TTT Terminal Learner

Act. Observe. Adapt.

A terminal agent that updates its fast weights from real terminal feedback—even at inference time. Built on Qwen3-8B, with native thinking and tool calling.

Task → Think & act → Execute in terminal → Learn from output → Next action

During training, observation loss updates fast weights in the last 9 of 36 layers. Expert action loss then backpropagates through those updates, teaching the model how to adapt. At inference, the backbone stays frozen; only fast weights change. No expert labels needed.

• Real feedback: teacher forcing on actual terminal output.
• Meta-learning: differentiable inner SGD, outer AdamW, and FlashBoB attention.
• Four-GPU training: layer model parallelism, resumable checkpoints, optional W&B.
• 445 expert trajectories: Terminus 2 + Claude Opus 4.6, including failed runs.

Quick start

Linux + CUDA, Python 3.11/3.12. The default configuration uses four visible GPUs in one process.

pip install -r requirements-training-cuda.txt
python scripts/prepare_data.py

# Validate the CUDA backend and cross-device gradients.
python -m pytest tests/test_model_parallel.py tests/test_flashbob.py -q

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py \
  --config configs/experiment/terminal_ttt.yaml

The data script downloads only the selected expert trajectories and verifies file hashes. Configure training in terminal_ttt.yaml; resume with --resume runs/agent.pt.

New training runs check complete action and observation examples against the model context capacity. Each trajectory keeps its valid prefix: the first oversized interaction round and all later rounds are excluded, without truncating text or changing downloaded data. The retained turn counts determine the learning-rate schedule and are saved for resume. Older checkpoints without this selection retain their original strict overflow behavior.

Run the agent

python inference.py serve --checkpoint runs/agent.pt \
  --model-parallel-devices cuda:0,cuda:1,cuda:2,cuda:3

Use the JSONL service or Python AgentSession: reset → act → execute externally → observe. Your terminal runner executes commands and returns their output.

Status: Research implementation. Full Qwen3-8B four-GPU memory usage and Terminal-Bench success rates have not yet been measured. Long-context meta-learning can still exceed GPU memory; a Harbor adapter is not included.

Training guide · Design notes (Chinese)

Inspired by TTT-Terminal-World.