Add validation error feedback for LLM retry attempts (#102)

[neoneye]

Summary

• Implements proposal docs: task complexity scoring + model routing proposal (post-mortem + rubric) #102: Error-Feedback Retries (Pipeline Intelligence feature fix: Add UTF-8 encoding for markdown file operations #1)
• When Pydantic validation fails, structured error details are extracted and made available to the caller's execute_function via LLMExecutor.validation_feedback
• The caller can inject this feedback into the prompt for the retry attempt, improving success rate

How it works

1. LLM returns invalid JSON that fails Pydantic validation
2. _extract_validation_feedback() walks the exception chain and formats the error
3. The executor stores feedback in validation_feedback and retries the same model
4. The caller's execute_function can check executor.validation_feedback and include it in the prompt

Changes

• worker_plan/worker_plan_internal/llm_util/llm_executor.py:
  • Add _extract_validation_feedback() — walks exception chains for Pydantic errors
  • Add max_validation_retries parameter to LLMExecutor
  • Add validation_feedback property
  • Retry loop for validation errors in run()

Test plan

• Verify Python syntax with ast.parse()
• Existing test_llm_executor.py tests pass (CI)
• Integration test: structured LLM call with intentionally bad schema → verify retry with feedback

🤖 Generated with Claude Code