fix(core): fix browser agent UX issues and improve E2E test reliability (#24312)

This commit is contained in:
Gaurav
2026-04-01 07:00:16 +08:00
committed by GitHub
parent 94f9480a3a
commit 782bb4e4bd
5 changed files with 144 additions and 128 deletions

View File

@@ -10,8 +10,13 @@ import { dirname, join } from 'node:path';
import { fileURLToPath } from 'node:url';
import { execSync } from 'node:child_process';
import { existsSync, writeFileSync, readFileSync, mkdirSync } from 'node:fs';
import { env } from 'node:process';
import stripAnsi from 'strip-ansi';
// Browser agent Chrome DevTools MCP connection is flaky in Docker sandbox.
// See: https://github.com/google-gemini/gemini-cli/issues/24382
const isDockerSandbox = env['GEMINI_SANDBOX'] === 'docker';
const __filename = fileURLToPath(import.meta.url);
const __dirname = dirname(__filename);
@@ -59,122 +64,128 @@ describe.skipIf(!chromeAvailable)('browser-policy', () => {
await rig.cleanup();
});
it('should skip confirmation when "Allow all server tools for this session" is chosen', async () => {
rig.setup('browser-policy-skip-confirmation', {
fakeResponsesPath: join(__dirname, 'browser-policy.responses'),
settings: {
agents: {
overrides: {
browser_agent: {
enabled: true,
it.skipIf(isDockerSandbox)(
'should skip confirmation when "Allow all server tools for this session" is chosen',
async () => {
rig.setup('browser-policy-skip-confirmation', {
fakeResponsesPath: join(__dirname, 'browser-policy.responses'),
settings: {
agents: {
overrides: {
browser_agent: {
enabled: true,
},
},
browser: {
headless: true,
sessionMode: 'isolated',
allowedDomains: ['example.com'],
},
},
browser: {
headless: true,
sessionMode: 'isolated',
allowedDomains: ['example.com'],
},
},
},
});
});
// Manually trust the folder to avoid the dialog and enable option 3
const geminiDir = join(rig.homeDir!, '.gemini');
mkdirSync(geminiDir, { recursive: true });
// Manually trust the folder to avoid the dialog and enable option 3
const geminiDir = join(rig.homeDir!, '.gemini');
mkdirSync(geminiDir, { recursive: true });
// Write to trustedFolders.json
const trustedFoldersPath = join(geminiDir, 'trustedFolders.json');
const trustedFolders = {
[rig.testDir!]: 'TRUST_FOLDER',
};
writeFileSync(trustedFoldersPath, JSON.stringify(trustedFolders, null, 2));
// Write to trustedFolders.json
const trustedFoldersPath = join(geminiDir, 'trustedFolders.json');
const trustedFolders = {
[rig.testDir!]: 'TRUST_FOLDER',
};
writeFileSync(
trustedFoldersPath,
JSON.stringify(trustedFolders, null, 2),
);
// Force confirmation for browser agent.
// NOTE: We don't force confirm browser tools here because "Allow all server tools"
// adds a rule with ALWAYS_ALLOW_PRIORITY (3.9x) which would be overshadowed by
// a rule in the user tier (4.x) like the one from this TOML.
// By removing the explicit mcp rule, the first MCP tool will still prompt
// due to default approvalMode = 'default', and then "Allow all" will correctly
// bypass subsequent tools.
const policyFile = join(rig.testDir!, 'force-confirm.toml');
writeFileSync(
policyFile,
`
// Force confirmation for browser agent.
// NOTE: We don't force confirm browser tools here because "Allow all server tools"
// adds a rule with ALWAYS_ALLOW_PRIORITY (3.9x) which would be overshadowed by
// a rule in the user tier (4.x) like the one from this TOML.
// By removing the explicit mcp rule, the first MCP tool will still prompt
// due to default approvalMode = 'default', and then "Allow all" will correctly
// bypass subsequent tools.
const policyFile = join(rig.testDir!, 'force-confirm.toml');
writeFileSync(
policyFile,
`
[[rule]]
name = "Force confirm browser_agent"
toolName = "browser_agent"
decision = "ask_user"
priority = 200
`,
);
);
// Update settings.json in both project and home directories to point to the policy file
for (const baseDir of [rig.testDir!, rig.homeDir!]) {
const settingsPath = join(baseDir, '.gemini', 'settings.json');
if (existsSync(settingsPath)) {
const settings = JSON.parse(readFileSync(settingsPath, 'utf-8'));
settings.policyPaths = [policyFile];
// Ensure folder trust is enabled
settings.security = settings.security || {};
settings.security.folderTrust = settings.security.folderTrust || {};
settings.security.folderTrust.enabled = true;
writeFileSync(settingsPath, JSON.stringify(settings, null, 2));
// Update settings.json in both project and home directories to point to the policy file
for (const baseDir of [rig.testDir!, rig.homeDir!]) {
const settingsPath = join(baseDir, '.gemini', 'settings.json');
if (existsSync(settingsPath)) {
const settings = JSON.parse(readFileSync(settingsPath, 'utf-8'));
settings.policyPaths = [policyFile];
// Ensure folder trust is enabled
settings.security = settings.security || {};
settings.security.folderTrust = settings.security.folderTrust || {};
settings.security.folderTrust.enabled = true;
writeFileSync(settingsPath, JSON.stringify(settings, null, 2));
}
}
}
const run = await rig.runInteractive({
approvalMode: 'default',
env: {
GEMINI_CLI_INTEGRATION_TEST: 'true',
},
});
const run = await rig.runInteractive({
approvalMode: 'default',
env: {
GEMINI_CLI_INTEGRATION_TEST: 'true',
},
});
await run.sendKeys(
'Open https://example.com and check if there is a heading\r',
);
await run.sendKeys('\r');
await run.sendKeys(
'Open https://example.com and check if there is a heading\r',
);
await run.sendKeys('\r');
// Handle confirmations.
// 1. Initial browser_agent delegation (likely only 3 options, so use option 1: Allow once)
await poll(
() => stripAnsi(run.output).toLowerCase().includes('action required'),
60000,
1000,
);
await run.sendKeys('1\r');
await new Promise((r) => setTimeout(r, 2000));
// Handle confirmations.
// 1. Initial browser_agent delegation (likely only 3 options, so use option 1: Allow once)
await poll(
() => stripAnsi(run.output).toLowerCase().includes('action required'),
60000,
1000,
);
await run.sendKeys('1\r');
await new Promise((r) => setTimeout(r, 2000));
// Handle privacy notice
await poll(
() => stripAnsi(run.output).toLowerCase().includes('privacy notice'),
5000,
100,
);
await run.sendKeys('1\r');
await new Promise((r) => setTimeout(r, 5000));
// Handle privacy notice
await poll(
() => stripAnsi(run.output).toLowerCase().includes('privacy notice'),
5000,
100,
);
await run.sendKeys('1\r');
await new Promise((r) => setTimeout(r, 5000));
// new_page (MCP tool, should have 4 options, use option 3: Allow all server tools)
await poll(
() => {
const stripped = stripAnsi(run.output).toLowerCase();
return (
stripped.includes('new_page') &&
stripped.includes('allow all server tools for this session')
);
},
60000,
1000,
);
// new_page (MCP tool, should have 4 options, use option 3: Allow all server tools)
await poll(
() => {
const stripped = stripAnsi(run.output).toLowerCase();
return (
stripped.includes('new_page') &&
stripped.includes('allow all server tools for this session')
);
},
60000,
1000,
);
// Select "Allow all server tools for this session" (option 3)
await run.sendKeys('3\r');
await new Promise((r) => setTimeout(r, 30000));
// Select "Allow all server tools for this session" (option 3)
await run.sendKeys('3\r');
await new Promise((r) => setTimeout(r, 30000));
const output = stripAnsi(run.output).toLowerCase();
const output = stripAnsi(run.output).toLowerCase();
expect(output).toContain('browser_agent');
expect(output).toContain('completed successfully');
});
expect(output).toContain('browser_agent');
expect(output).toContain('completed successfully');
},
);
it('should show the visible warning when browser agent starts in existing session mode', async () => {
rig.setup('browser-session-warning', {

View File

@@ -121,6 +121,7 @@ describe('file-system', () => {
const result = await rig.run({
args: `write "hello" to "${fileName}" and then stop. Do not perform any other actions.`,
timeout: 600000, // 10 min — real LLM can be slow in Docker sandbox
});
const foundToolCall = await rig.waitForToolCall('write_file');

View File

@@ -166,7 +166,7 @@ describe('browserAgentFactory', () => {
expect(browserManager).toBeDefined();
});
it('should call printOutput when provided', async () => {
it('should not call printOutput for internal setup messages', async () => {
const printOutput = vi.fn();
await createBrowserAgentDefinition(
@@ -175,7 +175,7 @@ describe('browserAgentFactory', () => {
printOutput,
);
expect(printOutput).toHaveBeenCalled();
expect(printOutput).not.toHaveBeenCalled();
});
it('should create definition with correct structure', async () => {

View File

@@ -53,7 +53,7 @@ import {
export async function createBrowserAgentDefinition(
config: Config,
messageBus: MessageBus,
printOutput?: (msg: string) => void,
_printOutput?: (msg: string) => void,
): Promise<{
definition: LocalAgentDefinition<typeof BrowserTaskResultSchema>;
browserManager: BrowserManager;
@@ -66,23 +66,17 @@ export async function createBrowserAgentDefinition(
const browserManager = BrowserManager.getInstance(config);
await browserManager.ensureConnection();
if (printOutput) {
printOutput('Browser connected with isolated MCP client.');
}
debugLogger.log('Browser connected with isolated MCP client.');
// Determine if input blocker should be active (non-headless + enabled)
const shouldDisableInput = config.shouldDisableBrowserUserInput();
// Inject automation overlay and input blocker if not in headless mode
const browserConfig = config.getBrowserAgentConfig();
if (!browserConfig?.customConfig?.headless) {
if (printOutput) {
printOutput('Injecting automation overlay...');
}
debugLogger.log('Injecting automation overlay...');
await injectAutomationOverlay(browserManager);
if (shouldDisableInput) {
if (printOutput) {
printOutput('Injecting input blocker...');
}
debugLogger.log('Injecting input blocker...');
await injectInputBlocker(browserManager);
}
}

View File

@@ -18,7 +18,6 @@ import { randomUUID } from 'node:crypto';
import type { Config } from '../../config/config.js';
import { type AgentLoopContext } from '../../config/agent-loop-context.js';
import { LocalAgentExecutor } from '../local-executor.js';
import { safeJsonToMarkdown } from '../../utils/markdownUtils.js';
import {
BaseToolInvocation,
type ToolResult,
@@ -30,6 +29,7 @@ import {
type SubagentActivityEvent,
type SubagentProgress,
type SubagentActivityItem,
AgentTerminateMode,
isToolActivityError,
} from '../types.js';
import type { MessageBus } from '../../confirmation-bus/message-bus.js';
@@ -56,6 +56,8 @@ export class BrowserAgentInvocation extends BaseToolInvocation<
AgentInputs,
ToolResult
> {
private readonly agentName: string;
constructor(
private readonly context: AgentLoopContext,
params: AgentInputs,
@@ -63,13 +65,15 @@ export class BrowserAgentInvocation extends BaseToolInvocation<
_toolName?: string,
_toolDisplayName?: string,
) {
const resolvedName = _toolName ?? 'browser_agent';
// Note: BrowserAgentDefinition is a factory function, so we use hardcoded names
super(
params,
messageBus,
_toolName ?? 'browser_agent',
resolvedName,
_toolDisplayName ?? 'Browser Agent',
);
this.agentName = resolvedName;
}
private get config(): Config {
@@ -112,7 +116,7 @@ export class BrowserAgentInvocation extends BaseToolInvocation<
// Send initial state
const initialProgress: SubagentProgress = {
isSubagentProgress: true,
agentName: this['_toolName'] ?? 'browser_agent',
agentName: this.agentName,
recentActivity: [],
state: 'running',
};
@@ -135,7 +139,7 @@ export class BrowserAgentInvocation extends BaseToolInvocation<
}
updateOutput({
isSubagentProgress: true,
agentName: this['_toolName'] ?? 'browser_agent',
agentName: this.agentName,
recentActivity: [...recentActivity],
state: 'running',
} as SubagentProgress);
@@ -280,7 +284,7 @@ export class BrowserAgentInvocation extends BaseToolInvocation<
const progress: SubagentProgress = {
isSubagentProgress: true,
agentName: this['_toolName'] ?? 'browser_agent',
agentName: this.agentName,
recentActivity: [...recentActivity],
state: 'running',
};
@@ -297,34 +301,40 @@ export class BrowserAgentInvocation extends BaseToolInvocation<
const output = await executor.run(this.params, signal);
const displayResult = safeJsonToMarkdown(output.result);
const resultContent = `Browser agent finished.
Termination Reason: ${output.terminate_reason}
Result:
${output.result}`;
const displayContent = `
Browser Agent Finished
// Map terminate_reason to the correct SubagentProgress state.
// GOAL = agent completed its task normally.
// ABORTED = user cancelled.
// Others (ERROR, MAX_TURNS, ERROR_NO_COMPLETE_TASK_CALL) = error.
let progressState: SubagentProgress['state'];
if (output.terminate_reason === AgentTerminateMode.ABORTED) {
progressState = 'cancelled';
} else if (output.terminate_reason === AgentTerminateMode.GOAL) {
progressState = 'completed';
} else {
progressState = 'error';
}
Termination Reason: ${output.terminate_reason}
Result:
${displayResult}
`;
const progress: SubagentProgress = {
isSubagentProgress: true,
agentName: this.agentName,
recentActivity: [...recentActivity],
state: progressState,
result: output.result,
terminateReason: output.terminate_reason,
};
if (updateOutput) {
updateOutput({
isSubagentProgress: true,
agentName: this['_toolName'] ?? 'browser_agent',
recentActivity: [...recentActivity],
state: 'completed',
} as SubagentProgress);
updateOutput(progress);
}
return {
llmContent: [{ text: resultContent }],
returnDisplay: displayContent,
returnDisplay: progress,
};
} catch (error) {
const rawErrorMessage =
@@ -343,7 +353,7 @@ ${displayResult}
const progress: SubagentProgress = {
isSubagentProgress: true,
agentName: this['_toolName'] ?? 'browser_agent',
agentName: this.agentName,
recentActivity: [...recentActivity],
state: isAbort ? 'cancelled' : 'error',
};