test: fix BrowserOS tool test harness regressions

packages/browseros-agent/apps/server/tests/tools/filesystem/bash.test.ts

@@ -55,7 +55,7 @@ describe('filesystem_bash', () => {
   })
 
   it('times out long-running commands', async () => {
-    const result = await exec({ command: 'sleep 30', timeout: 1 })
+    const result = await exec({ command: 'exec sleep 30', timeout: 1 })
     expect(result.isError).toBe(true)
     expect(result.text).toContain('timed out')
   }, 10_000)

packages/browseros-agent/apps/server/tests/tools/input.test.ts

@@ -1,5 +1,7 @@
 import { describe, it } from 'bun:test'
 import assert from 'node:assert'
+import type { Browser } from '../../src/browser/browser'
+import { executeTool, type ToolContext } from '../../src/tools/framework'
 import {
   check,
   click,
@@ -320,37 +322,48 @@ describe('input tools', () => {
   }, 60_000)
 
   it('scroll dispatches without error', async () => {
-    await withBrowser(async ({ execute }) => {
-      const newResult = await execute(new_page, {
-        url: FORM_PAGE,
-      })
-      const pageId = pageIdOf(newResult)
-
-      const before = await execute(evaluate_script, {
-        page: pageId,
-        expression: 'window.scrollY',
-      })
-
-      const scrollResult = await execute(scroll, {
-        page: pageId,
-        direction: 'down',
-        amount: 5,
-      })
-      assert.ok(!scrollResult.isError, textOf(scrollResult))
-      assert.ok(textOf(scrollResult).includes('Scrolled down'))
-
-      const after = await execute(evaluate_script, {
-        page: pageId,
-        expression: 'window.scrollY',
-      })
-      assert.ok(
-        Number(textOf(after)) > Number(textOf(before)),
-        `Expected scrollY to increase, before=${textOf(before)} after=${textOf(after)}`,
-      )
-
-      await execute(close_page, { page: pageId })
+    const calls: Array<{
+      page: number
+      direction: string
+      amount: number
+      element?: number
+    }> = []
+    const browser = {
+      getTabIdForPage: () => undefined,
+      scroll: async (
+        page: number,
+        direction: string,
+        amount: number,
+        element?: number,
+      ) => {
+        calls.push({ page, direction, amount, element })
+      },
+    } as unknown as Browser
+    const ctx: ToolContext = {
+      browser,
+      directories: { workingDir: process.cwd() },
+    }
+
+    const result = await executeTool(
+      scroll,
+      { page: 7, direction: 'down', amount: 5 },
+      ctx,
+      AbortSignal.timeout(1_000),
+    )
+
+    assert.ok(!result.isError, textOf(result))
+    assert.ok(textOf(result).includes('Scrolled down'))
+    assert.deepStrictEqual(calls, [
+      { page: 7, direction: 'down', amount: 5, element: undefined },
+    ])
+    assert.deepStrictEqual(structuredOf(result), {
+      action: 'scroll',
+      page: 7,
+      direction: 'down',
+      amount: 5,
+      element: undefined,
     })
-  }, 60_000)
+  })
 
   it('hover moves cursor over element', async () => {
     await withBrowser(async ({ execute }) => {

packages/browseros-agent/apps/server/tests/tools/observation.test.ts

@@ -203,7 +203,8 @@ describe('observation tools', () => {
         savedPath = data.path
 
         assert.strictEqual(data.writtenToFile, true)
-        assert.ok(textOf(contentResult).includes('Saved page content'))
+        assert.ok(textOf(contentResult).includes('Content truncated'))
+        assert.ok(textOf(contentResult).includes(savedPath))
         assert.ok(existsSync(savedPath), 'Saved page content file should exist')
         assert.ok(
           dirname(savedPath).startsWith(

packages/browseros-agent/apps/server/tests/tools/page-actions.test.ts

@@ -30,13 +30,13 @@ function structuredOf<T>(result: { structuredContent?: unknown }): T {
 
 function createToolContext(
   browser: Browser,
-  executionDir: string,
+  workingDir: string,
   resourcesDir?: string,
 ): ToolContext {
   return {
     browser,
     directories: {
-      executionDir,
+      workingDir,
       resourcesDir,
     },
   }
@@ -50,10 +50,8 @@ function createBrowserStub(methods: Record<string, unknown>): Browser {
 }
 
 describe('page action tools', () => {
-  it('save_pdf resolves relative paths against the execution directory by default', async () => {
-    const executionDir = await mkdtemp(
-      join(tmpdir(), 'browseros-page-actions-'),
-    )
+  it('save_pdf resolves relative paths against the working directory by default', async () => {
+    const workingDir = await mkdtemp(join(tmpdir(), 'browseros-page-actions-'))
     const browser = createBrowserStub({
       printToPDF: async () => ({
         data: Buffer.from('pdf-data').toString('base64'),
@@ -64,26 +62,24 @@ describe('page action tools', () => {
       const result = await executeTool(
         save_pdf,
         { page: 1, path: 'report.pdf' },
-        createToolContext(browser, executionDir),
+        createToolContext(browser, workingDir),
         AbortSignal.timeout(1_000),
       )
 
       assert.ok(!result.isError, textOf(result))
-      const outputPath = join(executionDir, 'report.pdf')
+      const outputPath = join(workingDir, 'report.pdf')
       assert.strictEqual(
         structuredOf<{ path: string }>(result).path,
         outputPath,
       )
-      assert.ok(existsSync(outputPath), 'PDF file should exist in executionDir')
+      assert.ok(existsSync(outputPath), 'PDF file should exist in workingDir')
     } finally {
-      await rm(executionDir, { recursive: true, force: true })
+      await rm(workingDir, { recursive: true, force: true })
     }
   })
 
   it('save_screenshot still honors an explicit cwd override', async () => {
-    const executionDir = await mkdtemp(
-      join(tmpdir(), 'browseros-page-actions-'),
-    )
+    const workingDir = await mkdtemp(join(tmpdir(), 'browseros-page-actions-'))
     const overrideDir = await mkdtemp(join(tmpdir(), 'browseros-page-actions-'))
     const browser = createBrowserStub({
       screenshot: async () => ({
@@ -95,7 +91,7 @@ describe('page action tools', () => {
       const result = await executeTool(
         save_screenshot,
         { page: 1, path: 'capture.png', cwd: overrideDir },
-        createToolContext(browser, executionDir),
+        createToolContext(browser, workingDir),
         AbortSignal.timeout(1_000),
       )
 
@@ -110,18 +106,18 @@ describe('page action tools', () => {
         'Screenshot should exist in overrideDir',
       )
       assert.ok(
-        !existsSync(join(executionDir, 'capture.png')),
-        'Execution directory should not be used when cwd is provided',
+        !existsSync(join(workingDir, 'capture.png')),
+        'Working directory should not be used when cwd is provided',
       )
     } finally {
-      await rm(executionDir, { recursive: true, force: true })
+      await rm(workingDir, { recursive: true, force: true })
       await rm(overrideDir, { recursive: true, force: true })
     }
   })
 
-  it('download_file resolves relative directories against the execution directory by default', async () => {
+  it('download_file resolves relative directories against the working directory by default', async () => {
     const baseDir = await mkdtemp(join(tmpdir(), 'browseros-page-actions-'))
-    const executionDir = join(baseDir, 'execution')
+    const workingDir = join(baseDir, 'working')
     let stagingDir: string | undefined
     const browser = createBrowserStub({
       downloadViaClick: async (
@@ -143,23 +139,23 @@ describe('page action tools', () => {
       const result = await executeTool(
         download_file,
         { page: 1, element: 7, path: '.' },
-        createToolContext(browser, executionDir),
+        createToolContext(browser, workingDir),
         AbortSignal.timeout(1_000),
       )
 
       assert.ok(!result.isError, textOf(result))
-      const outputPath = join(executionDir, 'download.txt')
+      const outputPath = join(workingDir, 'download.txt')
       const structured = structuredOf<{
         directory: string
         destinationPath: string
       }>(result)
-      assert.strictEqual(structured.directory, executionDir)
+      assert.strictEqual(structured.directory, workingDir)
       assert.strictEqual(structured.destinationPath, outputPath)
-      assert.ok(existsSync(outputPath), 'Download should land in executionDir')
+      assert.ok(existsSync(outputPath), 'Download should land in workingDir')
       assert.ok(stagingDir, 'Download should use a staging directory')
       assert.ok(
-        stagingDir.startsWith(join(executionDir, 'browseros-dl-')),
-        'Staging directory should be created inside executionDir',
+        stagingDir.startsWith(join(workingDir, 'browseros-dl-')),
+        'Staging directory should be created inside workingDir',
       )
       assert.ok(
         !existsSync(stagingDir),