The script auto-enumerates every scalar metric the run actually logged and pulls all of them — the user does not need to know what was logged in advance. Only pass --keys "k1,k2,..." if the user explicitly wants a subset (e.g. "I just want train and val loss"). If auto-enumeration fails (SwanLab API drift, auth issue), the script falls back to a generic key list and prints a warning; surface that warning, but the dump itself still succeeds.

If the user hasn't named a project and SWANLAB_PROJECT isn't set, ask once: "Which SwanLab project?"

Workflow

1. Run the script with Bash, capturing full stdout/stderr.

1. If the script exited non-zero, show the last 10 lines of stderr verbatim, then stop. Common signals and the right hint:
• ApiError mentioning auth / 401 / netrc → tell the user to run swanlab login
• ModuleNotFoundError: swanlab (or pandas / omegaconf) → pip install swanlab>=0.7.15 pandas omegaconf
• RuntimeError: No experiments in ... → user named the wrong project, or hasn't pushed any run yet
• 404 on the experiment ID → ID typo or experiment was deleted
• Network errors → relay verbatim, don't speculate

1. On success, read /brief.md — the script writes a fixed-format Markdown brief containing run metadata, the latest non-NaN value of every scalar with its step, and the dump's row × col count. Don't reinvent this with awk / one-liners over metrics.csv; the brief is the contract. If brief.md is missing or empty, that's a bug in the script — surface it.

1. Tell the user what's in brief.md, in their language (mirror Chinese if they wrote Chinese). Keep it tight — one heading for the run, one short list / table for the latest scalars, end with the absolute path of the output directory so they can cd or open it. You don't need to dump every scalar if there are many; pick the ones the user is likely to care about (typically a train/loss, epoch/avg_loss, val/, plus anything they named earlier in the conversation).

Idempotency

Re-running the script for the same run (same --exp-id, or --latest resolving to the same run) overwrites files in the matching swanlog_/ directory rather than accumulating duplicates. The timestamp comes from run.created_at, not the wall clock at fetch time, so the directory name is stable across re-fetches. Safe to call repeatedly to refresh metrics on a still-running experiment.

Boundaries

• This skill only reads from SwanLab. It never starts training, modifies configs, deletes runs, or pushes anything to the cloud.
• It does not assume any particular ML framework, dataset, or directory layout — it works against any SwanLab project.
• metrics.csv only contains scalar columns (FLOAT / INTEGER). IMAGE / AUDIO / TEXT artifacts are skipped — they don't fit a flat CSV. If the user needs those, point them to the WebUI URL from the brief.