monitor-experiment

name	monitor-experiment
description	Monitor running experiments, check progress, collect results. Use when user says "check results", "is it done", "monitor", or wants experiment output.
argument-hint	[server-alias or screen-name]
allowed-tools	Bash(ssh *), Bash(echo *), Read, Write, Edit

本整合包适配：涉及 mcp__codex__codex / mcp__manual_review__* 的评审调用，按 shared-references/reviewer-adapter.md 的后端优先级适配（MCP → codex/gemini CLI → Devin 子会话 → 新对话人工中转 → 同模型降级并标注）。引用 tools/experiment_queue/* 的脚本位于本包 tools/ 目录。未收录的 skill 引用见 shared-references/pack-mapping.md。

Monitor Experiment Results

⏱ External cadence is appropriate here. This skill waits on an external fact (job completion / progress), so it is a natural /loop / CronCreate surface: the wake reads status and self-judges only machine-checkable completion (exit code, file exists, epoch logged) — never quality. This is the additive external-wait shape in shared-references/external-cadence.md. If a scheduled wait here ends in a verdict step (e.g. then audit results), run that verdict once after the wait clears — not re-entered per tick.

Monitor: $ARGUMENTS

Workflow

Step 1: Check What's Running

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 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

Step 6: Feishu Notification (if configured)

After results are collected, check the optional notification config (platform-dependent; e.g. Feishu via ~/.claude/feishu.json on Claude Code — on other platforms skip unless an equivalent channel is configured):

• Send experiment_done notification: results summary table, delta vs baseline
• If config absent or mode "off": skip entirely (no-op)

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