feat(browser-agent): add type_text composite tool and improve prompt

- Add custom type_text tool that types a full string by internally
  calling press_key for each character, turning N model round-trips
  into 1. Dramatically speeds up text input in complex web apps.

- Move tool-specific usage rules from system prompt to individual
  tool descriptions via augmentToolDescription() for better
  organization and token efficiency.

- Add terminal failure handling instructions to system prompt
  (Chrome connection errors, browser crashes, repeated errors)
  with specific remediation steps.

- Add complex web app guidance (spreadsheets, rich editors) to
  system prompt, recommending type_text + keyboard navigation.

- Fix augmentToolDescription key ordering so more-specific keys
  (fill_form, click_at) match before shorter keys (fill, click).

- Remove non-existent tool references (scroll, type_text as MCP tool)
  and add click_at hint for vision tool.
This commit is contained in:
Gaurav Ghosh
2026-02-17 05:15:28 -08:00
parent 1c8a37379b
commit 2bc2945d14
2 changed files with 232 additions and 30 deletions
@@ -55,7 +55,6 @@ PARALLEL TOOL CALLS - CRITICAL:
- Do NOT make parallel calls for actions that change page state (click, fill, press_key, etc.)
- Each action changes the DOM and invalidates UIDs from the current snapshot
- Make state-changing actions ONE AT A TIME, then observe the results
- For typing text, prefer type_text tool instead of multiple press_key calls
OVERLAY/POPUP HANDLING:
Before interacting with page content, scan the accessibility tree for blocking overlays:
@@ -72,6 +71,24 @@ When you need to identify elements by visual attributes not in the AX tree (e.g.
3. Use the returned coordinates with click_at(x, y) or other tools yourself
4. If the analysis is insufficient, call it again with a more specific instruction
COMPLEX WEB APPS (spreadsheets, rich editors, canvas apps):
Many web apps (Google Sheets/Docs, Notion, Figma, etc.) use custom rendering rather than standard HTML inputs.
- fill does NOT work on these apps. Instead, click the target element, then use type_text to enter the value.
- Navigate cells/fields using keyboard shortcuts (Tab, Enter, ArrowDown) — more reliable than clicking cell UIDs.
- For spreadsheets: click a cell → type_text("value") → press_key("Enter") to confirm and move to the next cell.
- Use the Name Box (cell reference input, usually showing "A1") to jump to specific cells.
TERMINAL FAILURES — STOP IMMEDIATELY:
Some errors are unrecoverable and retrying will never help. When you see ANY of these, call complete_task immediately with success=false and include the specific remediation steps in your summary:
- "Could not connect to Chrome" — Include ALL of these instructions in your summary:
1. Open Chrome (version 144+)
2. Go to chrome://inspect/#remote-debugging and enable remote debugging
3. Or change sessionMode to "persistent" in settings.json to let the agent launch its own browser
- "Browser closed" or "Target closed" or "Session closed" — The browser process has terminated. Tell the user to restart and try again.
- "net::ERR_" network errors on the SAME URL after 2 retries — the site is unreachable. Report the URL and error.
- Any error that appears IDENTICALLY 3+ times in a row — it will not resolve by retrying.
Do NOT keep retrying terminal errors. Report them with actionable remediation steps and exit immediately.
CRITICAL: When you have fully completed the user's task, you MUST call the complete_task tool with a summary of what you accomplished. Do NOT just return text - you must explicitly call complete_task to exit the loop.`;
/**
@@ -116,6 +116,13 @@ class McpToolInvocation extends BaseToolInvocation<
};
} catch (error) {
const errorMsg = error instanceof Error ? error.message : String(error);
// Chrome connection errors are fatal — re-throw to terminate the agent
// immediately instead of returning a result the LLM would retry.
if (errorMsg.includes('Could not connect to Chrome')) {
throw error;
}
debugLogger.error(`MCP tool ${this.toolName} failed: ${errorMsg}`);
return {
llmContent: `Error: ${errorMsg}`,
@@ -126,6 +133,112 @@ class McpToolInvocation extends BaseToolInvocation<
}
}
/**
* Composite tool invocation that types a full string by calling press_key
* for each character internally, avoiding N model round-trips.
*/
class TypeTextInvocation extends BaseToolInvocation<
Record<string, unknown>,
ToolResult
> {
constructor(
private readonly browserManager: BrowserManager,
private readonly text: string,
messageBus: MessageBus,
) {
super({ text }, messageBus, 'type_text', 'type_text');
}
getDescription(): string {
return `type_text: "${this.text.substring(0, 50)}${this.text.length > 50 ? '...' : ''}"`;
}
protected override async getConfirmationDetails(
_abortSignal: AbortSignal,
): Promise<ToolCallConfirmationDetails | false> {
if (!this.messageBus) {
return false;
}
return {
type: 'mcp',
title: `Confirm Tool: type_text`,
serverName: 'browser-agent',
toolName: 'type_text',
toolDisplayName: 'type_text',
onConfirm: async (outcome: ToolConfirmationOutcome) => {
await this.publishPolicyUpdate(outcome);
},
};
}
protected override getPolicyUpdateOptions(
_outcome: ToolConfirmationOutcome,
): PolicyUpdateOptions | undefined {
return {
mcpName: 'browser-agent',
};
}
override async execute(): Promise<ToolResult> {
try {
const chars = [...this.text]; // Handle Unicode correctly
let successCount = 0;
let lastError: string | undefined;
for (const char of chars) {
// Map special characters to key names
const key = char === ' ' ? 'Space' : char;
const result: McpToolCallResult = await this.browserManager.callTool(
'press_key',
{ key },
);
if (result.isError) {
const errorText = result.content
?.filter(
(c: { type: string; text?: string }) =>
c.type === 'text' && c.text,
)
.map((c: { type: string; text?: string }) => c.text)
.join('\n');
lastError = errorText || 'Unknown error';
// Continue typing remaining characters on soft errors
debugLogger.warn(
`type_text: press_key("${key}") failed: ${lastError}`,
);
} else {
successCount++;
}
}
const summary =
successCount === chars.length
? `Successfully typed ${chars.length} characters: "${this.text}"`
: `Typed ${successCount}/${chars.length} characters. Last error: ${lastError}`;
return {
llmContent: summary,
returnDisplay: summary,
};
} catch (error) {
const errorMsg = error instanceof Error ? error.message : String(error);
// Chrome connection errors are fatal
if (errorMsg.includes('Could not connect to Chrome')) {
throw error;
}
debugLogger.error(`type_text failed: ${errorMsg}`);
return {
llmContent: `Error: ${errorMsg}`,
returnDisplay: `Error: ${errorMsg}`,
error: { message: errorMsg },
};
}
}
}
/**
* DeclarativeTool wrapper for an MCP tool.
*/
@@ -165,7 +278,56 @@ class McpDeclarativeTool extends DeclarativeTool<
}
/**
* Creates DeclarativeTool instances from dynamically discovered MCP tools.
* DeclarativeTool for the custom type_text composite tool.
*/
class TypeTextDeclarativeTool extends DeclarativeTool<
Record<string, unknown>,
ToolResult
> {
constructor(
private readonly browserManager: BrowserManager,
messageBus: MessageBus,
) {
super(
'type_text',
'type_text',
'Types a full text string into the currently focused element by pressing each key in sequence. ' +
'Much faster than calling press_key for each character individually. ' +
'Use this to enter text into form fields, spreadsheet cells, or any focused input. ' +
'The element must already be focused (e.g., after a click). ' +
'Does NOT press Enter at the end — call press_key("Enter") separately if needed.',
Kind.Other,
{
type: 'object',
properties: {
text: {
type: 'string',
description:
'The text to type. Each character will be pressed in sequence.',
},
},
required: ['text'],
},
messageBus,
/* isOutputMarkdown */ true,
/* canUpdateOutput */ false,
);
}
build(
params: Record<string, unknown>,
): ToolInvocation<Record<string, unknown>, ToolResult> {
return new TypeTextInvocation(
this.browserManager,
String(params['text'] ?? ''),
this.messageBus,
);
}
}
/**
* Creates DeclarativeTool instances from dynamically discovered MCP tools,
* plus custom composite tools (like type_text).
*
* These tools are registered in the browser agent's isolated ToolRegistry,
* NOT in the main agent's registry.
@@ -179,7 +341,7 @@ class McpDeclarativeTool extends DeclarativeTool<
export async function createMcpDeclarativeTools(
browserManager: BrowserManager,
messageBus: MessageBus,
): Promise<McpDeclarativeTool[]> {
): Promise<Array<McpDeclarativeTool | TypeTextDeclarativeTool>> {
// Get dynamically discovered tools from the MCP server
const mcpTools = await browserManager.getDiscoveredTools();
@@ -187,21 +349,31 @@ export async function createMcpDeclarativeTools(
`Creating ${mcpTools.length} declarative tools for browser agent`,
);
return mcpTools.map((mcpTool) => {
const schema = convertMcpToolToFunctionDeclaration(mcpTool);
// Augment description with uid-context hints
const augmentedDescription = augmentToolDescription(
mcpTool.name,
mcpTool.description ?? '',
);
return new McpDeclarativeTool(
browserManager,
mcpTool.name,
augmentedDescription,
schema.parametersJsonSchema,
messageBus,
);
});
const tools: Array<McpDeclarativeTool | TypeTextDeclarativeTool> =
mcpTools.map((mcpTool) => {
const schema = convertMcpToolToFunctionDeclaration(mcpTool);
// Augment description with uid-context hints
const augmentedDescription = augmentToolDescription(
mcpTool.name,
mcpTool.description ?? '',
);
return new McpDeclarativeTool(
browserManager,
mcpTool.name,
augmentedDescription,
schema.parametersJsonSchema,
messageBus,
);
});
// Add custom composite tools
tools.push(new TypeTextDeclarativeTool(browserManager, messageBus));
debugLogger.log(
`Total tools registered: ${tools.length} (${mcpTools.length} MCP + 1 custom)`,
);
return tools;
}
/**
@@ -223,31 +395,44 @@ function convertMcpToolToFunctionDeclaration(
}
/**
* Augments MCP tool descriptions with uid-context hints.
* Adds semantic guidance for tools that work with accessibility tree elements.
* Augments MCP tool descriptions with usage guidance.
* Adds semantic hints and usage rules directly in tool descriptions
* so the model makes correct tool choices without system prompt overhead.
*
* Actual chrome-devtools-mcp tools:
* Input: click, drag, fill, fill_form, handle_dialog, hover, press_key, upload_file
* Navigation: close_page, list_pages, navigate_page, new_page, select_page, wait_for
* Emulation: emulate, resize_page
* Performance: performance_analyze_insight, performance_start_trace, performance_stop_trace
* Network: get_network_request, list_network_requests
* Debugging: evaluate_script, get_console_message, list_console_messages, take_screenshot, take_snapshot
* Vision (--experimental-vision): click_at, analyze_screenshot
*/
function augmentToolDescription(toolName: string, description: string): string {
const uidHints: Record<string, string> = {
// More-specific keys MUST come before shorter keys to prevent
// partial matching from short-circuiting (e.g., fill_form before fill).
const hints: Record<string, string> = {
fill_form:
' Fills multiple standard HTML form fields at once. Same limitations as fill — does not work on canvas/custom widgets.',
fill: ' Fills standard HTML form fields (<input>, <textarea>, <select>) by uid. Does NOT work on custom/canvas-based widgets (e.g., Google Sheets cells, Notion blocks). If fill times out or fails, click the element first then use press_key with individual characters instead.',
click_at:
' Clicks at exact pixel coordinates (x, y). Use with analyze_screenshot to find coordinates of visual elements.',
click:
' Use the element uid from the accessibility tree snapshot (e.g., uid="87_4" for a button).',
fill: ' Use the element uid from the accessibility tree snapshot for input/select elements.',
' Use the element uid from the accessibility tree snapshot (e.g., uid="87_4"). UIDs are invalidated after this action — call take_snapshot before using another uid.',
hover:
' Use the element uid from the accessibility tree snapshot to hover over elements.',
type: ' Type text into the currently focused element.',
scroll:
' Scroll the page in the specified direction. Use after take_snapshot to see more content.',
take_snapshot:
' Returns the accessibility tree with uid values for each element. Call this first to see available elements.',
' Returns the accessibility tree with uid values for each element. Call this FIRST to see available elements, and AFTER every state-changing action (click, fill, press_key) before using any uid.',
navigate_page:
' Navigate to the specified URL. Call take_snapshot after to see the new page.',
new_page:
' Opens a new page/tab with the specified URL. Call take_snapshot after to see the new page.',
press_key:
' Press a keyboard key. Use for Enter, Tab, Escape, arrow keys, etc.',
' Press a SINGLE keyboard key (e.g., "Enter", "Tab", "Escape", "ArrowDown", "a", "8"). ONLY accepts one key name — do NOT pass multi-character strings like "Hello" or "A1\\nEnter". To type text, call press_key once per character.',
};
// Check for partial matches (e.g., "click" matches "click_element")
for (const [key, hint] of Object.entries(uidHints)) {
// Check for partial matches — order matters! More-specific keys first.
for (const [key, hint] of Object.entries(hints)) {
if (toolName.toLowerCase().includes(key)) {
return description + hint;
}