Alexander Whitestone
|
e63cdaf16f
|
feat: self-modifying agent that improves its own prompts (#813)
Docker Build and Publish / build-and-push (pull_request) Has been cancelled
Contributor Attribution Check / check-attribution (pull_request) Has been cancelled
Supply Chain Audit / Scan PR for supply chain risks (pull_request) Has been cancelled
Tests / test (pull_request) Has been cancelled
Tests / e2e (pull_request) Has been cancelled
Resolves #813. Agent analyzes session transcripts for failure
patterns and generates prompt patches to prevent future failures.
agent/self_modify.py (PromptLearner class):
- analyze_session(): detects 5 failure types from transcripts:
retry_loop, timeout, hallucination, context_loss, tool_failure
- generate_patches(): converts patterns to prompt patches with
confidence scoring (frequency-based)
- apply_patches(): appends learned rules to system prompt with
backup and rollback support
- learn_from_session(): full cycle analyze → patch → apply
Failures → patterns → patches → improved prompts → fewer failures.
Safety: patches only ADD rules (append-only), never remove.
Rollback: restores from timestamped backup.
|
2026-04-16 01:23:48 -04:00 |
|