No eval shortcut

[matheper]

This pull request introduces several enhancements to the debug-gym framework, focusing on customizable system prompts, improved agent behavior, and codebase simplifications. The changes include adding support for Jinja templates in system prompts, introducing new utility methods for agents, and simplifying the instructions property in various environments.

Customizable System Prompts

• Added support for Jinja templates to build system prompts, allowing users to customize the format and content of prompts. Custom filters such as to_pretty_json and trim_message were introduced to enhance template functionality. ([[1]](https://github.com/microsoft/debug-gym/pull/152/files#diff-769a128278122d779780df5a3a6a2d1ffa96e793d619dec71d091bf21368824eL75-R188), [[2]](https://github.com/microsoft/debug-gym/pull/152/files#diff-b335630551682c19a781afebcf4d07bf978fb1f8ac04c6bf87428ed5106870f5L140-R222), [[3]](https://github.com/microsoft/debug-gym/pull/152/files#diff-34ea40c6da6f40f445b3eafff912f32342156241356dd99a856f22ace2d24a06R36-R37), [[4]](https://github.com/microsoft/debug-gym/pull/152/files#diff-9ebc73ddba89145d2c3c0d2a7a35c2fc03a4306bc810bc01c3edac0e4dcd96a7R35-R36), [[5]](https://github.com/microsoft/debug-gym/pull/152/files#diff-8b1170dceb4870f28ae9f00425abdbd14207448da847fcfdc58087e399ff5cceR35-R36))
• Updated the documentation in README.md to explain how to use Jinja templates for system prompts, including examples and details about custom filters. ([README.mdL140-R222](https://github.com/microsoft/debug-gym/pull/152/files#diff-b335630551682c19a781afebcf4d07bf978fb1f8ac04c6bf87428ed5106870f5L140-R222))

Agent Enhancements

• Refactored the build_system_prompt method to use either a Jinja template or a default JSON-based format. Added utility methods like shortcut_features and _auto_eval_on_rewrite for better modularity. ([debug_gym/agents/base_agent.pyL75-R188](https://github.com/microsoft/debug-gym/pull/152/files#diff-769a128278122d779780df5a3a6a2d1ffa96e793d619dec71d091bf21368824eL75-R188))
• Introduced the _load_system_prompt_template method to handle loading and validating custom templates. ([debug_gym/agents/base_agent.pyL75-R188](https://github.com/microsoft/debug-gym/pull/152/files#diff-769a128278122d779780df5a3a6a2d1ffa96e793d619dec71d091bf21368824eL75-R188))

Environment Simplifications

• Simplified the instructions property across multiple environments (AiderBenchmarkEnv, MiniNightmareEnv, SWEBenchEnv) to return a plain string instead of a dictionary. ([[1]](https://github.com/microsoft/debug-gym/pull/152/files#diff-0fd30e53ec10b12a89fd527c14c2d8c898af0b24e565ac866877c8c6e322e0f5L15-R16), [[2]](https://github.com/microsoft/debug-gym/pull/152/files#diff-0c9839d137c9c2cedc011109ff42d57f42e819ec862b943a5df9da45f86782f8L25-R26), [[3]](https://github.com/microsoft/debug-gym/pull/152/files#diff-14c661bab81e165fba55704ca88ccbf72e3a1a1f5c528ff9857527106d3f34ddL54-R55))
• Updated the base environment to return an empty string for instructions by default. ([debug_gym/gym/envs/env.pyL275-R276](https://github.com/microsoft/debug-gym/pull/152/files#diff-dd500acaa499f131456c6cd77f9310117f349f19f869fc0fcb4279f71025f108L275-R276))

Configuration Updates

• Added an optional system_prompt_template_file field in configuration files to specify custom Jinja templates. ([[1]](https://github.com/microsoft/debug-gym/pull/152/files#diff-34ea40c6da6f40f445b3eafff912f32342156241356dd99a856f22ace2d24a06R36-R37), [[2]](https://github.com/microsoft/debug-gym/pull/152/files#diff-9ebc73ddba89145d2c3c0d2a7a35c2fc03a4306bc810bc01c3edac0e4dcd96a7R35-R36), [[3]](https://github.com/microsoft/debug-gym/pull/152/files#diff-8b1170dceb4870f28ae9f00425abdbd14207448da847fcfdc58087e399ff5cceR35-R36))
• Changed the default value of auto_eval_on_rewrite to False in multiple configuration files to give users more control over evaluation behavior. ([[1]](https://github.com/microsoft/debug-gym/pull/152/files#diff-34ea40c6da6f40f445b3eafff912f32342156241356dd99a856f22ace2d24a06L11-R11), [[2]](https://github.com/microsoft/debug-gym/pull/152/files#diff-9ebc73ddba89145d2c3c0d2a7a35c2fc03a4306bc810bc01c3edac0e4dcd96a7L11-R11), [[3]](https://github.com/microsoft/debug-gym/pull/152/files#diff-8b1170dceb4870f28ae9f00425abdbd14207448da847fcfdc58087e399ff5cceL11-R11))

Codebase Cleanup

• Removed unused imports and redundant code in various files to improve maintainability. ([[1]](https://github.com/microsoft/debug-gym/pull/152/files#diff-872c89cb5a43ebc1e36269426eb653c84080edcbb0134bcc946b3a788cf09b47L2), [[2]](https://github.com/microsoft/debug-gym/pull/152/files#diff-7f81de762adcd06df31fbac17be81b87332a901ff68f36e232e12d25b754ab92L1-L5))

[MarcCote]

Awesome, I love the new template for human mode thanks.