v1.3.1: Launching bluebox

What's Changed

• Guide Agent Uses Last Response Id [EOM] by @dimavrem22 in #98
• Guide agent guides discovery by @dimavrem22 in #100
• no calling discovery before we have cdp captures [EOM] by @dimavrem22 in #101
• msg-parsing fix for discovery request tool by @dimavrem22 in #104
• Small fixes to simplify agent JS generation by @dimavrem22 in #105
• Rename to bluebox by @rayruizhiliao in #106
• Update quickstart vid by @rayruizhiliao in #107
• Update README etc. by @rayruizhiliao in #108

Full Changelog: v1.3.0...v1.3.1

v1.3 - Guide Agent & Improved Validation

Highlights

This release introduces the Guide Agent - an AI-powered assistant that helps users create and refine routines interactively. Also includes significant improvements to routine validation.

New Features

• Guide Agent - A new conversational agent that helps users build and edit routines
  • Scaffolding and context documentation for the agent
  • Ability to edit routines based on user feedback
  • Persistent suggested edits with thread patching support
  • Resource base infrastructure for agent context

Improvements

• Better Routine Validation - Enhanced validation logic for routines, catching more edge cases and providing clearer error messages
• Documentation Updates - Improved README with clearer setup instructions

What's Changed

• better routine validation by @dimavrem22 in #84
• Update README.md by @dimavrem22 in #85
• Guide agent scaffolding by @alex-w-99 in #86
• Guide agent context docs business logic by @dimavrem22 in #89
• more business logic by @dimavrem22 in #91
• improved guide agent context by @dimavrem22 in #93
• agent can edit routine by @dimavrem22 in #94
• resource base | persist suggested edit | thread patch by @dimavrem22 in #95
• Beta v2026.01.22 by @alex-w-99 in #92

Full Changelog: v1.2.3...v1.3

v1.2.3

New Features:

Metadata is recorded for every routine operation during execution:

• All operations: type, duration, errors
• Fetch: request and response
• Download: request and response
• JS Evaluation: console logs and execution errors
• Click: selector, UI element
• Type: selector, UI element, length of typed string

Bugs Fixed:

• Consolidated JS code for Download and Fetch Operations
• Download Operation can now interpolate placeholders from storage and cookies in the request body

v1.2.2

What's Changed

• Agent benchmarks (test the routine discovery pipeline using s3 cdp captures, deterministic tests, and llm routine correctness assessment tests)

Bugs Fixed:

• url of download operation can now be interpolated using cookies, session storage, and local storage
• keys of headers and body can now be interpolated using cookies, session storage, and local storage

Full Changelog: v1.2.1...v1.2.2

v1.2.1

• cleaner prompts (shorter to avoid token waste and better examples)
• we are now using TOON where ever possible to avoid token waste from JSON
• allow 'async' in the js injection

v1.2

🚀 Release Summary

This release delivers major improvements across routine execution, observability, data models, SDK structure, and OpenAI integration, with a strong focus on reliability, performance, and future agentic workflows.

---

✅ Resolved Issues

• Fixed placeholder interpolation tracking for improved error logging and observability (#58)
• Made pylint and pytest required (non-optional) dependencies (#54)
• Adopted OpenAI Responses API structured outputs (#51)
• Prevented Routine Discovery from generating unused parameters (#46)
• Cleaned up incorrect and inconsistent file path references

---

✨ New Features

• Chunking for large outputs to prevent CDP timeouts
• More efficient fetch storage, removing unnecessary CDP round-trips
• JS-level parameter interpolation logging to improve routine observability
• Window property monitoring (v1):
  • Tracks window properties on navigation and periodically
  • Logs value changes and removals with timestamps and URLs
  • Non-blocking to avoid disrupting browser sessions
• New routine operations:
  • JavaScript Evaluation Operation
  • Download Operation
  • UI Operations (foundation for UI-driven routines)
• UI data models:
  • UI elements
  • UI interactions
  • Improved interaction collection aligned with new models

---

🔧 Improvements & Refactors

• Refactored routine execution to operate cleanly at the routine/operation level
• Updated and unified routines, params, operations, and endpoint data models
• Refactored the CDP module
• Added new data utilities
• Improved routine integrity:
  • Automatic escaping of quotes when required
  • Fixed validation bugs (fully covered by unit tests)
• Updated OpenAI Responses API parsing
• Improved prompts and LM response data models
• Expanded network monitoring support (more file types, .txt as default response body)
• Migrated to GPT-5.1 for faster and more reliable execution
• Verified that Responses API provides sufficient context for extended runtimes

---

🧪 Developer Experience

• Updated and expanded unit test coverage
• SDK fully refactored to use new data models
• Added a new quickstart.py that uses the SDK directly, making SDK testing significantly easier

---

🛣️ Looking Ahead

Future PRs will focus on:

• Running and evaluating generated routines before completion (#57)
• Leveraging UI and window property data for reverse-engineering agents
• Adding more UI-driven routine operations
• Further improving placeholder interpolation diagnostics (#58 follow-ups)

v1.1.2

What's Changed

• Window props by @dimavrem22 in #47
• Fix unexpected Chrome session termination by @rayruizhiliao in #49

Full Changelog: v1.1.0...v1.1.2

v1.1.0

Changes in v1.1.0

🚀 Major Changes

PyPI Package Release

• Package is now available on PyPI: pip install web-hacker
• Restructured as a proper Python package with web_hacker/ namespace
• Added CLI entry points: web-hacker-monitor, web-hacker-discover, web-hacker-execute

New High-Level SDK Interface

• Added WebHacker class for simplified usage:

  from web_hacker import WebHacker
  hacker = WebHacker()
  with hacker.monitor_browser(output_dir="./captures"):
      pass
  routine = hacker.discover_routine(task="Search for flights", cdp_captures_dir="./captures")
  result = hacker.execute_routine(routine, parameters={"origin": "NYC", "destination": "LAX"})

✨ New Features

• Quickstart Script: Interactive Python script (quickstart.py) to automate the entire workflow
• Session Storage Interpolation: Support for {{sessionStorage:...}}, {{localStorage:...}}, {{cookie:...}}, and {{meta:...}} placeholders in fetch URLs
• Network Improvements: Added support for HTML responses, data cleaning, and retry loop for failed requests
• Human Interactions: Added support for executing GUI interactions

🔧 Improvements

• Improved task input experience in discovery workflow
• Better error handling and graceful exit
• Enhanced logging and messaging
• Updated documentation and README

🔄 Breaking Changes

Import Paths Changed:

Before

from src.config import Config
python scripts/browser_monitor.py

After

from web_hacker.config import Config
web-hacker-monitor

Note: Core functionality unchanged. Only import paths and package structure changed.

📦 Installation

pip install web-hacker

View on PyPI: https://pypi.org/project/web-hacker/1.1.0/

Initial public release

[1.0.0] - 2025-11-05

Added

• Initial public release of web-hacker
• Core reverse-engineering and web automation modules
• Example routines under /example_routines
• Basic tests and GitHub Actions setup
• Apache 2.0 License and CODEOWNERS configuration