Conversation
Land the CodingAgentConfig parsing and the version and watermark helpers on their own, ahead of any change to launch or configure behavior, so the data-shape change is reviewable in isolation. Nothing here is called from the launch path yet; the stacked launch-apply PR wires it in. - managed_config: read the extended config shape (enabled_agents with per-agent AgentModels: names / model_service_location / provider, default_model, default_alias_models, http_headers; spend_tiers with a budget_policy fallback; update_time). Add managed_config_is_newer and managed_update_time, a spec_version forward-compat gate (a newer or malformed spec_version keeps the last-known-good cache), and the UCODE_MANAGED_CONFIG_STUB local-file read for pre-server testing. - mcp_servers and skills each parse to a names-or-location selector: an absolute names list of UC FQNs, or a unity_catalog_location (a catalog.schema) whose contents are discovered, enforced either/or. Both are UC mcp-services / skills only; there is no per-server type model. managed_setup serializes and validates the same shape (names are 3-part FQNs, location a 2-part catalog.schema, exactly one set), so serialize then normalize round-trips. - state: add get/set_applied_managed_update_time, the applied-version watermark the launch gate reads. - managed_export / managed_setup: parse and export updates; spec_version is server-owned on export so a supplied config cannot override the envelope value. - databricks: switch the managed-config update mask to spend_tiers; rename the user-facing "Unity AI Gateway" strings to "Unity Gateway". Co-authored-by: Isaac <no-reply@databricks.com>
With the config parsed (the stacked parse PR), apply the managed configuration to every enabled agent at launch, but only when it changed since the last apply (an update_time watermark) or on --refresh, so an unchanged launch never re-writes the OS-managed files or prompts for a password. Also remove the post-configure probe validation, wire the managed MCP servers and skills, and rename user-facing strings. - Gate the launch apply on update_time. refresh_managed_config always fetches; a launch re-applies the managed configuration (the only step that writes the OS-managed files and can prompt for a password) only when update_time is newer than the applied watermark, or on ug <agent> --refresh, and records the new watermark. An unchanged launch skips the apply and never prompts. ug configure always applies and records the watermark. The "was updated; re-applying it" note prints only on a genuine version change, not on a no-op --refresh. The per-tool launch write is wrapped in suppressed_managed_writes() so the apply-all owns the single OS write. - Apply the managed config's mcp_servers and skills to each enabled agent. Both are UC-services-only selectors: register the absolute names (as UC mcp-services) or download the named skills, plus everything discovered under a unity_catalog_location (a schema's mcp-services via list_mcp_services, or its skills), unioned. Nothing on the launch path turned the workspace's published servers or skills into agent registrations before, so they never showed up. - `ug configure` wires the managed config's MCP servers and skills at configure time (not only at launch), and suppresses the interactive MCP/skills/plugins prompt when the managed config already defines them (and never installs anything under --dry-run), so a managed workspace is set up without asking the developer to configure MCP. - Stop ug configure from validating agents with a probe message. Remove the post-configure validation and the --skip-validate flag: it is not configure's job to test models, the probe also reverted a freshly-written config on failure, and the gateway enforces model access regardless. The now-unused validate_all_tools / provider_permission_error / resolve_managed_for_tool helpers are removed; validate_tool stays for headless and e2e callers. - Drop the "Using managed config - launching <agent>" banner and the redundant post-panel "Starting <agent>" line from the bare-ug launch, so bare ug matches ug <agent>. - When a config we previously applied is gone (deleted, or the feature turned off), the launch does not auto-revert the OS-managed files; it notes that Unity Gateway managed configurations are still applied and points the developer to `ug revert`, but only when the config is definitively gone (NOT_FOUND or the feature is off), not on a transient fetch failure. Keyed on the applied watermark, so a workspace that never had a config still just uses the developer's own settings. - Rename the product string "CLI Managed Configuration" to "managed configuration" (lowercase) across user-facing lines, tighten the budget-tier recommendation line, and remove the --skip-validate docs. Co-authored-by: Isaac <no-reply@databricks.com>
Resolve the managed config's per-agent inputs into each agent's own config shape: the model allow-list or discovery location into Claude's /model picker and Codex's model catalog, and the custom http_headers into each agent's outbound requests. Stacked on the launch-apply PR (#593). - managed_resolve: add managed_static_models, managed_model_service_location, and managed_custom_headers; extend managed_state_overrides to layer claude_static_models / codex_static_models, the per-agent model_service_location, and {tool}_custom_headers into state. For codex a discovery location alone does not count as supplying models (no reader for it), so discovery still runs; for claude it does. - claude: a static list becomes availableModels + enforceAvailableModels + a modelPicker (exactly the named ids, non-Claude included); a discovery location turns on gateway discovery instead. Ownership markers prune only ug-owned picker keys on a static-to-discovery transition, and clear only when the OS-managed file was written. Managed http_headers append to ANTHROPIC_CUSTOM_HEADERS, with ug's fixed headers winning on conflict; values that would break the header serialization (newlines, and commas for Gemini) are dropped. - codex: a static list is written as a model_catalog_json catalog of full ModelPresets; a fetch failure preserves an existing catalog (or falls back to discovery when none exists yet), and a successful but empty result falls back to discovery. Managed http_headers merge into the provider http_headers, ug's fixed headers winning. - gemini: managed http_headers merge into GEMINI_CLI_CUSTOM_HEADERS, ug's fixed headers winning. - README: document the Claude picker, the Codex catalog, and the ucode-models.json managed file. Co-authored-by: Isaac <no-reply@databricks.com>
david-siqi-liu
force-pushed
the
david/AIGTWY-4573-model-picker
branch
11 times, most recently
from
September 15, 2026 13:42
329ebdd to
040523f
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Stacked on
david/AIGTWY-4573-model-picker(one additional commit).Replace the static-model catalog's authenticated
/ai-gateway/codex/v1/modelsfetch with client-side construction, inspired by databricks-eng/universe#2591694:debug models --bundledin an isolated temporary home, and validate the generated catalog with the same binary before writing configuration.none, native search disabled). Apply explicit capability overrides for known GLM and Kimi models.Validation
uv run --frozen pytest -q --ignore=tests/test_e2e_user_agent.py --deselect=tests/test_claude_smart_routing_v2.py::TestV2Launch::test_restores_model_captured_immediately_before_switchruff check .,ruff format --check src/ tests/,ty check src/, andgit diff --checkpass.The full-suite attempt exposed an unrelated Claude route display-name assertion (reproduced on the unchanged base branch) and two host-dependent
test_e2e_user_agent.pyfailures: the installed/managed Claude and Codex configurations prevented requests reaching the tests' localhost capture servers. The passing regression run excludes that E2E file and the confirmed baseline assertion.No dependency or lock-file changes.
This PR was created with GitHub MCP.