feat(appkit): add DatabricksAdapter.fromAiGateway for gateway chat completions

[MarioCadenas]

What

Lets AppKit agents consume models through the Databricks AI Gateway Chat Completions endpoint (/ai-gateway/mlflow/v1/chat/completions), naming the model in the request body (a UC system.* name, e.g. system.ai.claude-opus-5-5) instead of a per-endpoint URL. This is the migration path off the legacy databricks-* serving-endpoint names.

Two commits:

1. fromAiGateway adapter — DatabricksAdapter.fromAiGateway({ model }) explicit factory.
2. Automatic routing — a string model that starts with system. resolves to the gateway; anything else resolves to Model Serving as before. So the common case needs no factory:

createAgent({
  instructions: "…",
  model: "system.ai.claude-opus-5-5",   // -> AI Gateway automatically
});

createAgent({
  instructions: "…",
  model: "my-custom-endpoint",          // -> /serving-endpoints/my-custom-endpoint/invocations
});

How

The gateway speaks the same OpenAI Chat Completions wire format the adapter already used, so the change is deliberately small:

• connectors/serving/client.ts — new streamAiGateway(): hard-codes the gateway path, forces stream: true (mirrors the existing stream() serving helper; keeps streamPath's @internal hard-coded-path contract).
• agents/databricks.ts — optional model on the adapter options, injected into the body only when set; fromAiGateway factory; and adapterFromModelString() — the single decision point that both string-model resolution sites now share (plugins/agents/agents.ts + core/agent/run-agent.ts), so they can't drift.

Everything else is reused unchanged: the multi-step tool loop, SSE parsing, dot→__ tool-name mapping, thought-signature pass-through, DoS caps, generation params.

Non-breaking

• model is set on the request body only on the gateway path; fromModelServing / fromServingEndpoint requests are byte-identical to before (test-pinned).
• Serving-endpoint names are [a-zA-Z0-9_-] (no dots), so the system. prefix never matches an existing endpoint name — no current string changes routing. A system.* string never worked as a serving-endpoint name anyway.
• Escape hatch: to force a system.* name onto a serving endpoint, pass a pre-built DatabricksAdapter.fromServingEndpoint(...) as the agent's model.

Verification

• Live against the dogfood gateway: /ai-gateway/mlflow/v1/chat/completions 200s, model-in-body works (system.ai.claude-opus-5-5 → global.anthropic.claude-opus-5-5), streaming SSE + tool-calling return in the exact choices[0].delta / tool_calls[] shape the adapter parses.
• Tests: adapter + agents-plugin + core-agent suites green (465 passed). New tests cover: gateway path + model in body; serving path leaves model out; and the system.* vs non-system routing split.
• Typecheck: no new errors introduced (pre-existing baseline red is in unrelated genie / sql-warehouse / mock-workspace-client files only).

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 (+3.6 KB) — gzipped download (dist + bin; excludes release-only docs/NOTICE).

dist	raw	gzip
JS (runtime)	1.2 MB (+3.4 KB)	426 KB (+1.0 KB)
Type declarations	446 KB (+2.4 KB)	162 KB (+674 B)
Source maps	2.3 MB (+7.2 KB)	799 KB (+2.0 KB)
Other	11 KB	3.7 KB
Total	4.0 MB (+13 KB)	1.4 MB (+3.7 KB)

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

Entry	Initial (gz)	Lazy (gz)	Total (gz)	node_modules (min)	Own code (min)
.	96 KB	2.5 KB	98 KB	external	314 KB
./beta	93 KB (+170 B)	480 B (+23 B)	94 KB (+193 B)	external	281 KB (+540 B)
./testing	38 KB	30 KB	69 KB	external	201 KB
./tsdown	520 B	0 B	520 B	external	813 B
./type-generator	23 KB	0 B	23 KB	external	65 KB

Chunks:

Entry	Chunk	Load	Size (gz)
.	index.js	initial	92 KB
.	utils.js	initial	4.0 KB
.	remote-tunnel-manager.js	lazy	2.5 KB
./beta	beta.js	initial	77 KB
./beta	stream-manager.js	initial	5.8 KB
./beta	databricks.js	initial	3.3 KB
./beta	wide-event-emitter.js	initial	3.2 KB
./beta	configuration.js	initial	2.3 KB
./beta	service-context.js	initial	1.3 KB
./beta	client.js	initial	434 B
./beta	client-options.js	initial	219 B
./beta	supervisor-api.js	lazy	192 B
./beta	databricks.js	lazy	165 B
./beta	index.js	lazy	123 B
./testing	manifest.js	initial	26 KB
./testing	index.js	initial	10.0 KB
./testing	wide-event-emitter.js	initial	2.9 KB
./testing	index.js	lazy	26 KB
./testing	remote-tunnel-manager.js	lazy	2.5 KB
./testing	utils.js	lazy	1.2 KB
./tsdown	index.js	initial	520 B
./type-generator	index.js	initial	23 KB

@databricks/appkit-ui

npm tarball (packed): 350 KB — gzipped download (dist + bin; excludes release-only docs/NOTICE).

dist	raw	gzip
JS (runtime)	395 KB	132 KB
Type declarations	229 KB	84 KB
Source maps	766 KB	253 KB
CSS	16 KB	3.2 KB
Total	1.4 MB	472 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	20 B	0 B	20 B	0 B	0 B
./react	432 KB	49 KB	481 KB	1.3 MB	177 KB
./react/beta	1.0 KB	0 B	1.0 KB	0 B	1.9 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	20 B
./react	index.js	initial	430 KB
./react	tslib	initial	2.1 KB
./react	apache-arrow	lazy	49 KB
./react/beta	beta.js	initial	1.0 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 36422423929 -R databricks/appkit -n appkit-template-0.78.0-pr.0f7fbf2-feat-agents-ai-gateway-605 -D appkit-pr-605 \
  && unzip -o "appkit-pr-605/appkit-template-0.78.0-pr.0f7fbf2-feat-agents-ai-gateway-605.zip" -d "appkit-pr-605" \
  && databricks apps init --template "appkit-pr-605"

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.