lkml-digest

name: lkml-digest description: Summarize the last 24 hours (or a chosen window) of a lore.kernel.org mailing list. Use when the user asks for an LKML / kernel mailing-list digest, daily summary, top topics/subsystems, or "what happened on recently". Wraps the local lkml-digest CLI which reads from the lkml-tools cache.

lkml-digest — LKML daily digest

This skill turns the locally cached lore.kernel.org mirror into a summary by piping the lkml-digest CLI output into the model.

Arguments

The user may pass a language token as the first argument:

• en → write the final summary in English (default).
• zh → write the final summary in Traditional Chinese (繁體中文).

Additional free-form tokens after the language are treated as overrides for the list and/or window, e.g.:

• /lkml-digest zh — Chinese summary of the last 24h of lkml.
• /lkml-digest en linux-pm 7d — English summary of the last week of linux-pm.
• /lkml-digest zh linux-mm today — Chinese, today only, linux-mm.

If the first token isn't en or zh, fall back to: language = match the language of the user's invocation prompt; treat all tokens as list/window hints.

When to use

Trigger on prompts like:

• "summarize the last 24h of lkml"
• "what's hot on linux-pm this week"
• "give me a digest of "
• "top kernel subsystems in the last hours"

Strategy: filter cheap, then read deep

Don't pour every full mail body into context. Use the CLI's two-phase design:

1. --format compact prints one short metadata record per mail (subject, sender/recipient, time, reply count, Message-ID, commit) — no bodies. Scan this cheaply to decide what's worth reading.
2. --select-commit <commit>,… (or --select-msgid <id>,…) then re-fetches only the chosen mails in full, by the Commit: ids (read from the local mirror) or Message-ID: values (fetched from lore) the compact pass handed you. This keeps the expensive full-body read scoped to the handful of threads you'll actually summarize.

Steps

1. Build. Always run cargo build --release -p lkml-digest from the repo root first, so the binary tracks the latest source (it's a no-op rebuild when already current).

2. List & window. Default lkml / --since 24h. "this week" → 7d, "today" → --range today. Override on user request.

3. Phase 1 — compact scan (metadata only, no bodies):

   ./target/release/lkml-digest --list <LIST> --since <WINDOW> --format compact
   

   Output is a # … count=N header, then blank-line-separated records of Subject / From / To / Date / Replies / Message-ID / Commit. Replies = transitive in-window reply count for that subtree.

4. Rank & bucket. Reconstruct threads from Subject: ([PATCH vN M/K], Re:) and Replies:. Score by subject keywords, reply volume (>5 = hot), and notable maintainers in From: (Torvalds, Greg KH, akpm, Peter Zijlstra, tglx, Kicinski, Rafael, …). Drop bot traffic (syzbot, test robot) unless it's the story. Then spread picks across these buckets for breadth — don't let one hot subsystem dominate:

   • core — VFS, locking, generic kernel
   • mm — memory management
   • scheduler — sched core, load balancing, sched/pm interaction
   • pm / power — pm, cpufreq, cpuidle, thermal, OPP, suspend/resume
   • pci — pci, pcie, hotplug, ASPM
   • usb — usb, xhci, dwc3, typec
   • net — net(-next), mptcp, dsa, bpf
   • storage / fs — nvme, scsi, block, ext4/btrfs/xfs, nfs
   • other — rust, tracing, kvm, …

   3–5 mails per populated bucket. If a bucket has no hot thread, still pick one general, non-platform-specific mail (a pci: core change over a board DT patch; a usb: core fix over SoC phy glue). Skip a bucket only if truly empty — note the omission.

5. Phase 2 — fetch picks in full. Collect the chosen Commit: ids and pull their bodies in one call, same window:

   ./target/release/lkml-digest --list <LIST> --since <WINDOW> \
       --select-commit <commit1>,<commit2>,…
   

   (Use --select-msgid <id1>,<id2>,… instead when picking by Message-ID:.)

   Prints ========-separated blocks (headers, blank line, --, decoded body).

6. Summarize by bucket with the language template below. Headings and prose in the chosen language; technical identifiers (functions, hashes, subjects, maintainer names) verbatim.

   English (en):

   # Linux Kernel Mailing List Daily Digest — <YYYY-MM-DD>
   
   > Window: <start> — <end> (UTC+8) · list: <list> · <N> mails
   
   ## 🔴 Today's Highlights
   3–5 sentences on the day's most notable discussions or technical trends.
   
   ## <Subsystem>
   
   ### <original English subject>
   - **Importance**: 🔴 High / 🟡 Medium / 🟢 Low
   - **Topic**: 1–2 sentences on what's being discussed.
   - **Progress**: patch state, point of contention, or conclusion.
   - **Key participants**: A, B, C
   

   Traditional Chinese (zh):

   # Linux Kernel Mailing List 每日摘要 — <YYYY-MM-DD>
   
   > 涵蓋時間：<起始> — <結束>（UTC+8）· 來源：<list> · <N> 封信
   
   ## 🔴 今日亮點
   3–5 句話，說明當天最值得關注的討論或技術趨勢。
   
   ## <子系統 / List 名稱>
   
   ### <英文 subject 原文>
   - **重要性**：🔴 高 / 🟡 中 / 🟢 低
   - **核心議題**：（1–2 句，說明在討論什麼問題）
   - **進展**：（patch 狀態、爭議點、或結論）
   - **主要參與者**：A、B、C
   

   Importance dots map to the ranking above: 🔴 high = strong keyword hit and high reply count or notable maintainer; 🟡 medium = one strong signal; 🟢 low = included for breadth or because the user asked.

Notes

• The CLI only walks the latest local epoch. For longer windows (>1 quarter) some history may be missing — flag that to the user if count looks too small for the window.