Monitor Experiment Results

Monitor: $ARGUMENTS

Workflow

Step 1: Check What's Running

Queue-batch context (when invoked as part of an /experiment-queue batch). If the caller invokes this skill with a job-ID that belongs to an active /experiment-queue batch, the batch's queue_state.json (under $REMOTE_RUN_DIR/queue_state.json on the SSH host) is the authoritative source for that job's status, started, and completed timestamps — the queue scheduler has already verified expected_output exists. Phase 3.6 below still runs (it's how cost.json gets finalized), but Step 1 / 2 / 3 may use queue_state.json to skip the screen -ls + log-parsing path for jobs the scheduler has already marked terminal. For standalone use (no active queue), Step 1's screen -ls is the canonical check.

SSH server:

ssh <server> "screen -ls"

Vast.ai instance (read ssh_host, ssh_port from vast-instances.json):

ssh -p <PORT> root@<HOST> "screen -ls"

Also check vast.ai instance status:

vastai show instances

Modal (when gpu: modal in CLAUDE.md):

modal app list         # List running/recent apps
modal app logs <app>   # Stream logs from a running app

Modal apps auto-terminate when done — if it's not in the list, it already finished. Check results via modal volume ls <volume> or local output.

Step 2: Collect Output from Each Screen

For each screen session, capture the last N lines:

ssh <server> "screen -S <name> -X hardcopy /tmp/screen_<name>.txt && tail -50 /tmp/screen_<name>.txt"

If hardcopy fails, check for log files or tee output.

Step 3: Check for JSON Result Files

ssh <server> "ls -lt <results_dir>/*.json 2>/dev/null | head -20"

If JSON results exist, fetch and parse them:

ssh <server> "cat <results_dir>/<latest>.json"

Step 3.5: Pull W&B Metrics (when wandb: true in CLAUDE.md)

Skip this step entirely if wandb is not set or is false in CLAUDE.md.

Pull training curves and metrics from Weights & Biases via Python API:

# List recent runs in the project
ssh <server> "python3 -c \"
import wandb
api = wandb.Api()
runs = api.runs('<entity>/<project>', per_page=10)
for r in runs:
    print(f'{r.id}  {r.state}  {r.name}  {r.summary.get(\"eval/loss\", \"N/A\")}')
\""

# Pull specific metrics from a run (last 50 steps)
ssh <server> "python3 -c \"
import wandb, json
api = wandb.Api()
run = api.run('<entity>/<project>/<run_id>')
history = list(run.scan_history(keys=['train/loss', 'eval/loss', 'eval/ppl', 'train/lr'], page_size=50))
print(json.dumps(history[-10:], indent=2))
\""

# Pull run summary (final metrics)
ssh <server> "python3 -c \"
import wandb, json
api = wandb.Api()
run = api.run('<entity>/<project>/<run_id>')
print(json.dumps(dict(run.summary), indent=2, default=str))
\""

What to extract:

• Training loss curve — is it converging? diverging? plateauing?
• Eval metrics — loss, PPL, accuracy at latest checkpoint
• Learning rate — is the schedule behaving as expected?
• GPU memory — any OOM risk?
• Run status — running / finished / crashed?

W&B dashboard link (include in summary for user):

https://wandb.ai/<entity>/<project>/runs/<run_id>

This gives the auto-review-loop richer signal than just screen output — training dynamics, loss curves, and metric trends over time.

Step 3.6: Finalize cost manifest (mandatory whenever a run completes)

For every run detected as completed (or failed) in Step 1, finalize its runs/<run-id>/cost.json (/run-experiment Step 5.5 wrote the initial stub with status: running):

1. Compute or fetch:
   • ended_at — UTC ISO 8601 timestamp at which the run terminated (from screen log timestamps, modal API, or the result JSON's writeback time).
   • wall_clock_seconds — ended_at - started_at in seconds (from the cost.json's started_at field). Round to 1 decimal.
   • gpu_seconds — for gpu_provider in {remote, local, vast}: wall_clock_seconds * len(gpu_ids). For modal: read from the Modal API's billing record (authoritative) — fall back to wall_clock_seconds * <tier_count> only if the API call fails.
   • gpu_hours — gpu_seconds / 3600, rounded to 2 decimals.
   • status — completed on clean exit, failed if the screen log shows a non-zero exit or the result file is missing.
2. Rewrite runs/<run-id>/cost.json with all fields populated (keep the started_at / gpu_ids / gpu_provider from the initial stub — do not invent new values).
3. Log [cost] run=<run-id> wall=<H>h gpu=<H>h status=<status>.

This file is the canonical record consumed by /auto-experiment Phase 5, /auto-iteration-loop Phase C bookkeeping (incrementing runs_total / gpu_hours_total), and /auto-verify budget tracking. Do not let downstream consumers parse free-form deploy logs — they read this file.

If cost.json is missing entirely (legacy run launched before this contract was active), write a best-effort version: derive gpu_hours from screen log start/end timestamps and label the file with "legacy": true so consumers can choose to trust it loosely.

Step 4: Summarize Results

Present results in a comparison table:

| Experiment | Metric | Delta vs Baseline | Status |
|-----------|--------|-------------------|--------|
| Baseline  | X.XX   | —                 | done   |
| Method A  | X.XX   | +Y.Y              | done   |

Step 5: Interpret

• Compare against known baselines
• Flag unexpected results (negative delta, NaN, divergence)
• Suggest next steps based on findings

Key Rules

• Always show raw numbers before interpretation
• Compare against the correct baseline (same config)
• Note if experiments are still running (check progress bars, iteration counts)
• If results look wrong, check training logs for errors before concluding
• Vast.ai cost awareness: When monitoring vast.ai instances, report the running cost (hours * $/hr from vast-instances.json). If all experiments on an instance are done, remind the user to run /vast-gpu destroy <instance_id> to stop billing
• Modal cost awareness: Modal auto-scales to zero — no idle billing. When reporting results from Modal runs, note the actual execution time and estimated cost (time * $/hr from the GPU tier used). No cleanup action needed