BrowserRuntime Changelog

1.9.6

Patch Changes

8b06340: Add missing symbolic key mappings to KeyboardUtils to fix "Unknown key" errors when using keys like minus, plus, slash, etc. in keyboard actions.

1.9.5

Patch Changes

58a2632: Fix critical message history management issues preventing 400 errors and infinite loops

Three Major Fixes:
1. Extended Thinking Loop Prevention: Fixed infinite loop where agents would repeat the same action (e.g., repeatedly calling goto). When Claude doesn't emit a thinking block, we now add a placeholder by reusing the most recent thinking block from history, instead of removing messages which caused context loss.
2. Extended Thinking Block Validation: When thinkingBudget is enabled, the API requires every assistant message to start with a thinking or redacted_thinking block. The fix adds placeholder thinking blocks to responses when Claude doesn't emit them, preventing 400 errors while preserving agent context.
3. Tool Use/Result Pairing: Fixed "unexpected tool_use_id found in tool_result blocks" error. The API requires each tool_result to have its corresponding tool_use in the IMMEDIATELY PREVIOUS message, not just anywhere in history. Rewrote cleanMessageHistory() to validate pairing on a per-message basis.
Errors Fixed:
- Infinite loops with extended thinking (agent repeating same actions)
- "Expected thinking or redacted_thinking, but found text"
- "unexpected tool_use_id found in tool_result blocks: [id]. Each tool_result block must have a corresponding tool_use block in the previous message"
Testing:
- Extended thinking test passes with multiple tool uses in 27.5s (no loops)
- Properly handles missing thinking blocks by adding placeholders
- Message history stays clean with proper tool_use/tool_result pairing

1.9.4

Patch Changes

2a28db5: Fix critical message history management issues preventing 400 errors

Two Major Fixes:
1. Extended Thinking Block Validation: When thinkingBudget is enabled, the API requires every assistant message to start with a thinking or redacted_thinking block. Added ensureThinkingBlocksForExtendedThinking() to filter out assistant messages without thinking blocks and their corresponding user messages to maintain conversation flow.
2. Tool Use/Result Pairing: Fixed "unexpected tool_use_id found in tool_result blocks" error. The API requires each tool_result to have its corresponding tool_use in the IMMEDIATELY PREVIOUS message, not just anywhere in history. Rewrote cleanMessageHistory() to validate pairing on a per-message basis.
Errors Fixed:
- "Expected thinking or redacted_thinking, but found text"
- "unexpected tool_use_id found in tool_result blocks: [id]. Each tool_result block must have a corresponding tool_use block in the previous message"
Testing: Extended thinking test passes with 15+ tool calls across multiple turns without errors.

1.9.3

Patch Changes

d1e5842: Fix extended thinking 400 error by ensuring all assistant messages start with thinking blocks when thinkingBudget is enabled

1.9.2

Patch Changes

d87ac39: Add workflow_dispatch trigger to release workflow for manual release triggering

1.9.1

Patch Changes

3947312: Fix thinking block handling to prevent 400 errors when using extended thinking

Problem: Using thinkingBudget caused 400 errors from Anthropic's API with the message: "Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block."

Root Cause: The BetaThinkingBlock type incorrectly defined thinking as a config object instead of a string containing the actual thinking content.

Changes:
- Fixed BetaThinkingBlock type: thinking is now correctly typed as string
- Added BetaRedactedThinkingBlock type for handling redacted thinking responses
- Updated responseToParams to properly parse both thinking and redacted_thinking blocks
- Added explicit block ordering when constructing assistant messages to ensure thinking blocks always come first (API requirement)
- Added test examples for extended thinking validation
Usage: Extended thinking now works correctly across multi-turn conversations:
```
const result = await agent.execute(
  "Complex task requiring reasoning",
  undefined,
  {
    thinkingBudget: 4096,
    maxTokens: 16384,
  }
);
```

1.9.0

Minor Changes

ebe5c2f: Expose browser context and createPage to custom tools

Custom tools can now access the browser infrastructure via ToolExecutionContext, allowing them to perform browser operations while leveraging existing session configuration (proxies, anti-detection, cookies, etc.).

New properties on ToolExecutionContext:
- browserContext: Direct access to the Playwright BrowserContext for advanced operations
- createPage(): Creates new pages with anti-detection scripts automatically applied
Features:
- Pages created via createPage() are tracked and automatically cleaned up on errors
- Custom init scripts can be passed to ComputerUseAgent constructor via initScripts option
- Main page remains unaffected by tool operations
Usage:
```
class MyCustomTool implements ComputerUseTool {
  async call(
    params: Record<string, unknown>,
    ctx?: ToolExecutionContext
  ): Promise<ToolResult> {
    if (!ctx?.createPage) {
      return { error: "Browser context not available" };
    }

    const page = await ctx.createPage();
    try {
      await page.goto(url, { waitUntil: "domcontentloaded" });
      // ... perform operations
      return { output: "Success" };
    } finally {
      await page.close();
    }
  }
}
```
This is useful for tools that need to:
- Navigate to URLs (activation links, OAuth callbacks)
- Capture screenshots or PDFs in isolation
- Execute JavaScript in a separate page context
- Perform operations without affecting the main agent page

1.8.1

Patch Changes

3a1092d: Fix missing shift modifier mapping

Fixes "Unknown key: shift" error that occurred when the AI agent sent lowercase "shift" in key combinations. The modifier was not being mapped to the correct Playwright key name "Shift".

Changes:
- Added shift: "Shift" to modifierKeyMap
- Updated tests to verify shift key mapping works correctly
Bug Fix: This resolves production errors like:
```
keyboard.down: Unknown key: "shift"
```
The agent can now correctly handle key combinations with shift:
- "shift+a" → ["Shift", "a"] ✅
- "Shift+Tab" → ["Shift", "Tab"] ✅

1.8.0

Minor Changes

b8a418b: Add support for space-separated keyboard sequences and repetition syntax

Fixes keyboard action handling to support space-separated key sequences like "Down Down Down" which previously failed with Playwright errors. The agent can now handle repeated key presses naturally.

Changes:
- Added parseKeySequence() method to handle space-separated keys
- Support for repetition syntax (e.g., Down*3 = press Down 3 times)
- Automatic detection between key sequences (space-separated) and combinations (plus-separated)
- Enhanced error messages for invalid sequences
Examples:
```
// Space-separated sequences (sequential key presses)
"Down Down Down"; // Press Down arrow 3 times
"Tab Tab Enter"; // Press Tab twice, then Enter

// Repetition syntax
"Down*3"; // Press Down 3 times
"Tab*5 Enter"; // Press Tab 5 times, then Enter

// Traditional key combinations (simultaneous)
"Ctrl+C"; // Hold Ctrl and press C
"Ctrl+Shift+A"; // Hold Ctrl+Shift and press A
```
Technical Details:
- Sequences use keyboard.press() for sequential actions
- Combinations use keyboard.down()/keyboard.up() for simultaneous keys
- Repetition count limited to 1-100 for safety
- Case-insensitive key name handling maintained

1.7.0

Minor Changes

461584d: Add tool execution context support

Adds optional toolExecutionContext parameter to pass runtime state to custom tools during execution. Custom tools can access this context via the second parameter in their call() method. Built-in tools ignore the context, ensuring no breaking changes.

Changes:
- New ToolExecutionContext interface for passing arbitrary context data
- Tools can optionally read context during execution
- Available through agent.execute() options
Usage:
```
await agent.execute("task", undefined, {
  toolExecutionContext: {
    /* your data */
  },
});
```

1.6.2

Patch Changes

02c36a5: Improve pause/resume behavior to prevent unnecessary loop continuation

Fixes an issue where resuming from a paused end_turn state would cause an unnecessary additional LLM API call. When the agent is paused after task completion (end_turn response), resuming now correctly ends execution instead of continuing the loop.

Changes:
- Remove stepIndex++ and continue after resume on end_turn
- Agent now properly ends when task is already complete
- Prevents wasteful extra API calls after resume
- Fixes potential step counter confusion in logs
Behavior:
- Before: Resume on end_turn → continue loop → extra LLM call → finally end
- After: Resume on end_turn → end immediately (task already complete)

1.6.1

Patch Changes

f40804a: Fix pause/resume race condition in agent execution loop

Fixes a race condition where pause signals sent after LLM responds with end_turn were ignored, causing the agent to complete execution despite being paused. This was particularly noticeable with Claude 4-5 model due to faster response times.

Changes:
- Added pause state checking after LLM end_turn response but before loop exit
- Agent now waits for resume signal when paused, even at completion
- Prevents premature execution completion when pause signal arrives after end_turn
- Maintains proper cancellation handling during pause state
Behavior:
- Before: Agent would exit immediately on end_turn, ignoring late pause signals
- After: Agent checks pause state and waits for resume before completing execution
This fix eliminates timing-dependent behavior and ensures pause/resume works reliably across all Claude model versions.

1.6.0

Minor Changes

bcdb2a2: feat: Add preferIPv4 option to resolve Tailscale/VPN IPv6 connectivity issues
- Add preferIPv4 option to RetryConfig interface (renamed from forceIPv4)
- Implement IPv4-only DNS resolution using custom HTTP agents
- Fix ENETUNREACH errors when IPv6 addresses can't be reached on VPN networks
- Clean implementation without global environment variable modifications
- Update documentation with Tailscale/VPN usage guidance
- Add comprehensive example demonstrating retry configuration with IPv4 preference

1.5.0

Minor Changes

5c4206a: feat: Add retry configuration for handling connection errors
- Add configurable retry logic with exponential backoff for API calls
- Create RetryConfig interface with customizable retry parameters
- Implement withRetry wrapper function for automatic retries on connection errors
- Support custom retry configuration in ComputerUseAgent constructor
- Default retry behavior: 3 attempts, 1s initial delay, 2x backoff multiplier
- Retryable errors include: Connection error, ECONNREFUSED, ETIMEDOUT, ECONNRESET, socket errors
- Add example demonstrating retry configuration usage
- Export RetryConfig type from main index

1.4.6

Patch Changes

946f793: Fix: Handle negative scroll amounts without throwing errors
- Convert negative scroll amounts to positive values with appropriate direction
- Override conflicting scroll directions when negative amounts are provided
- Add guidance in system prompt to use positive scroll amounts with direction parameter
- Supports both vertical (up/down) and horizontal (left/right) scrolling with negative amounts

1.4.5

Patch Changes

5d03d6d: Fix CI workflow to properly publish with pnpm

This fixes the "spawn pnpm ENOENT" error by updating the GitHub Actions workflow to use pnpm consistently instead of mixing npm and pnpm commands. The package should now properly publish to npm.

1.4.4

Patch Changes

04f63bf: Trigger npm publish for zod peer dependencies fix

This changeset ensures that version 1.4.3 with the zod peer dependencies fix gets properly published to npm, as the previous CI/CD run failed during the publish step due to git conflicts.

1.4.3

Patch Changes

05a7f04: Fix zod-to-json-schema compatibility issues by moving zod and zod-to-json-schema to peer dependencies

This change resolves the "def.shape is not a function" error that occurs when there are version mismatches between the browseragent's bundled zod/zod-to-json-schema and the consuming application's versions.

Breaking Change Note: Applications using this library now need to install zod and zod-to-json-schema as direct dependencies if they haven't already.

Benefits:
- Eliminates version conflicts between bundled vs application versions
- Ensures schema objects are created and processed by the same library versions
- Reduces bundle size by avoiding duplicate dependencies
- Gives consumers control over the exact versions used
Migration: Add these dependencies to your application's package.json:
```
{
  "dependencies": {
    "zod": "^3.25.0",
    "zod-to-json-schema": "^3.23.0"
  }
}
```

1.4.2

Patch Changes

71b268c: Enhanced playwright goto capability with configurable waitUntil parameter

Added support for users to specify waitUntil conditions through prompts when navigating to URLs. This enhancement allows agents to handle different website loading patterns more effectively.

New features:
- Optional waitUntil parameter in goto method arguments: ["url", "waitUntil"]
- Support for all Playwright waitUntil options:
  - "domcontentloaded" - Fast navigation, waits for DOM ready
  - "load" - Waits for all resources to load
  - "networkidle" - Waits for no network activity (default)
  - "commit" - Minimal wait, returns on network response
- Backward compatible - existing calls work unchanged
- Enhanced error handling and validation
- Updated documentation with examples and use cases
Benefits:
- Faster navigation for enterprise portals using "domcontentloaded"
- Better reliability for sites with continuous background activity
- Reduced timeouts on complex web applications
- User control over navigation timing strategy

1.4.1

Patch Changes

62919e1: Add exports for ComputerUseTool interface, PlaywrightTool, ComputerTool implementations, and related types to enable external tool usage

1.4.0

Minor Changes

e8f03a4: Add logger system
45ddbc6: Fix thinking budget causing API errors by removing default value

Remove the default value of 1024 for thinkingBudget parameter. This fixes 400 errors from Anthropic's API when thinking blocks aren't properly formatted.

Changes:
- thinkingBudget no longer defaults to 1024 in computerUseLoop
- Thinking mode is now disabled by default (when undefined)
- Users must explicitly set thinkingBudget to enable thinking
Fixes:
- Resolves "Expected thinking or redacted_thinking, but found text" API errors
- Prevents unintended thinking mode activation
- Improves reliability for standard agent usage
Migration:
```
// To enable thinking (if previously relying on default)
await agent.execute("task", undefined, { thinkingBudget: 1024 });
```

Patch Changes

6094998: Fix thinking budget default causing API errors

Remove the default value of 1024 for thinkingBudget parameter in computerUseLoop function. Previously, thinking was always enabled even when not explicitly requested, causing 400 errors from Anthropic's API due to improper thinking block formatting.

What changed:
- thinkingBudget parameter no longer defaults to 1024
- Thinking is now disabled by default (when thinkingBudget is undefined)
- Updated documentation to reflect new behavior
Breaking change: Users who relied on the default thinking budget (1024) will now need to explicitly set it:
```
// Before (thinking was enabled by default)
await agent.execute("task");

// After (to enable thinking, explicitly set thinkingBudget)
await agent.execute("task", undefined, { thinkingBudget: 1024 });
```
Fixes:
- Resolves 400 API errors: "Expected thinking or redacted_thinking, but found text"
- Prevents unintended thinking mode activation
- Improves API reliability for standard use cases

Uh oh!

FilesExpand file tree

CHANGELOG.md

Latest commit

History

CHANGELOG.md

File metadata and controls

BrowserRuntime Changelog

1.9.6

Patch Changes

1.9.5

Patch Changes

1.9.4

Patch Changes

1.9.3

Patch Changes

1.9.2

Patch Changes

1.9.1

Patch Changes

1.9.0

Minor Changes

1.8.1

Patch Changes

1.8.0

Minor Changes

1.7.0

Minor Changes

1.6.2

Patch Changes

1.6.1

Patch Changes

1.6.0

Minor Changes

1.5.0

Minor Changes

1.4.6

Patch Changes

1.4.5

Patch Changes

1.4.4

Patch Changes

1.4.3

Patch Changes

1.4.2

Patch Changes

1.4.1

Patch Changes

1.4.0

Minor Changes

Patch Changes

1.2.1

Patch Changes

1.2.0

Minor Changes

Patch Changes