version: "1.0.0" name: luban-log description: Download logs from the Luban platform via luban-log-downloader. Use when the user wants to fetch/pull/download or analyze logs for a Luban deployment (appid like k8s-sv1/sv0/sv2/k8s-job/k8s-vj/k8s-serverless), with optional pod names and time range. disable-model-invocation: true

ref: https://cooper.didichuxing.com/didocs/2209761958693

Luban Log Downloader

Fetches serving/job logs from the Luban platform (ES-backed, paginated).

Credentials

Real credentials live in ~/.claude/skills/luban-log/.env.local.

Every Bash call is a new shell, so ALWAYS inline the env prefix:

source ~/.claude/skills/luban-log/.env.local && luban-log-downloader <args>

What the user provides

Only two things are required:

1. appid (required), e.g. k8s-sv1-fuhdba-1789476287590-decode
2. pod name(s) (optional but recommended, comma-separated to skip pod discovery)

Optional: time range (-s/-e), output dir (-o), keyword search (--search).

Standard command

source ~/.claude/skills/luban-log/.env.local && luban-log-downloader \
  -a <appid> \
  --pod-name "<pod1>,<pod2>,<pod3>" \
  -s "YYYY-MM-DD HH:MM:SS" -e "YYYY-MM-DD HH:MM:SS" \
  -o logs_<label>

Time formats accepted by -s/-e: epoch ms, "2026-02-26 06:14:05", ISO "2026-02-26T06:14:05", or date only. No time range = defaults to last 3 days.

PD serving naming convention (this project)

A disaggregated serving deployment splits into multiple appids:

k8s-sv1-<name>-<ts>-prefill   → prefill pods (e.g. 3 replicas)
k8s-sv1-<name>-<ts>-decode    → decode pods
k8s-sv1-<name>-<ts>-lb        → router/lb pods (sglang-router, Rust)

Pod names look like <appid>-<hash>-master-0. If the user gives an appid without pods and discovery is unreliable, list pods with kubectl-style naming or ask.

Rules

1. Always use a distinct `-o` dir per time window (e.g. logs, logs_early).

Output files are named by pod name only — re-downloading the same pod for a different window into the same dir silently overwrites the previous file.

1. Run in background when downloading multiple pods or a window > 1 hour

(each pod can be 50–60 MB per 3 h; pagination takes a while).

1. Verify after download: ls -lh <dir> plus head -1 / tail -1 of one

file to confirm the time range actually covered.

1. If exact pod names are unknown, omit --pod-name to let the tool discover

pods (it prints them); or ask the user.

Post-download analysis quick wins

• Every line has TWO timestamps: [collection time CST] [container time UTC]

— they differ by 8 h. Always use the first (collection, Beijing time).

• Error triage:

``bash grep -icE 'error|exception|fail|timeout|refused|abort' <log> grep -iE 'error|exception|fail' <log> | sed -E "s/[0-9]+/N/g" | sort | uniq -c | sort -rn # dedup message types ``

• Timeline of a keyword per minute: grep '<keyword>' <log> | grep -oE '^\[[0-9-]+ [0-9]{2}:[0-9]{2}' | uniq -c
• Prefer writing intermediate analysis output to /tmp files and reading them,

instead of piping huge grep results straight to stdout.

Examples

One decode pod, 3-hour window

source ~/.claude/skills/luban-log/.env.local && luban-log-downloader -a k8s-sv1-xxx-decode \ --pod-name "k8s-sv1-xxx-decode-cb5b02-master-0" \ -s "2026-09-18 17:00:00" -e "2026-09-18 20:00:00"

All 3 prefill pods at once, separate output dir

source ~/.claude/skills/luban-log/.env.local && luban-log-downloader -a k8s-sv1-xxx-prefill \ --pod-name "pod-a,pod-b,pod-c" -s "2026-09-18 17:00:00" -e "2026-09-18 20:00:00" \ -o logs_prefill

Search mode: find keyword, then download ±2 h context around matches

source ~/.claude/skills/luban-log/.env.local && luban-log-downloader -a k8s-sv1-xxx-decode --search "OOM" --context-hours 2

Reference

• luban-log-downloader -h for full flags (--page-size, --region, --luban-url)
• Default API: http://api-ml.intra.xiaojukeji.com (env LUBAN_URL to override)