ContinualCode: Claude Code That Learns from Corrections

Source: https://x.com (tweet) Author: unknown Date: 2025-2026

Summary

Inspired by SDPO, a minimal Claude Code fork that learns from user corrections in real-time using Tinker. When a diff is denied, the model uses the correction as context, takes a gradient step on LoRA weights, and retries. Minimal implementation of online continual learning for coding agents.

Key Claims

  • Real-time learning from rejected diffs
  • Correction → context → LoRA gradient step → retry
  • Built on Tinker (RL training framework)
  • Demonstrates continual learning without catastrophic forgetting (LoRA isolation)
  • Minimal implementation — proves the concept is simple to build

Connection to Other Sources

Implements the concept from Continual Learning Problem (catastrophic forgetting). Uses LoRA to avoid forgetting. Relates to Cursor Real-Time RL — same idea but user-driven.

Concepts