[xi-db]

🥞 Stacked PR

Use this link to review incremental changes.

• stack/copilot-auth-proxy [Files changed] [MERGED]
  • stack/copilot-model-discovery [Files changed] [MERGED]
    • stack/copilot-provider-services [Files changed] [MERGED]
      • stack/copilot-model-location [Files changed]
        • stack/copilot-foundry [Files changed] ← this PR

---

What did you change, and why?

Change: ug copilot --provider <catalog>.<schema>.<name> now accepts Microsoft Foundry (microsoft_foundry) Model Provider Services, as Codex already does. Before, Copilot refused them ("can't route to ... microsoft_foundry").

• Wire per target. Foundry hosts more than OpenAI models (Claude, DeepSeek, GPT), and its targets declare only openai/v1/chat/completions, including the Claude ones.
  • Claude targets use the Anthropic wire (dbx-anthropic on /ai-gateway/anthropic), whatever they declare. Foundry answers Claude on Anthropic Messages and Responses but returns 404 on Chat Completions. On the Anthropic wire Copilot also sends its Claude request settings, and limits are capped at 200K as for other provider services.
  • GPT 5+ and o-series targets use Responses, and everything else uses Chat Completions, on /ai-gateway/openai/v1, as for OpenAI and Azure OpenAI services.
• Selection. ug configure now offers Foundry services for Copilot under External Models, and a managed model_provider_service of this type applies to Copilot.

How do you know it works?

Testing:

• Unit:
  • classifier: a Foundry Claude target on Anthropic, gpt-5-nano on Responses, DeepSeek on Chat Completions;
  • catalog: Foundry targets that declare only Chat Completions keep Claude on Anthropic;
  • service_usable_for_tool and supports_provider_service accept microsoft_foundry for Copilot.
  • just lint is clean.
• Integration (tests/integration/test_ug_copilot_providers.py, non-blocking Copilot lane): a file task through the existing Foundry MPS fixture on gpt-5-nano (the one Codex's Foundry journey uses), on dbx-responses, with the answering model checked from Copilot's session events.
• Live (ug from this branch, three Foundry services on a Databricks workspace, API-key and Entra auth):

Foundry target	Copilot wire	Result
claude-opus-4-8, and the default start model	Anthropic	✅ OK
deepseek	Chat Completions	✅ OK
gpt-5.4-nano	Responses	✅ OK

A direct request per gateway route for the same targets: Claude answered on Anthropic Messages and Responses, and 404'd on Chat Completions; DeepSeek answered on Chat Completions and Responses, and 404'd on Anthropic Messages.

Stacked on #1256.

This pull request and its description were written by Isaac.

[github-actions]

Polly AI Review

Blocking issues

• P2 — first-run location launch fails without global models. can we forward the selected schema through Copilot bootstrap? a valid custom location currently fails global availability before discovery.
  Evidence: src/ucode/cli.py:4147 introduces the location handoff, but bootstrap at src/ucode/cli.py:3279 forwards only provider. With fresh state and no global Copilot models, configuration rejects the workspace before scoped discovery. The shared bootstrap gap predates this PR, but this Copilot entry point is new.
  Regression: add tests/test_cli.py::TestAutoConfigureOnFirstRun.test_first_run_configures_through_launch_model_location, retaining real bootstrap and availability checks with an empty global catalog.

Security vulnerabilities

None identified in the changed code. This is a static assessment, not live security validation.

Non-blocking notes

• can we parameterize test_copilot_provider_and_model_location_are_mutually_exclusive with its existing Claude/Codex counterparts, retaining both assertions?
  tests/test_cli.py:1259
• is Copilot model-location support intentionally included from the parent stack? it adds configure, persistence, and launch behavior beyond Foundry-provider acceptance.
  src/ucode/cli.py:4133, src/ucode/agents/copilot.py:388

Approach

The Foundry approach is consistent with existing provider-support and wire-classification patterns. No materially simpler implementation is apparent. The location feature also reuses existing discovery; its bootstrap handoff needs completion.

Summary

Diff size (computed): 14 files; +498 / -55 lines (553 changed text lines); 0 binary files.

Category	Files	Added	Deleted	Share of changed text lines
Tests and test support	7	371	12	69.3%
Documentation	1	8	5	2.4%
Dependencies and lockfiles	0	0	0	0.0%
Implementation / other	6	119	38	28.4%

Shares use additions + deletions, including generated text and lockfiles. Binary files have no line count. Tests include fixtures/helpers under test paths and colocated test/spec files; categories are path-based, not coverage measurements.

Changes

The intended outcome is to allow Copilot to use Microsoft Foundry provider services, selecting Anthropic for Claude, Responses for GPT-5+/o-series, and Chat Completions otherwise.

The diff contains implementation, unit/component tests, live integration scenarios, and documentation. It also includes stacked Copilot model-location support: schema discovery, saved and managed source selection, and transient launch state. No dependency or generated-file changes are included.

The Foundry changes are proportional to the outcome. The location work explains much of the additional breadth, subject to the scope question above. The concrete reduction opportunity is shared CLI conflict-test setup rather than removing distinct lifecycle coverage.

Tests

Test-by-test assessment

Test / case	Behavior protected	Layer	Needed?	Action / rationale
tests/integration/README.md — journey descriptions, inventory, fixture entry, and count updates	Documents execution inputs and coverage	Documentation	keep	Supports the added journeys; collected-case totals were not independently verified.
tests/integration/test_ug_copilot_model_location.py::test_ug_copilot_runs_a_task_on_a_model_location	Fresh CLI launch, selected full model name, inference, cleanup	Live process/transport	keep	Exercises the real discovery-to-Copilot boundary using existing primitives. Its workspace does not isolate absence of global models.
tests/integration/test_ug_copilot_providers.py::test_ug_copilot_runs_a_task_through_a_microsoft_foundry_provider	Foundry GPT inference through Responses, model correlation, cleanup	Live process/transport	keep	Establishes the external Foundry transport boundary, beyond mocked classification.
tests/test_agent_copilot.py::_stub_location, nested anthropic_catalog, and AnthropicModelCatalog import	Explicit catalog inputs and schema observation	Unit support	keep	Small reusable discovery seam; no additional global fixture machinery.
tests/test_agent_copilot.py::TestClassifyProviderModel.test_provider_type_and_id_decide_the_wire — Foundry Claude, GPT-5-nano, DeepSeek parameters	Anthropic, Responses, and Chat Completions classification	Unit	keep	Places the provider/model matrix at the lowest reliable layer.
tests/test_agent_copilot.py::TestBuildProviderCatalog.test_foundry_claude_uses_anthropic_whatever_it_declares	Chat-only declarations cannot misroute Foundry Claude; mixed catalog remains usable	Unit	keep	Protects the specific declaration-override regression beyond name-only classification.
tests/test_agent_copilot.py::TestLocationCatalog.test_registers_the_schemas_services_by_full_name	Alias removal, full names, wire selection, filtering, schema forwarding	Unit	keep	Covers discovery transformation not supplied by provider-catalog tests.
tests/test_agent_copilot.py::TestLocationCatalog.test_an_empty_schema_says_so	Empty discovery produces a location-specific error	Unit	keep	Covers a material failure path without live setup.
tests/test_agent_copilot.py::TestLaunch.test_model_location_launch_registers_the_schemas_services	Registry composition, stale-model override, no provider-service header, token cap, start model, marker consumption	Component	keep	Protects launch composition beyond isolated catalog conversion.
tests/test_cli.py::TestSubcommandRouting.test_copilot_launch_uses_saved_model_location	Saved location reaches launch	CLI component	keep	Covers persistence-to-launch handoff.
tests/test_cli.py::TestSubcommandRouting.test_copilot_explicit_provider_overrides_saved_location	Explicit provider suppresses saved location	CLI component	keep	Checks source precedence and absence of a competing marker.
tests/test_cli.py::TestSubcommandRouting.test_copilot_explicit_model_location_overrides_saved_provider	Explicit location suppresses saved provider	CLI component	keep	Protects the opposite source transition.
tests/test_cli.py::TestSubcommandRouting.test_copilot_saved_model_location_is_ignored_under_managed_config	Developer state cannot override managed configuration	CLI component	keep	Covers administrator/developer configuration ownership.
tests/test_cli.py::TestSubcommandRouting.test_copilot_provider_and_model_location_are_mutually_exclusive	Conflicting source flags fail with actionable output	CLI component	consolidate	Parameterize with the existing Claude/Codex cases, preserving exit-code and error assertions.
tests/test_cli.py::TestSubcommandRouting.test_invalid_model_location_is_rejected[copilot]	Malformed schema is rejected	CLI component	keep	Extends the existing input matrix without new setup.
tests/test_cli.py::TestClaudeModelFlag.test_model_location_sets_transient_copilot_launch_marker	Explicit location reaches launch without configuring a global model	CLI component	keep	Checks option/configure plumbing, but mocked setup cannot expose the bootstrap failure.
tests/test_cli.py::TestCopilotLaunchModel.test_managed_default_is_the_launch_model_under_a_managed_source — both source parameters	Managed default survives provider-service and location selection	CLI component	keep	Both named-source branches need the launch-model handoff.
tests/test_cli.py::TestCopilotLaunchModel.test_managed_source_overrides_saved_provider[mps] and [uc-parent]	Managed sources supersede saved sources and bypass global discovery	CLI component	keep	Retains distinct source-specific marker and ownership assertions.
tests/test_cli.py::TestConfigureAgentsSelection.test_picker_offers_location_for_copilot	Picker offers location and clears the competing provider	Component	keep	Covers the newly enabled configure entry point.
tests/test_cli.py::TestConfigureAgentsSelection.test_managed_copilot_parent_is_passed_to_generic_configure	Multi-agent managed configuration forwards the schema	CLI component	keep	Protects the generic configure entry point against global-availability dependence.
tests/test_cli.py::TestConfigureAgentsSelection.test_single_copilot_agent_passes_managed_parent_directly	Single-agent configuration applies managed schema and refresh ownership	CLI component	keep	Separate public entry point with additional refresh and displayed-source assertions.
tests/test_databricks.py::TestServiceUsableForTool.test_copilot_routes_anthropic_and_openai_services[microsoft_foundry]	Shared filtering accepts Foundry for Copilot	Unit	keep	Protects discoverability and validation independently of wire classification.
tests/test_databricks.py::TestServiceUsableForTool.test_copilot_skips_other_provider_types — removed Foundry negative parameter	Remaining unsupported providers stay rejected	Unit	keep	Foundry moves to positive coverage; Gemini Enterprise and custom-provider rejection remain.
tests/test_managed_setup.py::TestProviderServiceSupport.test_copilot_supports_anthropic_openai_bedrock_and_foundry — renamed existing case	Managed setup accepts Foundry and still rejects unsupported providers	Unit	keep	Extends the existing scenario; replaces the obsolete Foundry rejection with Gemini Enterprise.

Existing unchanged tests cover provider headers, native endpoints, bearer handling, Claude token limits, catalog failures, model fallback, and launch cleanup. These provide useful coverage without duplicating every assertion in live tests.

The concrete missing regression is the first-run location path identified above. Existing first-run tests mock configuration and therefore miss its availability failure. A Copilot-only component test with real bootstrap is sufficient; another live scenario is not required for that bug.

The new Foundry live scenario exercises GPT/Responses, not Claude or DeepSeek transport. Classification and declaration-override assertions cover those routing decisions locally; the reported manual transport checks were not independently repeated.

The frozen snapshot checksum and both required policy files were verified. No tests, collection, lint, or typecheck were executed; no passing result is claimed.

To confirm the bootstrap fix manually, use an isolated fresh UG profile and a workspace with no global Copilot models but a usable service in the selected schema:

ug copilot --workspace <workspace> --model-location <catalog.schema> --model <service-full-name> -p "reply ok"

The selected schema should permit startup instead of producing the global availability error.

Scope

Foundry acceptance, classification, selection, and associated documentation support the stated outcome. Copilot model-location support is coherent with the parent-stack context but is not explicitly included in the Foundry description; its inclusion needs clarification. No clearly unrelated changes were established.

---

_{Automated review by Polly · workflow run}

[xi-db]

Thanks for the review.

• P2, first-run --model-location without global models: leaving this as is. _auto_configure_tool is shared, so ug claude --model-location and ug codex --model-location hit the same availability check on a fresh setup today; Copilot matches them. Fixing the bootstrap for all three agents is better as its own change than inside the Foundry PR.
• Parametrize the --provider + --model-location test: done in e9f4e4e (on copilot: run ug copilot on a Unity Catalog model location #1256): one test now covers Claude, Codex and Copilot.
• Model-location scope: that comes from the parent PR, copilot: run ug copilot on a Unity Catalog model location #1256. This PR is stacked on it and targets main, so the diff includes it; the "Files changed" link in the stack section shows only the Foundry change.