Agent replies without thinking aloud or tool names; chart-links E2E names its failure (#199)

A run of the real prompt against the local model showed reasoning written as chat text, internal tool names in a capabilities answer, and only a data fetch when asked to build a tool. The working principles now say that everything written is shown, to describe abilities in plain words, and that a request for a tool means building one.

The chart-links spec now waits for the chart or its error and reports browser errors, so the recurring 90-second timeout names its cause. The ZIP upload spec matches the file chip exactly, since the agent reply can quote the name.

apps/web/e2e/chart-links.spec.ts

@@ -1,15 +1,47 @@
-import { expect, test } from "@playwright/test";
+import { expect, type Locator, type Page, test } from "@playwright/test";
 import { captureScreenshot } from "./helpers";
 
 const fixture = "/e2e/fixtures/chart-links.html";
 
+/** Browser errors per page, so a chart that never renders fails with its cause. */
+const browserErrors = new WeakMap<Page, string[]>();
+
+test.beforeEach(({ page }) => {
+  const errors: string[] = [];
+  browserErrors.set(page, errors);
+  page.on("pageerror", (error) => errors.push(error.message));
+  page.on("console", (message) => {
+    if (message.type() === "error") errors.push(message.text());
+  });
+  page.on("response", (response) => {
+    if (response.status() >= 400) errors.push(`${response.status()} ${response.url()}`);
+  });
+});
+
+/** Waits for the plotted dot, failing fast with the chart error and browser errors instead. */
+async function expectChart(page: Page, scope: Locator | Page = page) {
+  const circle = scope.locator("svg circle").first();
+  const failure = scope.getByText("Chart failed to render", { exact: false }).first();
+  try {
+    // Every navigation renders Plot again, which takes seconds on a loaded CI runner.
+    await expect(circle.or(failure).first()).toBeVisible({ timeout: 90_000 });
+    if (await failure.isVisible()) throw new Error(await failure.innerText());
+  } catch (error) {
+    const errors = browserErrors.get(page) ?? [];
+    throw new Error(
+      `${error instanceof Error ? error.message : String(error)}\nBrowser errors:\n${errors.join("\n") || "none"}`,
+    );
+  }
+  return circle;
+}
+
 test.describe
   .serial("chart links", () => {
     test("cannot execute script when clicked or expanded", async ({ page }, testInfo) => {
       test.setTimeout(120_000);
       // ChartCanvas lazy-loads Observable Plot; pay the import cost once before the matrix below.
       await page.goto(`${fixture}?${new URLSearchParams({ href: "#details", inspect: "0" })}`);
-      await expect(page.locator("svg circle").first()).toBeVisible({ timeout: 90_000 });
+      await expectChart(page);
       for (const inspect of ["0", "1"]) {
         for (const href of [
           'javascript:void(document.body.dataset.plotExecuted="yes")',
@@ -19,9 +51,7 @@ test.describe
           "file:///example.txt",
         ]) {
           await page.goto(`${fixture}?${new URLSearchParams({ href, inspect })}`);
-          const circle = page.locator("svg circle").first();
+          const circle = await expectChart(page);
-          // Every navigation renders Plot again, which takes seconds on a loaded CI runner.
-          await expect(circle).toBeVisible({ timeout: 90_000 });
           const assertInert = async () => {
             const links = page.locator("svg a");
             expect(await links.count()).toBeGreaterThan(0);
@@ -43,7 +73,7 @@ test.describe
           await assertInert();
           await page.getByRole("button", { name: "Expand", exact: true }).click();
           const dialog = page.getByRole("dialog");
-          await expect(dialog.locator("svg circle")).toBeVisible({ timeout: 90_000 });
+          await expectChart(page, dialog);
           await dialog.locator("svg circle").click();
           await assertInert();
           await expect(page).toHaveURL(`${fixture}?${new URLSearchParams({ href, inspect })}`);
@@ -62,7 +92,7 @@ test.describe
         route.fulfill({ body: "Chart destination" }),
       );
       await page.goto(`${fixture}?${new URLSearchParams({ href: "#details" })}`);
-      await expect(page.locator("svg circle").first()).toBeVisible({ timeout: 90_000 });
+      await expectChart(page);
       for (const href of [
         "https://example.test/chart?a=1&b=2",
         "mailto:chart@example.test",
@@ -73,7 +103,7 @@ test.describe
         "java%73cript:void(0)",
       ]) {
         await page.goto(`${fixture}?${new URLSearchParams({ href })}`);
-        await expect(page.locator("svg circle").first()).toBeVisible({ timeout: 90_000 });
+        await expectChart(page);
         const link = page.locator("svg a").first();
         await expect(link).toBeAttached();
         expect(

apps/web/e2e/media-pipeline.spec.ts

@@ -104,7 +104,8 @@ test("a repository ZIP above the inline cap streams up as a file", async ({ page
   const response = await upload;
   expect(response.status()).toBe(201);
   expect(response.request().headers()["content-type"]).toBe("application/zip");
-  await expect(page.getByTestId("transcript").getByText("repo.zip")).toBeVisible({
+  // Exact: the agent's reply can quote the file name too.
+  await expect(page.getByTestId("transcript").getByText("repo.zip", { exact: true })).toBeVisible({
     timeout: 30_000,
   });
   await captureScreenshot(page, testInfo, "zip-upload");

packages/adapters/src/executor-instructions.test.ts

@@ -359,6 +359,8 @@ describe("run instructions", () => {
     );
     expect(instructions).toContain("Say something is done only after a tool confirmed it");
     expect(instructions).toContain("Reply in the language the user writes in.");
+    expect(instructions).toContain("do not think aloud");
+    expect(instructions).toContain("never by tool names or ids");
     expect(instructions).toContain("No apps are connected yet.");
     expect(instructions).not.toContain("A schedule started this run");
   });

packages/adapters/src/executor-prompt.ts

@@ -55,10 +55,11 @@ export function formatWorkingPrinciples(toolNames: ReadonlySet<string>): string
       : undefined,
     "- Before saying you cannot do something, check your tools, the connected apps and the skills listed here.",
     canBuildTools
-      ? "- When no tool does what the task needs, build one: write a script in your workspace, test it, and save how to run it as a skill with skill_create so you and the user can reuse it."
+      ? "- When the user asks for a tool, or none of yours does what the task needs, build one: write a script in your workspace, test it, and save how to run it as a skill with skill_create so you and the user can reuse it."
       : undefined,
     "- Ask only when the answer changes the outcome and there is no sensible default; otherwise pick the default and say which.",
     "- Reply in the language the user writes in. Lead with the result and keep it short.",
+    "- Everything you write is shown to the user: do not think aloud, and before a tool call write at most one short sentence. Describe what you can do in plain words, never by tool names or ids.",
   ]
     .filter(Boolean)
     .join("\n");