setattr("hermes_cli.setup.prompt_yes_no", lambda *a, **kw: False)
+    monkeypatch.setattr("hermes_cli.auth.detect_external_credentials", lambda: [])
+    monkeypatch.setattr("hermes_cli.main._save_custom_provider", lambda *a, **kw: None)
+    monkeypatch.setattr(
+        "hermes_cli.models.probe_api_models",
+        lambda api_key, base_url: {"models": None, "probed_url": base_url + "/models"},
+    )
+
+    setup_model_provider(config)
+    save_config(config)
+
+    reloaded = load_config()
+    assert isinstance(reloaded["model"], dict)
+    assert reloaded["model"]["provider"] == "custom"
+    assert reloaded["model"]["base_url"] == "http://192.168.1.50:8080/v1"
 
 
 def test_codex_setup_uses_runtime_access_token_for_live_model_list(tmp_path, monkeypatch):