chore: refactor and updates to crucible challenge example

[GangGreenTemperTatum]

Notes

• a small update on the existing crucible challenge example (for prompt injection challenges since theres no tool calling etc in this specific example)
• fixes the platform domain name change also
• removes the weird crucible generator setup and turned it into asyncio process
• the model at the moment does not submit the flag and is done within the code itself

---

Generated Summary

• Enhanced the SYSTEM_PROMPT to provide clearer guidance on interaction with the challenge, emphasizing techniques for prompt injection and flag extraction.
• Implemented CrucibleState dataclass to track attempts, successful techniques, and potential flags, improving state management between attempts.
• Added methods for querying and submitting flags to the challenge API, streamlining interactions with the backend.
• Introduced a new logic in main to vary the temperature parameter during attempts to explore different strategies, potentially increasing success rates.
• Updated command line interface to include a random temperature option and configurable maximum steps, allowing for greater flexibility during execution.
• Revised logging for more detailed progress reporting, including success rates and information regarding encountered flags.
• Removed the old CrucibleGenerator class in favor of simplified function calls, enhancing code clarity and maintainability.
• Improved error handling for API requests with clearer feedback for failure cases.

These changes significantly improve the functionality, usability, and robustness of the crucible example, setting a solid foundation for further development and experimentation.

This summary was generated with ❤️ by rigging

Generated Summary

• Enhanced the SYSTEM_PROMPT with additional context on the challenge and guidance for participants, potentially improving their approach.
• Introduced CrucibleState dataclass for tracking the state during attempts, including metrics for successful and failed techniques, which aids in strategizing.
• Added detailed logging for tracking attempts, including success rates and temperature adjustments for strategy exploration, improving debugging capabilities.
• Refactored the way responses and flags are handled during the challenge, introducing error handling for API interactions to improve robustness.
• Updated cli function to use the new challenge URL format and added options for temperature randomization and maximum steps, enhancing configurability.
• Removed the CrucibleGenerator class in favor of direct asynchronous API calls, simplifying the code and improving maintainability.
• Overall, these changes elevate the challenge experience by providing clearer guidance and a more structured approach to flag extraction.

This summary was generated with ❤️ by rigging