[Feature] Post-Process Optimization for External Agents like Copilot/Codex

[christso]

Problem

Black-box agents (e.g., GitHub Copilot, legacy Codex) can't be directly optimized, but users want DSPy-like tuning for their outputs (e.g., refine generated code against guidelines).

Proposed Solution

Introduce --mode external-wrapper in bbeval opt:

• Workflow: Mock external agent output → DSPy post-processor (e.g., ChainOfThought verifier + refiner) optimized on testset.
• Input: API mock for Copilot (e.g., via VS Code extension hook); attach guidelines.
• Optimization: Tune refiner prompt (≤5 trials) for tasks like "Fix Copilot's SQL per rules."
• Output: JSON wrapper script (e.g., Python callable) + optimized refine prompt.
• Validation: Use code_execution for output testing.