Jsa Park

jsa-park — park the session before you walk away

Part of jsa-park (https://github.com/kansokusha2026/jsa-park), a companion to the JSA measurement scripts (https://github.com/kansokusha2026/jsa).

The CLI sibling jsa_handoff.py builds the same note after the fact, from logs — and its judgment fields stay blank, because logs record operations, not decisions. You are running inside the live session: the decisions are still in context, so you can draft those fields too. That is this skill's entire reason to exist. Use this skill while the session is alive; use the CLI when the session is already gone (usage limit, crash, closed window).

Respond, and write RESUME.md, in the language of the conversation (Japanese conversation → Japanese note).

Step 0 — break-even check

First determine (from what the user said, or by asking once): will they be back within about 1 hour?

• Back within ~1 hour → parking is usually not worth it. On a Claude subscription within plan usage, Claude Code asks for the one-hour cache TTL, so the cache typically survives short breaks and continuing the same session costs mostly cheap cache reads. Say so, and stop — unless the user explicitly wants to park anyway. (On an API key, usage credits, or a cloud provider the default TTL is five minutes, so the threshold worth parking at is much shorter.)
• Longer than ~1 hour, overnight, unknown, or a usage-limit lockout → proceed. After a long gap the cache has expired, and the next message in this session would re-register the whole conversation (hundreds of thousands of tokens in heavy sessions).

Step 1 — draft the note

Use this two-part structure — the same field layout as jsa_handoff.py, so the two tools' output stays interchangeable.

Part 1 — the operations (you know these directly from the session):

• Source line: parked from this live session, with date and time
• Where the work stands: tasks done / in progress / not started
• Files edited in this session
• The state at parking time (what was just finished, what was mid-flight)

Part 2 — the judgments (the part the CLI can never fill):

• Goal / 目的 — what this work ultimately achieves. One sentence.
• Decided / 決まったこと — ONLY items the user explicitly approved. For each item, add a short pointer to the user's approving words. If you cannot point to an approval, the item does not belong here.
• Proposed, not approved / 保留中の提案 — everything you proposed that the user has not (yet) approved. When in doubt, put it here, never in Decided.
• Next action / 次にやること — ONE concrete step, not "continue".

Optional fields — include only when the conversation actually contains them: Rejected alternatives (with the reason each was rejected) / 検討して採らなかった案, Changed premises / 前提が変わったこと, Unresolved flags / 未解決の指摘・違和感, Files not to touch / 触らないファイル.

End the file with:

• How to resume: start a fresh session and read this file first, confirming its contents before any work. (With jsa-resume installed, saying "再開します" / "resume" is enough.)
• Security note: the note contains file paths and work details, which can include credentials or internal URLs — check the contents before sharing.

Step 2 — confirm the one thing that matters

Show the draft (at minimum the four core fields) and ask the user to check exactly one thing: is anything under "Decided" actually still just a proposal? Move items per their answer. This mix-up is the main way a resumed session goes wrong, and it is the reason the two fields exist.

Step 3 — write, with backup, and stamp for the meter

Write the note to RESUME.md in the project root. If one already exists, rename it to RESUME.md.bak first (one generation kept, same convention as the CLI).

Then stamp the note so the savings can be measured at resume time:

python3 "<this skill's base directory>/jsa_meter.py" park --write

The stamp records this session's final context size (read locally from ~/.claude/projects/; nothing leaves the machine). At resume time, jsa-resume compares it against the fresh session's cold start and shows the user what the park saved, in tokens. If the script fails, say so and continue — the note itself is the deliverable.

Step 4 — release the session

Tell the user, plainly:

• This session can now be closed. Continuing to type here after a long break would rewrite the whole conversation.
• To resume: open a fresh session in this folder and say "再開します" / "resume" (with jsa-resume installed), or paste: "Read RESUME.md and confirm its contents first. Do not start working yet."
• Don't commit RESUME.md by accident — keep it (and RESUME.md.bak) in .gitignore.

What this skill deliberately does not do

• Keep the cache warm with periodic pings. A keep-alive re-send bills at the cache-read rate rather than the full input rate, so on models with cheap cache reads it can undercut a cold restart for a fairly long break. It never undercuts parking, though: pinging pays repeatedly to preserve the context that parking stops carrying, and the bill grows with every hour away.
• Detect your absence automatically — invoke it before you leave.
• Close the session for you.