databricks · andy-xu-db · Sep 16, 2026
diff --git a/scripts/run_integration.py b/scripts/run_integration.py
@@ -105,6 +105,11 @@ def arguments():
         default="main.ucode.ci_e2e_anthropic_relay_mps",
         help="Existing relayed (subscription-relay) Anthropic MPS for the hybrid-routing CUJ.",
     )
+    parser.add_argument(
+        "--claude-provider-model",
+        default="claude-haiku-4-5-20251001",
+        help="Only model exposed by the Anthropic MPS discovery fixture.",
+    )
     parser.add_argument(
         "--codex-provider",
         default="main.ucode.ci_openai_mps",
@@ -120,6 +125,16 @@ def arguments():
         default="main.ucode",
         help="Schema containing the dedicated model-discovery Model Services.",
     )
+    parser.add_argument(
+        "--claude-parent-model",
+        default="main.ucode.ci_e2e_claude",
+        help="Claude-compatible Model Service in --parent-schema.",
+    )
+    parser.add_argument(
+        "--codex-parent-model",
+        default="main.ucode.ci_e2e_codex",
+        help="Codex-compatible Model Service in --parent-schema.",
+    )
     parser.add_argument("--python", default=sys.executable, help="Python 3.12+ path or uv version.")
     parser.add_argument("--dependency", action="append", default=[], metavar="PACKAGE==VERSION")
     parser.add_argument("--constraints", type=Path, help="Replay a previous dependencies.txt.")
@@ -293,9 +308,12 @@ def run(command, *, cwd=output, env=base_env, timeout=600) -> str:
             "codex_model": args.codex_model,
             "claude_provider": args.claude_provider,
             "claude_relayed_provider": args.claude_relayed_provider,
+            "claude_provider_model": args.claude_provider_model,
             "codex_provider": args.codex_provider,
             "codex_provider_model": args.codex_provider_model,
             "parent_schema": args.parent_schema,
+            "claude_parent_model": args.claude_parent_model,
+            "codex_parent_model": args.codex_parent_model,
             "dependencies": args.dependency,
             "workspace": args.workspace,
         },
@@ -534,9 +552,12 @@ def run(command, *, cwd=output, env=base_env, timeout=600) -> str:
                 "UG_INTEGRATION_CLAUDE_PROVIDER": args.claude_provider,
                 "UG_INTEGRATION_CLAUDE_RELAYED_PROVIDER": args.claude_relayed_provider,
                 "UG_INTEGRATION_CLAUDE_OAUTH_TOKEN": oauth_token,
+                "UG_INTEGRATION_CLAUDE_PROVIDER_MODEL": args.claude_provider_model,
                 "UG_INTEGRATION_CODEX_PROVIDER": args.codex_provider,
                 "UG_INTEGRATION_CODEX_PROVIDER_MODEL": args.codex_provider_model,
                 "UG_INTEGRATION_PARENT_SCHEMA": args.parent_schema,
+                "UG_INTEGRATION_CLAUDE_PARENT_MODEL": args.claude_parent_model,
+                "UG_INTEGRATION_CODEX_PARENT_MODEL": args.codex_parent_model,
                 "UCODE_TEST_WORKSPACE": args.workspace or "",
                 "DATABRICKS_BEARER": bearer,
             }

diff --git a/tests/README.md b/tests/README.md
@@ -38,6 +38,18 @@ All tests live directly in `integration/`; shared mechanics live in `utils/`.
 | `test_ug_configure_codex_databricks` | Configure Databricks Hosted; execute the generated auth helper; open Codex TUI and read a file | Generated helper invokes `ug` with clean token stdout; completed assistant answer contains the file value; normal exit and reopen |
 | `test_ug_configure_codex_openai_mps` | Select OpenAI MPS in the real configure picker; launch Codex | Saved provider in status; completed TUI file task; normal exit |
 | `test_ug_claude_custom_oauth_cli_boots`, `test_ug_codex_custom_oauth_cli_boots` | Launch with `ENABLE_CUSTOM_OAUTH_FROM_CLI=1`, `--workspace`, and `--client-id databricks-cli` | Real TUI reaches a usable prompt, accepts keyboard input, exits normally, and saves `client_id = databricks-cli` in its generated CLI profile; Claude also reads the OS-managed settings and requires a profile-only `apiKeyHelper` |
+| `test_case_13_configured_claude_reuses_saved_model_location` | Configure Claude with a model location, then launch without options | The saved parent supplies the discovered catalog |
+| `test_case_14_configured_codex_reuses_saved_model_location` | Configure Codex with a model location, then launch without options | The saved parent supplies the discovered catalog |
+| `test_case_15_fresh_claude_uses_system_models_when_discovery_disabled` | Launch fresh Claude with discovery disabled | Workspace `system.ai` models appear; Claude Code creates no cache after picker launch |
+| `test_case_16_fresh_codex_uses_system_models_when_discovery_disabled` | Launch fresh Codex with discovery disabled | Workspace `system.ai` models appear without a scoped catalog |
+| `test_case_17_*` | Launch configured and fresh Claude with a provider | Automatic discovery supplies exactly the provider catalog |
+| `test_case_18_*` | Launch configured and fresh Codex with a provider | The provider supplies exactly its model catalog |
+| `test_case_19_*` | Launch configured and fresh Claude with a model location | The explicit parent supplies exactly its picker catalog |
+| `test_case_20_*` | Launch configured and fresh Codex with a model location | The explicit parent supplies exactly its model catalog |
+| `test_case_21_*` | Launch configured and fresh Claude with a provider and discovery disabled | Claude uses native families; Claude Code creates no cache after picker launch |
+| `test_case_22_*` | Launch configured and fresh Codex with a provider and discovery disabled | Codex uses its native catalog and ug writes no scoped catalog |
+| `test_case_23_*` | Launch configured and fresh Claude with a parent and discovery disabled | Claude uses native families; Claude Code creates no cache after picker launch |
+| `test_case_24_*` | Launch configured and fresh Codex with a parent and discovery disabled | Codex uses its native catalog and ug writes no scoped catalog |
 | `test_ug_claude_headless_prompt_argument`, `test_ug_claude_headless_prompt_stdin`, `test_ug_claude_headless_prompt_after_separator` | Run Claude from a script using each prompt form | Structured final answer contains the file value; exit zero; no routing |
 | `test_ug_codex_headless_prompt_argument`, `test_ug_codex_headless_prompt_stdin`, `test_ug_codex_headless_prompt_after_separator` | Run Codex from a script using each prompt form | Completed turn and final answer contain the file value; exit zero; no routing |
 | `test_ug_claude_headless_explicit_model_bypasses_routing` | Pass `--model VALUE` / `--model=VALUE` with routing enabled | Real file task completes; no routing wrapper |
@@ -64,16 +76,16 @@ All tests live directly in `integration/`; shared mechanics live in `utils/`.
 | `test_ug_and_ucode_auth_helpers_emit_only_the_supplied_bearer` | Run both auth helper commands with the public bearer override, with and without forced refresh | Exact token-only stdout, no warnings or ANSI escapes; no workspace authentication or saved state |
 | `test_ug_and_ucode_web_search_helpers_preserve_mcp_stdio` | Initialize and list tools through both web-search helper commands | Exactly the MCP JSON-RPC responses; no text/ANSI contamination; existing server/tool identities preserved; no model request |
 
-With both agents selected there are **42 live cases** (6 interactive TUI cases),
-**4 managed-workspace cases** (marker `managed`, run against a separate workspace that
-publishes a CodingAgentConfig), **28 managed-fixture cases** (marker `managed_fixture`), and
-**5 installation checks**. The 12 numbered scenarios fetch the published config once per agent,
-replace that agent's static model source with its dedicated test MPS, drop its incompatible static
-defaults, and reuse the result across 24 explicit configured/fresh journeys. The other four
-collected cases, from three test functions, inject focused model and MCP shapes. Parametrization
-varies argument spelling or routing mode, never hides the
-agent/provider in the test name. Duplicate boot-only cases are incorporated into the Databricks
-configuration TUI journeys.
+With both agents selected there are **62 live cases** (16 interactive TUI cases),
+**4 un-stubbed managed-workspace cases** (marker `managed`), **28 managed-config fixture cases**
+(marker `managed_fixture`), and **5 installation checks**. The 24 numbered scenarios comprise
+**44 explicit journeys**, and all groups total **99 executions**. Cases 1–12 fetch the published
+config once per agent, replace that agent's static model source with its dedicated test MPS, drop
+its incompatible static defaults, and reuse the result across 24 configured/fresh journeys. The
+other four managed-fixture executions inject focused model and MCP shapes. The un-stubbed cases
+retain coverage of the config fetch/wire contract. Parametrization varies argument spelling or
+routing mode, never hides the agent/provider in the test name. Duplicate boot-only cases are
+incorporated into the Databricks configuration TUI journeys.
 Generated-file cleanup and strict app-server stdout assertions remain enforced.
 
 ug no longer runs a post-configure agent probe; the deprecated `--skip-validate`
@@ -94,7 +106,7 @@ dependency graph to reproduce a user's combination. Every relevant same-reposito
 PR and push to `main` runs both smoke and the full CUJ suite. Smoke covers the
 Databricks Hosted configure/TUI, custom OAuth CLI TUI, and headless argument
 journeys for both agents, in two parallel jobs. After smoke finishes, the full
-suite runs all 41 cases across two parallel agent jobs: one Claude VM and one
+suite runs all 62 live cases across two parallel agent jobs: one Claude VM and one
 Codex VM, each running its configure, headless, and commands/lifecycle cases
 serially. Each agent is installed once for the full suite, and no two full jobs
 for the same agent overlap within a run.
@@ -126,7 +138,8 @@ pending. The descriptive jobs provide the actual coverage and diagnostics.
 | --- | --- |
 | Live MCP and skills functionality | Deferred; installation tests cover the local web-search MCP handshake and tool listing, not upstream proxying or a real search request |
 | Broad configure flags, tracing, multiple workspaces, and PAT flows | Deferred while focusing on basic CUJs |
-| Provider switching, relayed/subscription MPS | Not covered by the four provider journeys |
+| Relayed/subscription MPS discovery | Not covered by the scoped discovery journeys |
+| Fresh provider/parent validation and mixed Bedrock filtering | Not covered after removing the duplicate model-discovery suites |
 | TUI initial prompt supplied on the launch command line | Not yet covered; headless prompt arguments are covered |
 | Follow-up turns and conversation resume | Not covered; reopen proves startup, not conversation resume |
 | Claude/Codex interactive smart routing | Deferred at the user's request; routing jobs and live journeys removed. Unit/component routing tests remain, but do not establish live routing behavior. |

diff --git a/tests/integration/README.md b/tests/integration/README.md
@@ -100,6 +100,8 @@ test_ug_configure_claude_lifecycle.py   # repeat setup, revert, rejected credent
 test_ug_configure_codex_lifecycle.py    # repeat setup, revert, rejected credentials
 test_ug_claude_managed_model_discovery.py # fetched/reused Claude MPS policy cases
 test_ug_codex_managed_model_discovery.py  # fetched/reused Codex MPS policy cases
+test_ug_claude_model_discovery.py       # Tests-tab cases 13, 15, 17, 19, 21, 23
+test_ug_codex_model_discovery.py        # Tests-tab cases 14, 16, 18, 20, 22, 24
 test_ug_configure_managed.py            # managed workspace: static model list, no agent selector
 test_ug_configure_managed_models.py     # injected model lists: pickers and Codex fallback metadata
 test_ug_configure_managed_mcp.py        # injected managed MCP list
@@ -151,27 +153,31 @@ MPS CUJs select the existing services already used by e2e:
 - Codex: `main.ucode.ci_openai_mps`, using its allowed `gpt-5-nano` model.
 
 Use `--claude-provider` / `--claude-relayed-provider` / `--codex-provider` to
-reproduce another existing service. Use `--codex-provider-model` when that OpenAI
-service allows a different model. Those choices are recorded in `versions.json`.
+reproduce another existing service. Use `--claude-provider-model` /
+`--codex-provider-model` when it allows a different model. Those choices are recorded in `versions.json`.
 No service is created or modified. A missing service, permission, or OAuth token
 fails the selected CUJ, rather than skipping it.
 
-There are **42 live cases** (including 6 TUI journeys) and **5 installation
-checks** with both agents. A separate **4 managed-workspace cases** (one per agent,
-an idempotent re-configure, and cache reuse within the TTL; marker `managed`) run against a
-workspace that publishes a CodingAgentConfig; see "Managed-workspace journeys" below. A further
-**28 `managed_fixture` cases** use `UCODE_MANAGED_CONFIG_STUB`: 24 explicit configured/fresh
-managed-discovery journeys fetch the published config once per agent module, replace that agent's
-static model source with its dedicated MPS, drop its incompatible static defaults, and reuse the
-result; four collected cases from three test functions inject focused model/MCP shapes.
-See the named coverage and gaps matrix in
+Scoped discovery additionally requires Model Services
+`main.ucode.ci_e2e_claude` and `main.ucode.ci_e2e_codex`. Override them with
+`--parent-schema`, `--claude-parent-model`, or `--codex-parent-model`. The tests
+consume but never create or modify them.
+
+There are **62 live cases** (including 16 TUI journeys), **4 un-stubbed managed-workspace cases**
+(marker `managed`), **28 managed-config fixture cases** (marker `managed_fixture`), and
+**5 installation checks** with both agents. The 24 numbered scenarios comprise 44 explicit
+journeys, and all groups total 99 executions. Cases 1–12 fetch the published config once per agent
+module, replace that agent's static model source with its dedicated MPS, drop its incompatible
+static defaults, and reuse it across 24 configured/fresh journeys. Four additional fixture
+executions inject focused model/MCP shapes. The un-stubbed cases retain the config fetch/wire
+contract; see "Managed-workspace journeys" below and the named coverage and gaps matrix in
 [../README.md](../README.md).
 
 ```bash
 # Append one of these selections to the runner command:
 -- -m live         # default: all live user journeys
 -- -m smoke        # six Hosted, custom OAuth CLI TUI, and headless journeys
--- -m tui          # six interactive TUI journeys
+-- -m 'live and tui'  # sixteen interactive live configuration/model-discovery journeys
 -- -k test_ug_codex_app_server_client_initializes  # one named journey and its variants
 # Use --installation-only before -- for package checks without credentials.
 ```
@@ -235,20 +241,20 @@ cannot receive those secrets.
 
 The workspace check requires the secret to match
 `https://eng-ml-inference-team-us-east-1.cloud.databricks.com` (a trailing slash
-is accepted). It never changes the secret or switches workspaces. There is no CI
-model-discovery or model-selection job. Real `ug configure` performs its normal
-workspace discovery inside each test; only explicit-model scenarios choose and
-record a discovered `system.ai` model as a test argument.
+is accepted). It never changes the secret or switches workspaces. There is no
+separate CI model-selection job; the full agent lanes include scoped model
+discovery. Real `ug configure` performs its normal workspace discovery inside
+each test; only explicit-model scenarios choose and record a discovered
+`system.ai` model as a test argument.
 Every same-repository PR and push to `main` runs **Smoke journeys**, followed by
-**Full journeys** even if smoke fails. Smoke runs the Hosted configure/TUI,
-headless argument, and custom OAuth CLI TUI journeys for each agent (six cases,
-two agent jobs). Full runs all 42 live cases, including those smoke cases, in two
-disjoint agent lanes:
+**Full journeys** even if smoke fails. Smoke runs the Hosted configure/TUI, headless argument,
+and custom OAuth CLI TUI journey for each agent (six cases, two agent jobs). Full runs all 62
+live cases, including those smoke cases, in two disjoint agent lanes:
 
 | Agent lane | Marker | Cases |
 | --- | --- | --- |
-| Claude | `live and claude` | 17 |
-| Codex | `live and codex` | 25 |
+| Claude | `live and claude` | 27 |
+| Codex | `live and codex` | 35 |
 
 Each lane installs only its agent CLI, once, and runs all its configure, headless,
 commands, lifecycle, and applicable app-server journeys. Cases remain serial
@@ -329,7 +335,7 @@ The workflow consumes the stored bearer; it does not mint or refresh credentials
 For a manual run, use **Actions → Integration → Run workflow**, select the branch,
 and choose `full` (default), `smoke`, `tui`, or `installation`. `live` remains an
 alias for `full`. Manual subsets are explicit: `smoke` runs just the six smoke
-cases; `tui` adds `and tui` to each agent lane's marker and runs all six TUI cases. Installation
+cases; `tui` adds `and tui` to each agent lane's marker and runs all 16 live TUI cases. Installation
 checks always run. Set the ug/agent versions. From the CLI:
 
 ```bash
@@ -500,7 +506,7 @@ uv run --no-project --python 3.12 python scripts/run_integration.py \
 unset DATABRICKS_BEARER
 ```
 
-This runs all 42 live cases. For the five installation checks, run the same
+This runs all 62 live cases. For the five installation checks, run the same
 runner/version/index arguments with `--installation-only` and omit `-- -m live`;
 no bearer or workspace is needed. Results remain under `.integration-runs/`.
 Each invocation needs a new output directory; an existing one is rejected.

diff --git a/tests/integration/conftest.py b/tests/integration/conftest.py
@@ -106,6 +106,11 @@ def claude_oauth_token():
     return token
 
 
+@pytest.fixture(scope="session")
+def claude_provider_model():
+    return os.environ["UG_INTEGRATION_CLAUDE_PROVIDER_MODEL"]
+
+
 @pytest.fixture(scope="session")
 def codex_provider():
     return os.environ["UG_INTEGRATION_CODEX_PROVIDER"]
@@ -119,3 +124,13 @@ def codex_provider_model():
 @pytest.fixture(scope="session")
 def parent_schema():
     return os.environ["UG_INTEGRATION_PARENT_SCHEMA"]
+
+
+@pytest.fixture(scope="session")
+def claude_parent_model():
+    return os.environ["UG_INTEGRATION_CLAUDE_PARENT_MODEL"]
+
+
+@pytest.fixture(scope="session")
+def codex_parent_model():
+    return os.environ["UG_INTEGRATION_CODEX_PARENT_MODEL"]