[MarioCadenas]

What

Adds the new Databricks Apps model_service resource (a Unity Catalog securable of type MODEL_SERVICE, from databricks-eng/universe#2571640) and makes it the agents plugin's default-model resource.

Changes

• Resource type: model_service in resourceTypeSchema, permission EXECUTE. In databricks.yml it binds as uc_securable with securable_type: MODEL_SERVICE, mapping name to securable_full_name. Generated types, JSON schemas, plugin create defaults and docs updated.
• Agents plugin: the optional default-LLM resource is now model_service (agents-model-service, env DATABRICKS_MODEL_SERVICE_NAME, value catalog.schema.name). The agents plugin and standalone runAgent read that env first and still fall back to DATABRICKS_SERVING_ENDPOINT_NAME, so existing apps keep working.
• Routing: adapterFromModelString sends any dotted name to the AI Gateway (fromAiGateway), not only system.*. Model services are catalog.schema.name, and serving endpoint names cannot contain dots, so no existing endpoint changes route. To force a dotted name onto a serving endpoint, pass a pre-built adapter.
• appkit plugin add picker: opens on the system.ai model services, with "Browse other catalogs..." for a catalog, schema, model service drill-down. Uses databricks ai-gateway list-model-services, so it needs a CLI that has that command; otherwise it falls back to manual entry. --set / --from-plan accept any catalog.schema.name.
• appkit doctor: existence check for model_service (GET /api/2.1/unity-catalog/model-services/{name}).

Not changed

• The serving plugin stays on serving_endpoint. It is deprecated, and it calls /serving-endpoints/{name}/invocations, which a model service is not.
• No SCOPE_BY_TYPE entry for model_service, so the agents resource no longer carries a user API scope. Agents that run on behalf of the user may need ai-gateway; not added until confirmed.

Why the picker is a drill-down

A flat walk over every catalog and schema did not finish in 10 minutes on dogfood (1158 catalogs), and finding system by paging is slow, so the picker opens on system.ai and leaves the rest one pick away.

Testing

• Unit tests: schema and binding, picker steps (open on system.ai, browse, --set), doctor check, routing, agents env default.
• pnpm check, pnpm -r typecheck, pnpm test, docs build: green.
• Live on dogfood: doctor passes for system.ai.gpt-5-mini and reports NOT_FOUND for a made-up name; the system.ai listing returns 70 services; runAgent with DATABRICKS_MODEL_SERVICE_NAME=system.ai.gpt-5-mini answered through the AI Gateway.

Task: KIT-34 (app-kit/34)

This pull request and its description were written by Isaac.

[github-actions]

📦 Bundle size report

Compared against bundle-size-baseline.json (main).

@databricks/appkit

npm tarball (packed): 1.2 MB (+262 B) — gzipped download (dist + bin; excludes release-only docs/NOTICE).

dist	raw	gzip
JS (runtime)	1.3 MB (+520 B)	460 KB (+112 B)
Type declarations	468 KB (+3.3 KB)	169 KB (+70 B)
Source maps	2.5 MB (+995 B)	859 KB (+80 B)
Other	11 KB	3.7 KB
Total	4.2 MB (+4.8 KB)	1.5 MB (+262 B)

Per-entry composition (own code — deps external (as shipped))

Entry	Initial (gz)	Lazy (gz)	Total (gz)	node_modules (min)	Own code (min)
.	106 KB (+57 B)	2.5 KB	109 KB (+57 B)	external	344 KB (+234 B)
./beta	103 KB (+79 B)	483 B (+1 B)	104 KB (+80 B)	external	310 KB (+341 B)
./testing	45 KB (+43 B)	32 KB	77 KB (+43 B)	external	224 KB (+232 B)
./tsdown	520 B	0 B	520 B	external	813 B
./type-generator	26 KB	0 B	26 KB	external	74 KB

Chunks:

Entry	Chunk	Load	Size (gz)
.	index.js	initial	102 KB
.	utils.js	initial	4.6 KB
.	remote-tunnel-manager.js	lazy	2.5 KB
./beta	beta.js	initial	83 KB
./beta	modular.js	initial	6.0 KB
./beta	stream-manager.js	initial	5.9 KB
./beta	service-context.js	initial	4.2 KB
./beta	databricks.js	initial	3.3 KB
./beta	client.js	initial	647 B
./beta	index.js	initial	20 B
./beta	supervisor-api.js	lazy	192 B
./beta	databricks.js	lazy	177 B
./beta	index.js	lazy	114 B
./testing	manifest.js	initial	32 KB
./testing	index.js	initial	10 KB
./testing	wide-event-emitter.js	initial	2.9 KB
./testing	index.js	lazy	28 KB
./testing	remote-tunnel-manager.js	lazy	2.5 KB
./testing	utils.js	lazy	1.8 KB
./tsdown	index.js	initial	520 B
./type-generator	index.js	initial	26 KB

@databricks/appkit-ui

npm tarball (packed): 397 KB — gzipped download (dist + bin; excludes release-only docs/NOTICE).

dist	raw	gzip
JS (runtime)	429 KB	147 KB
Type declarations	267 KB	98 KB
Source maps	853 KB	286 KB
CSS	16 KB	3.2 KB
Total	1.5 MB	534 KB

Per-entry composition (consumer bundle — deps bundled, peerDeps external)

Entry	Initial (gz)	Lazy (gz)	Total (gz)	node_modules (min)	Own code (min)
./js	5.3 KB	49 KB	55 KB	208 KB	14 KB
./js/beta	2.1 KB	0 B	2.1 KB	0 B	5.0 KB
./react	433 KB	49 KB	481 KB	1.3 MB	177 KB
./react/beta	4.9 KB	0 B	4.9 KB	0 B	12 KB

Chunks:

Entry	Chunk	Load	Size (gz)
./js	index.js	initial	5.2 KB
./js	chunk	initial	120 B
./js	apache-arrow	lazy	49 KB
./js/beta	beta.js	initial	2.1 KB
./react	index.js	initial	431 KB
./react	tslib	initial	2.1 KB
./react	apache-arrow	lazy	49 KB
./react/beta	beta.js	initial	4.9 KB

[github-actions]

🤖 AppKit PR bot

🔬 Run evals

Start an eval for this PR from the evals-monitor app: Go to Evals Monitor →

📦 Try this PR's app template

Scaffolds a new app from this PR's SDK build. Run it in any folder (requires the GitHub CLI — gh auth login — and the Databricks CLI):

gh run download 41849434868 -R databricks/appkit -n appkit-template-0.90.0-pr.8907544-work-34-add-model-service-resource-rep-705 -D appkit-pr-705 \
  && unzip -o "appkit-pr-705/appkit-template-0.90.0-pr.8907544-work-34-add-model-service-resource-rep-705.zip" -d "appkit-pr-705" \
  && databricks apps init --template "appkit-pr-705"

The template pins @databricks/appkit and @databricks/appkit-ui to tarballs built from this branch, so the scaffolded app runs against this PR's code.

[MarioCadenas]

Review 6 · Fail · 1 critical, 1 minor

Does the change do what the task asked? No. Gate: major and worse must be fixed.

The model_service resource type, its uc_securable MODEL_SERVICE/EXECUTE binding, the picker, the doctor check and the agents plugin switch are built, and the 262 related tests pass. The serving plugin is still on serving_endpoint, but the task says plainly to replace it in both plugins. The closing note cites a human decision to keep it, but there is no human message in the task log that confirms this.

Required

• critical: The serving plugin still declares serving_endpoint, not model_service, but the task says to replace it in both the serving and agents plugins. (packages/appkit/src/plugins/serving/manifest.json:11)
  Proof: The manifest has "type": "serving_endpoint" at line 11; the diff since 05c51e2 has no file under plugins/serving; said-by.jsonl has no human message saying serving stays as is (only the agent's closing-note says so). If that human decision is real, downgrade this.

Not required (below the gate)

• minor: Model services are listed by drilling down through the CLI rather than through a paginated ListModelServices in the workspace client, which is what item 2 asked for. (packages/shared/src/cli/commands/registry/workspace-picker.ts)
  Proof: The diff stat shows no file under packages/shared/src/workspace-client.

[MarioCadenas]

Review 7 · Pass · 1 minor, 1 trivial

Does the change do what the task asked? Yes. Gate: major and worse must be fixed.

model_service is added to the schema, DABs binding (uc_securable, MODEL_SERVICE, EXECUTE), doctor, the CLI picker (catalog > schema > ai-gateway list-model-services, prefix stripped) and the agents manifest and runtime env fallback, with tests that pass. The serving plugin staying on serving_endpoint and the CLI-based lister are human-settled decisions, so I don't count them as gaps.

Not required (below the gate)

• minor: Default-model routing widened from the system.* prefix to any dotted string, so any dotted value in DATABRICKS_SERVING_ENDPOINT_NAME or model now goes to AI Gateway instead of Model Serving. (packages/appkit/src/agents/databricks.ts adapterFromModelString)
  Proof: Diff: model.startsWith("system.") became model.includes("."); serving endpoint names cannot contain dots, so the impact is limited to invalid inputs.
• trivial: run-agent's no-model fallback still surfaces the DATABRICKS_SERVING_ENDPOINT_NAME error text, while the agents plugin now names DATABRICKS_MODEL_SERVICE_NAME. (packages/appkit/src/core/agent/run-agent.ts resolveAdapter)
  Proof: Read the diff: the comment and fromModelServing error path are unchanged.