agentleFS
namedtoasternamedtoaster/The-Last-Mile/skills/session-end/SKILL.md

session-end

Skill39 starsChanged 22 days agoMITView on GitHub

Wrap up a working session automatically. Runs the project's checks, adds manual test items to the release checklist, records new gotchas and decisions, updates the handoff file and docs, checks progress against the current goal, commits only this session's files and pushes. Use when the user says they are done, wrapping up, or asks to end the session.

28 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use session-start.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • bitwize-music-studioSkill

    session-start

    Runs the session startup procedure - verifies setup, loads config and state, checks skill models, and reports project status. Use at the beginning of a fresh session.

    Also in 11 projects that use this

The file

We split this skill into the rules it gives an agent

12 rules. 0 are common, 8 are distinctive, and 4 are very specific to this project.

  • 0

    Common

    Most of the 21 edited versions we read keep these.

  • 8

    Distinctive

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  • could be automatic: a setting could enforce it
  1. Very specific:

    Run the check command [npm run check, or lint, typecheck and tests].

    line 16
  2. Distinctive:

    If this session changed anything a person has to see on a device or in a browser (a screen, layout, animation, navigation path, purchase, …

    line 25
  3. Distinctive:

    If something this session cost real time (it took more than one attempt to diagnose, or it would have caught the next person out), add a …

    line 43
  4. Distinctive:

    Skill problems: If a skill from The Last Mile did the wrong thing this session (a step that misled you, a missing step, a wrong default), …

    line 62
  5. Very specific:

    Always rewrite CURRENT_WORK.md: status, where things stand, next steps, and anything waiting on someone.

    line 75
  6. Distinctive:

    If a doc claims something is outstanding and you know it is done, correct it and write the date.

    line 78
  7. Distinctive:

    Run git status --short and confirm each file on your list is modified or new.

    line 100
  8. Distinctive:

    If one hunk mixes your lines with someone else's, or git apply refuses the patch, stop and tell the user instead of committing over their …

    line 111
  9. Distinctive:

    Never git add -A, git add ., git commit -a, or git add on a shared file.

    line 113
  10. Very specific:

    Run git diff --cached --name-only and confirm it lists exactly your files.

    line 114
  11. Very specific:

    Run git show HEAD and confirm it holds only your changes, and git diff to confirm the edits you left out are still there.

    line 119
  12. Distinctive:

    Push rejected because the remote moved: run git pull --rebase, then push again.

    line 128

“Common” is measured: kept by at least 6 in 10 of the 21 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Session end

Fully automatic. Do not ask wrap up questions. Do the work, then finish with a short summary. Read the Settings table in the project's rules file (CLAUDE.md or AGENTS.md) first; it says which check command to run, which docs to review and where to push. If there is no settings table, use the defaults in brackets below and say so in the summary.

1. Run the checks

Run the check command [npm run check, or lint, typecheck and tests]. Fix what fails and run it again until it is clean. Do not report the session as done while anything is red.

These are the automated checks only. Do not hand test features now; that is batched into the release checklist in step 2.

2. Release checklist

If this session changed anything a person has to see on a device or in a browser (a screen, layout, animation, navigation path, purchase, notification, deep link, or anything that behaves differently in a release build), add an item to Pending in RELEASE_TEST_CHECKLIST.md:

  • Surface: where it is.
  • Steps: how to reach it and what to do.
  • Expected: the result.
  • _(added YYYY-MM-DD)_

Tick any item verified this session, by the user or by you, with the evidence written into the item. Never clear the file wholesale. Never add an item that no available device can satisfy. For a web only project, the release is the deploy: word each item so it is checked on the live site once the change is deployed. Skip this step if nothing user facing changed.

3. Gotchas and decisions

If something this session cost real time (it took more than one attempt to diagnose, or it would have caught the next person out), add a dated entry to Gotchas in the rules file: the symptom, the cause, the fix and the general rule. Normal tool behaviour that worked as expected is not a gotcha. Most sessions add none. If the user made a decision that should not be re-litigated, add it to Standing decisions with the date and one line of why. Keep entries specific; a vague gotcha helps nobody.

Write what was observed, and mark a guess as a guess. A gotcha is read later as settled fact, and so is the handoff file, so both have to separate the two. If the cause was measured, say what was measured. If the fix has not been seen to work, because the checks were not re-run or the build that would prove it has not finished, say that in the entry and name the evidence that would settle it. Never write a theory as the cause before anything has confirmed it. The next person reasons from your entry instead of from the symptom, so a confident wrong entry costs more time than no entry at all, and it can send a later session to defend the wrong fix. A plausible mechanism, a precise sounding number nobody measured, and an untested fix are the three that read as fact and are not.

Skill problems. If a skill from The Last Mile did the wrong thing this session (a step that misled you, a missing step, a wrong default), and the user already asked for it to be fixed, run the improve-skill skill now. Otherwise list it in the summary under Skill problems, one line each: the skill, what it did, what it should have done. A project setting that was missing is a rules file fix, not a skill problem.

4. Docs

Review each doc listed in the settings [CURRENT_WORK.md, BACKLOG.md, FEATURES.md, docs/ALGORITHMS.md]. Update only what this session changed. Skip any listed doc that does not exist; do not create it just because it is listed.

Always rewrite CURRENT_WORK.md: status, where things stand, next steps, and anything waiting on someone. Keep it under 80 lines.

If a doc claims something is outstanding and you know it is done, correct it and write the date. Stale status docs send people chasing finished work.

5. Goals

Read GOALS.md.

  • No file, or no current cycle: say "No current goal set." in the summary, once, and move on.
  • Current cycle: tick any must do finished this session. If the end date is near and must dos are open, say so plainly in one line.
  • Cycle ended: give one short reminder in the summary that it is time to close this cycle and set the next one with the set-goals skill. Do not block or nag beyond that one line.

6. Commit only this session's files

The user may run several sessions in the same repo at once, sharing one working tree. A broad commit would sweep another session's changes into yours.

  1. Build the file list from your own edits in this conversation, including the doc updates above. Do not build it from git status alone.
  2. Run git status --short and confirm each file on your list is modified or new. Files that are dirty but not on your list belong to someone else; leave them.
  3. For each file, run git diff -- <file> and compare every hunk with your own edits. A file where every hunk is yours gets staged whole. A file that also has hunks you did not write (someone else's edits, or edits that were there before the session) gets staged one hunk at a time, so their edits stay in the working tree uncommitted: git add -- <whole file> <whole file> ... git diff -- <shared file> > /tmp/full.patch, then copy it to /tmp/own.patch keeping the diff/---/+++ header and only your hunks (each starts at an @@ line), then git apply --cached /tmp/own.patch. If one hunk mixes your lines with someone else's, or git apply refuses the patch, stop and tell the user instead of committing over their work. Never git add -A, git add ., git commit -a, or git add on a shared file.
  4. Run git diff --cached --name-only and confirm it lists exactly your files. Unstage anything else with git restore --staged -- <file>. Then commit what is staged: git commit -m "<plain sentence describing the change>". Do not name files after git commit: with file names, git commits the whole working copy of those files and sweeps in the edits you just left out.
  5. Run git show HEAD and confirm it holds only your changes, and git diff to confirm the edits you left out are still there. If something else got in, undo with git reset --soft HEAD~1 and commit again correctly.

7. Push and confirm the checks pass

Push to the destination in the settings [a branch].

  • No remote: skip the push and say so in the summary.
  • Push rejected because the remote moved: run git pull --rebase, then push again. Only your own unpushed commits get replayed. Never force push and never rewrite commits that are already on the remote. If the rebase hits a conflict, stop and tell the user.

Then check whether the repo runs checks on push (for example a .github/workflows/ folder). If it does not, skip this part. If it does, find the run for your commit (gh run list and gh run watch when the GitHub CLI is installed, otherwise a GitHub tool if one is connected) and wait for it. If it fails, read the actual failure, fix it, commit the fix the same way and push again. Stop after three failed attempts and tell the user what is still failing. If you cannot see the run at all, say so in the summary instead of claiming it passed.

8. Summary

Finish with a short summary, in the voice the rules file asks for. Keep any prose to a line or two, then end with one table, so the next action is visible at a glance. Rows that need the user come first and say so plainly; a question for the user is its own row, never buried in a paragraph.

ItemCategoryNext action
Rotate the API keyNeeds youOpen the dashboard and create a new key
Profile screen fixDoneNone

Categories: Needs you, Decision (the user picks), Done, FYI. The rows cover:

  • What was done this session.
  • The goal line from step 5.
  • Whether a version bump is due, and which kind (patch for fixes, minor for new features), if the app ships to a store.
  • Skill problems from step 3, if any, with the offer to fix them.
  • Anything the user needs to do by hand.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.