continualcode

Status: Unstable — results will be posted soon

A minimal coding agent that trains itself from your corrections, built on Tinker and SDPO. It has tools (read, write, edit, edit_lines, glob, grep, bash), presents one tool call at a time, and you approve or deny each one. When you deny with a correction, it takes a gradient step on LoRA and retries with updated weights.

When you deny, the same model re-scores its own tokens with your correction as extra context. The logprob difference at each position is a per-token advantage — O(N) bits of signal from one correction, not the O(1) you get from a scalar reward. No separate reward model or critic: the model conditioned on your feedback is the teacher (SDPO, Hübotter et al. 2026).

You: "fix the test"
Agent: write(test.py, ...)       # overwrites the file
You: n → "use edit_lines; don't overwrite"
  → gradient step, retry
Agent: edit_lines(test.py, 14, 17, ...)
You: y

pip install continualcode
export TINKER_API_KEY=<your-key>
continualcode

Training runs on Tinker's API so you don't need a local GPU. You can also run inference-only with enable_training=false.

continualcode model_name=Qwen/Qwen3-30B-A3B-Instruct-2507
continualcode lora_rank=64 save_every=10
continualcode teacher_regularization=trust-region teacher_update_rate=0.1

• read — read a file with line numbers
• write — write content to a file, creating it if needed
• edit — find and replace a string (must be unique unless all=true)
• edit_lines — replace a line range with new content
• glob — find files by pattern
• grep — search files for a regex
• bash — run a shell command

• train.py — builds the teacher prompt from your correction, scores tokens under both contexts, computes per-token advantages from the logprob gap, IS-weighted gradient step through LoRA
• tui.py — interactive CLI where you approve/deny/edit tool calls
• tools.py — tool implementations
• benchmarks/auto_train.py — automated evaluation on LiveCodeBench

Full explanation →

• SDPO — Hübotter et al. 2026
• SDFT — Shenfeld, Damani, Guestrin 2026
• GKD — Agarwal et al. 2023
• Tinker — training API
• Design doc — full reasoning