Run Experiment

Run the provided command in a named tmux window with output logged to a file. Prevents context pollution and survives disconnects.

Naming Convention

Use sortable timestamps so alphabetical = chronological:

• •Session: experiments (shared across all experiments)
• •Window: <task>-<MMDD>-<HHMM> like train-gpt2-0127-1430, eval-mmlu-0215-0915
• •Log: tmp/<window-name>.log

Instructions

1. •

   Derive experiment name from the command + sortable timestamp:

   code

   "python train.py --model gpt2" → "train-gpt2-0127-1430"
   "uv run eval.py --dataset mmlu" → "eval-mmlu-0215-0915"
   "python sweep.py --lr 1e-4" → "sweep-lr-0303-2200"
   

2. •

   Create session and named window:

   bash

   EXP_NAME="<task>-$(date -u +%m%d-%H%M)"  # e.g., train-gpt2-0127-1430
   tmux has-session -t experiments 2>/dev/null || tmux new-session -d -s experiments -n default
   tmux new-window -t experiments -n "$EXP_NAME"
   

3. •

   Create log path and run:

   bash

   mkdir -p tmp
   LOG="tmp/${EXP_NAME}.log"
   PANE_ID="experiments:${EXP_NAME}.1"
   tmux-cli send "cd $(pwd) && <user-command> 2>&1 | tee $LOG" --pane="$PANE_ID"
   

4. •

   Report to user:

   code

   Experiment running: experiments:<EXP_NAME>
   Log: tmp/<EXP_NAME>.log
   
   Monitor with:
   - `tmux-cli capture --pane=experiments:<EXP_NAME>.1` - recent output
   - `tmux attach -t experiments` - live view (Ctrl+B, D to detach)
   - `tail -f tmp/<EXP_NAME>.log` - follow log
   

5. •

   Status Check (if requested):

   • •Use tmux-cli capture --pane=experiments:<EXP_NAME>.1 for recent output
   • •Use tail -50 tmp/<EXP_NAME>.log for log contents

Example

User: /run-experiment uv run python train.py --model gpt2-small --epochs 100

You run:

# 1. Set experiment name (sortable: MMDD-HHMM)
EXP_NAME="train-gpt2-$(date -u +%m%d-%H%M)"  # e.g., train-gpt2-0127-1430

# 2. Ensure session exists and create named window
tmux has-session -t experiments 2>/dev/null || tmux new-session -d -s experiments -n default
tmux new-window -t experiments -n "$EXP_NAME"

# 3. Run with logging
mkdir -p tmp
tmux-cli send "cd /path/to/project && uv run python train.py --model gpt2-small --epochs 100 2>&1 | tee tmp/${EXP_NAME}.log" --pane="experiments:${EXP_NAME}.1"

Then tell user:

Experiment running: experiments:train-gpt2-0127-1430
Log: tmp/train-gpt2-0127-1430.log

Monitor with:
- `tmux-cli capture --pane=experiments:train-gpt2-0127-1430.1` - recent output
- `tmux attach -t experiments` - live view (Ctrl+B, D to detach)

Multiple Experiments

Each gets its own named window:

# First experiment
tmux new-window -t experiments -n "train-base"
tmux-cli send "python train.py --model base | tee tmp/train-base.log" --pane="experiments:train-base.1"

# Second experiment
tmux new-window -t experiments -n "train-large"
tmux-cli send "python train.py --model large | tee tmp/train-large.log" --pane="experiments:train-large.1"

# List all experiments
tmux list-windows -t experiments

Cleanup

tmux kill-window -t experiments:train-gpt2  # Kill specific experiment
tmux kill-session -t experiments            # Kill all experiments

Notes

• •Named windows: Easy to identify which experiment is which
• •Persistent: Survives Claude session disconnects
• •Works inside/outside tmux: Same commands work in both modes
• •Monitorable: tmux attach -t experiments shows all experiment windows
• •Hydra experiments: logs also auto-save to out/*/main.log