task-progress-stream

Use this skill when the user wants real-time progress updates inside the OpenClaw chat UI for a long-running task such as:

• model training
• evaluation
• inference
• data preprocessing
• long shell jobs
• benchmark runs

It works in two modes:

1. run
   Start a command, capture stdout/stderr, parse common metrics, and periodically inject progress summaries into chat.

2. tail
   Monitor an existing log file and periodically inject parsed progress summaries into chat.

What it extracts

It tries to recognize common fields from logs:

• epoch / max_epoch
• step / max_step
• loss
• learning rate
• validation metric
• best metric
• ETA
• speed

Typical usage

Start a training job and stream progress into chat

node ./scripts/task_progress_stream.js run \
  --session main \
  --label train \
  --cwd /path/to/project \
  --cmd "python src/train.py --config configs/train.yaml" \
  --interval-sec 20