diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 3a3a0f2b4..e116cf2cb 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -117,6 +117,17 @@ jobs:
             echo "::error::The VS Code extension bundle is stale. Run \`just vscode-build\` and commit the result."
             exit 1
           }
+      # Last, so the offline checks fail fast before this spends anything on real models. It calls
+      # the gateway, so it can fail on provider quota or rate limits; that must not turn the required
+      # `test` check red, and the step is still marked failed in the run.
+      - name: Check the requests against a real gateway
+        continue-on-error: true
+        if: ${{ github.event_name != 'merge_group' }}
+        env:
+          UCODE_TEST_WORKSPACE: ${{ secrets.UCODE_TEST_WORKSPACE }}
+          # See the e2e shards: a hosted runner has no cached login, so CI supplies a pre-fetched bearer.
+          DATABRICKS_BEARER: ${{ secrets.DATABRICKS_BEARER }}
+        run: npm run live
 
   e2e-shards:
     name: ${{ matrix.name }} · ${{ matrix.os == 'windows-latest' && 'Windows' || 'Linux' }}
diff --git a/vscode-extension/.gitignore b/vscode-extension/.gitignore
index 13f124c2b..707d70413 100644
--- a/vscode-extension/.gitignore
+++ b/vscode-extension/.gitignore
@@ -1,3 +1,4 @@
 node_modules/
 *.vsix
 .test-out/
+.live-out/
diff --git a/vscode-extension/README.md b/vscode-extension/README.md
index 8abee96ef..a8de86e00 100644
--- a/vscode-extension/README.md
+++ b/vscode-extension/README.md
@@ -21,3 +21,25 @@ Token fetches and failed requests are logged to the **Unity Gateway** output cha
 
 `npm ci`, then `npm run typecheck`, `npm test` and `npm run build`. The build writes the bundle and
 `package.json` to `../src/ucode/vscode/extension/`, which ug packs from; commit them (`just vscode-build`).
+
+`npm run live` sends the requests this extension builds to a real AI Gateway and checks the replies,
+covering what a mocked test can't: the gateway accepting each route's request shape, tool calls,
+images inside tool results, prompt caching, and that a Model Provider Service's models need its
+routing header. It takes the workspace and models from the environment:
+
+```
+UG_LIVE_WORKSPACE=https://<workspace> \
+UG_LIVE_AUTH_COMMAND='ug auth-token --host https://<workspace>' \
+UG_LIVE_MPS= UG_LIVE_OPENAI_MPS= \
+npm run live
+```
+
+CI runs it as the last step of the `VS Code extension` job, from `UCODE_TEST_WORKSPACE` and
+`DATABRICKS_BEARER`; with neither set the suite skips itself, so a fork's pull request stays green.
+`UG_LIVE_WORKSPACE` and `UG_LIVE_TOKEN` override them for a one-off local run.
+
+The models and Model Provider Services default to the ones the integration CUJs use, so CI passes
+only the two secrets. Any of `UG_LIVE_CLAUDE_MODEL`, `UG_LIVE_GPT_MODEL`, `UG_LIVE_MPS`,
+`UG_LIVE_MPS_MODEL`, `UG_LIVE_OPENAI_MPS` and `UG_LIVE_OPENAI_MPS_MODEL` overrides one; setting it to
+the empty string leaves that case out, which is what another workspace needs for the services it
+doesn't have. Each case needs its model or service to be set; the rest are left out. A run costs a few cents.
diff --git a/vscode-extension/package.json b/vscode-extension/package.json
index dc5d88801..85e177dc9 100644
--- a/vscode-extension/package.json
+++ b/vscode-extension/package.json
@@ -33,7 +33,8 @@
   "scripts": {
     "build": "esbuild src/extension.ts --bundle --minify --legal-comments=none --outfile=../src/ucode/vscode/extension/out/extension.js --external:vscode --format=cjs --platform=node --target=node22 && node -e \"const m = require('./package.json'); delete m.scripts; delete m.devDependencies; require('fs').writeFileSync('../src/ucode/vscode/extension/package.json', JSON.stringify(m, null, 2) + '\\n')\"",
     "typecheck": "tsc",
-    "test": "esbuild test/*.test.ts --bundle --platform=node --format=cjs --target=node22 --outdir=.test-out --alias:vscode=./test/vscode.ts --log-level=warning && node --test .test-out/*.test.js"
+    "test": "esbuild test/*.test.ts --bundle --platform=node --format=cjs --target=node22 --outdir=.test-out --alias:vscode=./test/vscode.ts --log-level=warning && node --test .test-out/*.test.js",
+    "live": "esbuild test/live/requests.live.ts --bundle --platform=node --format=cjs --target=node22 --outdir=.live-out --alias:vscode=./test/vscode.ts --log-level=warning && node --test .live-out/requests.live.js"
   },
   "devDependencies": {
     "@types/node": "^22.20.3",
diff --git a/vscode-extension/test/live/requests.live.ts b/vscode-extension/test/live/requests.live.ts
new file mode 100644
index 000000000..0de45076e
--- /dev/null
+++ b/vscode-extension/test/live/requests.live.ts
@@ -0,0 +1,208 @@
+// Sends the requests the extension builds to a real AI Gateway and checks the replies it gets back.
+// Run with `npm run live`; the offline suite globs `test/*.test.ts` and never collects this.
+//
+// Needs UCODE_TEST_WORKSPACE and DATABRICKS_BEARER, as the Python e2e tests do; without them the
+// suite skips rather than fails, so a fork's pull request stays green. The models and Model Provider
+// Services come from the UG_LIVE_* variables below; a case whose variables are unset isn't
+// registered, so the CI job decides what gets covered.
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import { Tokens } from "../../src/auth";
+import {
+  anthropicMessagesReplies,
+  anthropicMessagesRequest,
+  chatCompletionsReplies,
+  chatCompletionsRequest,
+  responsesReplies,
+  responsesRequest,
+  type Reply,
+  type Turn,
+} from "../../src/formats";
+import { PROVIDER_HEADER } from "../../src/extension";
+import { streamEvents } from "../../src/gateway";
+import { ASSISTANT, call, collect, message, rejectsWith, result, SYSTEM, text, USER } from "../support";
+import { LanguageModelDataPart, LanguageModelTextPart, LanguageModelToolCallPart } from "../vscode";
+
+const env = (name: string): string => process.env[name] ?? "";
+
+// The same variables the Python e2e tests take, with UG_LIVE_* overrides for running one by hand.
+const WORKSPACE = (env("UG_LIVE_WORKSPACE") || env("UCODE_TEST_WORKSPACE")).replace(/\/+$/, "");
+const TOKEN_VAR = env("UG_LIVE_TOKEN") ? "UG_LIVE_TOKEN" : "DATABRICKS_BEARER";
+const AUTH_COMMAND = env("UG_LIVE_AUTH_COMMAND") || (env(TOKEN_VAR) && `printf %s "$${TOKEN_VAR}"`);
+
+const MLFLOW = "/ai-gateway/mlflow/v1";
+const OPENAI = "/ai-gateway/openai/v1";
+const MESSAGES = "/ai-gateway/anthropic/v1/messages";
+const ANTHROPIC_VERSION = { "anthropic-version": "2023-06-01" };
+
+const tokens = new Tokens(AUTH_COMMAND);
+
+const send = (path: string, body: object, headers: Record<string, string> = {}) =>
+  streamEvents(WORKSPACE, path, body, tokens, { ...headers }, new AbortController().signal);
+
+const turnOf = (over: Partial<Turn> & { model: string }): Turn => ({
+  messages: [],
+  tools: [],
+  toolRequired: false,
+  ...over,
+});
+
+const textOf = (replies: Reply[]): string =>
+  replies
+    .filter((reply): reply is LanguageModelTextPart => reply instanceof LanguageModelTextPart)
+    .map((part) => part.value)
+    .join("");
+
+const callsOf = (replies: Reply[]): LanguageModelToolCallPart[] =>
+  replies.filter((reply): reply is LanguageModelToolCallPart => reply instanceof LanguageModelToolCallPart);
+
+// A 16x16 red PNG, so a model asked about the image has something real to look at.
+const RED_PNG = "iVBORw0KGgoAAAANSUhEUgAAABAAAAAQCAIAAACQkWg2AAAAFklEQVR4nGO4IyJCEmIY1TCqYfhqAAACcQQQFwFJQgAAAABJRU5ErkJggg==";
+const redImage = () => new LanguageModelDataPart(Buffer.from(RED_PNG, "base64"), "image/png");
+
+const viewImageTool = { name: "view_image", description: "Show the image at a path", inputSchema: { type: "object", properties: { path: { type: "string" } } } };
+
+/** The model must echo this back, so a reply can't pass on stray text alone. */
+const marker = () => `UG${Math.floor(Math.random() * 1e6)}`;
+
+const echoTurn = (model: string, mark: string): Turn =>
+  turnOf({
+    model,
+    messages: [message(SYSTEM, text("Answer in as few words as possible.")), message(USER, text(`Reply with exactly: ${mark}`))],
+  });
+
+const imageTurn = (model: string): Turn =>
+  turnOf({
+    model,
+    tools: [viewImageTool],
+    messages: [
+      message(USER, text("Look at a.png with the tool, then name its main colour in one lowercase word.")),
+      message(ASSISTANT, call("c1", "view_image", { path: "a.png" })),
+      message(USER, result("c1", text("a.png:"), redImage())),
+    ],
+  });
+
+const anthropic = (turn: Turn, maxTokens: number, headers: Record<string, string> = {}) =>
+  collect(anthropicMessagesReplies(send(MESSAGES, anthropicMessagesRequest(turn, maxTokens), { ...ANTHROPIC_VERSION, ...headers })));
+
+const responses = (turn: Turn, base = MLFLOW, headers: Record<string, string> = {}) =>
+  collect(responsesReplies(send(`${base}/responses`, responsesRequest(turn), headers)));
+
+const chat = (turn: Turn, base = MLFLOW, headers: Record<string, string> = {}) =>
+  collect(chatCompletionsReplies(send(`${base}/chat/completions`, chatCompletionsRequest(turn), headers)));
+
+// What to exercise, defaulted to the models and services the integration CUJs use (the same values
+// scripts/run_integration.py defaults its --claude-provider and --codex-provider flags to). Override
+// any of these for another workspace, or set one to "" to leave that case out.
+const setting = (name: string, fallback: string): string => (name in process.env ? env(name) : fallback);
+
+const CLAUDE = setting("UG_LIVE_CLAUDE_MODEL", "system.ai.claude-haiku-4-5");
+const GPT = setting("UG_LIVE_GPT_MODEL", "system.ai.gpt-5-4-nano");
+// No default: every OpenAI-type model on the CI workspace is GPT-5 or newer, which ug routes to
+// Responses, so nothing there exercises Chat Completions.
+const CHAT_GPT = env("UG_LIVE_CHAT_MODEL");
+const MPS = setting("UG_LIVE_MPS", "main.ucode.ci_e2e_anthropic_nonrelay_mps");
+const MPS_MODEL = setting("UG_LIVE_MPS_MODEL", "claude-haiku-4-5-20251001");
+const OPENAI_MPS = setting("UG_LIVE_OPENAI_MPS", "main.ucode.ci_openai_mps");
+const OPENAI_MPS_MODEL = setting("UG_LIVE_OPENAI_MPS_MODEL", "gpt-5-nano");
+
+if (!WORKSPACE || !AUTH_COMMAND) {
+  it.skip("set UCODE_TEST_WORKSPACE and DATABRICKS_BEARER to run the live tests", () => {});
+}
+
+describe("the workspace's own models", () => {
+  if (CLAUDE && WORKSPACE && AUTH_COMMAND) {
+    it("streams a Claude answer on Anthropic Messages, and reads the prompt cache on a repeat", async () => {
+      const mark = marker();
+      // Anthropic caches a prompt only above a minimum length, 2048 tokens for Haiku models, so this
+      // has to clear that with room to spare; the cache point itself sits on the system prompt.
+      const system = text(`Answer in as few words as possible. ${"Ignore this sentence. ".repeat(1500)}`);
+      const turn = turnOf({ model: CLAUDE, messages: [message(SYSTEM, system), message(USER, text(`Reply with exactly: ${mark}`))] });
+
+      assert.match(textOf(await anthropic(turn, 64)), new RegExp(mark));
+      // The second turn is identical, so its prompt should be served from the cache.
+      const events = await collect(send(MESSAGES, anthropicMessagesRequest(turn, 64), ANTHROPIC_VERSION));
+      const start = events.find((event) => (event as { type?: string }).type === "message_start") as
+        | { message?: { usage?: Record<string, number> } }
+        | undefined;
+      assert.ok((start?.message?.usage?.cache_read_input_tokens ?? 0) > 0, JSON.stringify(start?.message?.usage));
+    });
+
+    it("gets a Claude tool call, and an image inside its result reaches the model", async () => {
+      const forced = turnOf({
+        model: CLAUDE,
+        tools: [viewImageTool],
+        toolRequired: true,
+        messages: [message(USER, text("Look at a.png."))],
+      });
+      const [only, ...rest] = callsOf(await anthropic(forced, 256));
+      assert.equal(rest.length, 0);
+      assert.equal(only?.name, "view_image");
+      assert.equal(typeof only?.input, "object");
+
+      assert.match(textOf(await anthropic(imageTurn(CLAUDE), 256)).toLowerCase(), /red/);
+    });
+
+    it("retries with the output limit the gateway allows instead of failing the turn", async () => {
+      // No Claude model serves this many output tokens, so the first request is refused.
+      assert.ok(textOf(await anthropic(echoTurn(CLAUDE, marker()), 999_999)).length > 0);
+    });
+  }
+
+  if (GPT && WORKSPACE && AUTH_COMMAND) {
+    it("streams a GPT answer and a tool call on Responses", async () => {
+      const mark = marker();
+      assert.match(textOf(await responses(echoTurn(GPT, mark))), new RegExp(mark));
+
+      const forced = turnOf({ model: GPT, tools: [viewImageTool], toolRequired: true, messages: [message(USER, text("Look at a.png."))] });
+      const [only] = callsOf(await responses(forced));
+      assert.equal(only?.name, "view_image");
+    });
+
+    it("sends a tool result's image in the request Responses accepts", async () => {
+      assert.match(textOf(await responses(imageTurn(GPT))).toLowerCase(), /red/);
+    });
+  }
+
+  if (CHAT_GPT && WORKSPACE && AUTH_COMMAND) {
+    it("streams an answer on Chat Completions", async () => {
+      const mark = marker();
+      assert.match(textOf(await chat(echoTurn(CHAT_GPT, mark))), new RegExp(mark));
+    });
+  }
+});
+
+describe("a Model Provider Service", () => {
+  if (MPS && MPS_MODEL && WORKSPACE && AUTH_COMMAND) {
+    const header = { [PROVIDER_HEADER]: MPS };
+
+    it("serves its Claude models on Anthropic Messages, with tool calls", async () => {
+      const mark = marker();
+      assert.match(textOf(await anthropic(echoTurn(MPS_MODEL, mark), 256, header)), new RegExp(mark));
+
+      const forced = turnOf({ model: MPS_MODEL, tools: [viewImageTool], toolRequired: true, messages: [message(USER, text("Look at a.png."))] });
+      const [only] = callsOf(await anthropic(forced, 256, header));
+      assert.equal(only?.name, "view_image");
+    });
+
+    it("needs the routing header, and isn't served on the mlflow route", async () => {
+      const turn = echoTurn(MPS_MODEL, marker());
+      // Without this the gateway could be answering from system.ai, and a pass would prove nothing.
+      await rejectsWith(anthropic(turn, 64), "AI Gateway returned");
+      await rejectsWith(chat(turn, MLFLOW, header), "AI Gateway returned");
+    });
+  }
+
+  if (OPENAI_MPS && OPENAI_MPS_MODEL && WORKSPACE && AUTH_COMMAND) {
+    const header = { [PROVIDER_HEADER]: OPENAI_MPS };
+
+    it("serves its OpenAI models on the gateway's OpenAI route", async () => {
+      const mark = marker();
+      assert.match(textOf(await responses(echoTurn(OPENAI_MPS_MODEL, mark), OPENAI, header)), new RegExp(mark));
+    });
+
+    it("isn't served on the mlflow route", async () => {
+      await rejectsWith(responses(echoTurn(OPENAI_MPS_MODEL, marker()), MLFLOW, header), "AI Gateway returned");
+    });
+  }
+});