Imported from MusserLab/lab-claude-skills (skills/done/SKILL.md). Install upstream with npx skills add MusserLab/lab-claude-skills --skill done. Copyright stays with the author.

End of Session Wrap-up

When the user invokes /done, perform these end-of-session tasks in order.

0. Detect Project Type

Check the project's .claude/CLAUDE.md for a project-type: field:

• data-science — Full wrap-up including all steps
• general (or no field found) — Skip steps marked [Data Science only]

If no field exists, infer: renv.lock, outs/, or numbered XX_*.qmd scripts → data science. Otherwise → general.

Collaborator mode — <!-- done-mode: collaborator --> (optional):

• For repos you don't own (e.g. a student's project you're reviewing). Set it in your personal CLAUDE.local.md so it applies only to you, not the repo owner.

• In collaborator mode, /done keeps your session-management artifacts private to your clone:

  • Session log (Step 1b) → appended to CLAUDE.local.md, not the shared .claude/CLAUDE.md.
  • No auto-push (Step 4) → commit locally, then /done does the push-vs-hold triage for you and recommends a set to push (one-line rationale each); you approve or adjust rather than adjudicating file-by-file. Nothing is pushed without your OK, and nothing pushable is silently left behind.
  • Private plans/reports (.claude/local/) → update as normal but never commit them (the folder is gitignored locally); register them in CLAUDE.local.md, not the shared .claude/CLAUDE.md.

  Push-vs-hold triage — Claude runs this classification and recommends a push set; the user approves or adjusts (they don't decide each file). Default to a recommendation; only ask about genuinely ambiguous items:

  • Lean push (accurate, ready, helps the student/repo): shared docs (.claude/CLAUDE.md, project_plan.md/project_summary.md, docs/feedback/), shared inputs (data/), reference/lookup files, finished scripts or analysis, and doc-accuracy fixes.
  • Lean hold (not ready / personal): work-in-progress scripts/<you>/, drafts, prototypes — anything you'd only share once polished.
  • Already private (gitignored — never in a commit anyway): .claude/local/, outs/<you>/, CLAUDE.local.md.
  • Ask only about the genuinely ambiguous items — recommend the clear ones. The final call is always the user's.

• Granular alternative: <!-- done-autopush: false --> disables only the auto-push and leaves the rest normal.

Also read CLAUDE.local.md (gitignored, personal) for these markers — it overrides .claude/CLAUDE.md, so a collaborator can set personal wrap-up behavior without touching the shared repo.

0b. Load Project Extensions

Check if .claude/done_extensions.md exists in the project root. If it does, read it and incorporate its steps into the wrap-up at the points it specifies (e.g., "run after Step 2" or "run before Step 4"). Extension steps are additional — they never replace core steps.

If no extension file exists, skip silently.

1. Summarize Work and Decisions

Briefly list what was completed this session:

• Scripts created or modified
• Data files created or modified
• Documentation changes
• Key analytical or design decisions made

1b. Update Session Log in Project CLAUDE.md

Append a new entry to the Session Log section at the bottom of the project's .claude/CLAUDE.md. This is a rolling log of the last 5 sessions — it's the primary place future sessions look for "what to do next."

Collaborator mode (Step 0): append the session log to your private CLAUDE.local.md instead of the shared .claude/CLAUDE.md — same format and same last-5 trim — so your entries stay in your clone and don't reach the repo owner.

Format

## Session Log
<!-- Maintained by /done. Most recent first. Keep last 5 entries. -->

### YYYY-MM-DD — Short title
- **Plans:** [plan name(s) worked on, or "None"]
- **Work:** [1-2 sentences on what was done]
- **Next:** [bullet list of follow-up items for future sessions]

Rules

• Most recent first — new entry goes at the top of the list
• Same-day updates — if an entry for today's date already exists (from an earlier /done run in this session), replace it rather than adding a duplicate. Merge the work descriptions and update the Next items. This allows /done to be run multiple times per session without cluttering the log.
• Trim to 5 entries — delete the oldest entry if there are more than 5
• Short title should disambiguate from parallel sessions (e.g., "Metadata exploration", "WGCNA figures", "Environment setup")
• Same-day session numbering — when multiple sessions occur on the same date, number them: ### 2026-03-22 (session 1) — Title, ### 2026-03-22 (session 2) — Title, etc. This makes the log scannable when intensive work produces multiple sessions per day. The number reflects chronological order within that day.
• Plans line — list which planning documents were worked on (e.g., "figure_plan.md"). Write "None" if no plans were involved. This helps track work that falls outside of plans.
• Next items — these are the main value. Be specific and actionable. They accumulate across sessions until actually done — don't repeat items already listed in a recent entry unless context changed.
• If the Session Log section doesn't exist yet, create it at the bottom of the file (before any closing comments).

2. Update Relevant Planning Documents

Only update planning documents directly relevant to this session's work. Do NOT read all registered documents — identify the 1-2 that matter from session context.

For each relevant planning document:

a. Status tables — Mark completed phases/tasks. Add new phases if needed. When a phase is marked complete, collapse its detailed content to a 3-5 line summary (what was done, key outcomes). Before collapsing, ensure any forward-looking information needed by future phases is already captured elsewhere in the plan.

b. Script and file tracking — Add new scripts, mark replaced ones as legacy, update modified entries.

c. Task lists — Check/uncheck items as appropriate.

Show proposed changes before editing.

Also check Claude Code plan files

If a plan file from plan mode exists in ~/.claude/plans/ and is relevant:

• Update status tables, file tracking, task checklists
• If it contains significant multi-session tracking, suggest promoting to a registered .claude/ document

3. Update Project Documentation (if needed)

Only do these checks if the session actually changed something relevant. Skip silently otherwise.

Canonical Data Files Registry [Data Science only]

If canonical data files were created, replaced, or had format changes:

• Update CURRENT/Legacy status entries
• Add new files that should be tracked
• Show proposed changes before editing

Project CLAUDE.md

If new conventions, gotchas, or important patterns were discovered this session, propose additions. Only record things that are surprising or counter-default — skip anything Claude would infer from the codebase.

MEMORY.md

If reusable lessons or gotchas were discovered, or existing entries proved wrong/outdated, propose updates. Keep under 200 lines.

New .claude/*.md files

If new .claude/*.md files were created this session, add them to the Project Document Registry.

CHANGELOG.md

If the project has a CHANGELOG.md in its root directory:

• Review what was done this session
• Propose a changelog entry under today's date, using the existing format (Added/Changed/Fixed/Removed sections as appropriate)
• If today's date already has an entry, append to it rather than creating a duplicate
• Show proposed changes before editing
• If no CHANGELOG.md exists, skip silently — do not suggest creating one

3b. Capture SLURM Resource Profiles [Cluster only]

Run only on the HPC cluster. Guard: if command -v sacct fails (e.g., on local macOS), skip this step silently — it does not apply off-cluster.

If SLURM batch jobs were submitted and completed this session, capture their observed usage so future jobs can be right-sized instead of re-discovering it:

1. For each distinct tool / job type run this session, pull empirical usage: jobstats <jobid> (or seff <jobid> / sacct -j <jobid> --format=Elapsed,MaxRSS,ReqMem,AllocCPUS,State). Record peak memory, CPU efficiency, elapsed walltime, and the input scale (e.g. read/sequence count) — the scale is what makes the number reusable.
2. Record the right-sized CPUs / Memory / Time / Partition (with observed peak mem + efficiency + input scale) where your project keeps its resource notes — e.g. a comment in the batch script or a project planning doc — so the next run starts from the right ballpark.
3. Only bother when the numbers refine what you already knew — don't churn notes for trivial re-runs or jobs you didn't right-size.

4. Git Commit

Project repository

Run git status to check for uncommitted changes.

If there are changes:

• CRITICAL: Only include files actually created or modified during THIS session
• CRITICAL: Always stage specific files by name (git add file1 file2), NEVER use git add . or git add -A for the project repo — broad staging will pick up changes from parallel Claude Code sessions
• Identifying session files — use multiple sources, not just git diff:
  1. Conversation context is the primary record of what you did — including compacted/summarized earlier portions of the conversation
  2. Git diff (initial gitStatus vs current git status) is a cross-check — files appearing in current but not initial are candidates but not certainties
  3. Parallel conversations can create files — another Claude Code session running simultaneously may have added files that show up in your git diff but were NOT created by this conversation
  4. When uncertain about a file's origin, ask the user — do not silently include or exclude files you can't account for from conversation context
• Files already in initial gitStatus should NOT be included unless you worked on them
• Show the user session-relevant files and suggest a commit message
• If approved, commit only those files
• After committing, push to remote automatically (git push) — unless collaborator mode or done-autopush: false is set (see Step 0; typically in CLAUDE.local.md for collaborative repos). If auto-push is off, do NOT push automatically — but DO surface and pre-triage the decision: gather every local commit not yet on the remote (git log @{u}..HEAD --oneline), this session's and any earlier unpushed ones, classify each with the push-vs-hold heuristics (Step 0), and present a recommended push set and a hold set with a one-line reason each. Ask the user to approve or tweak the recommendation — Claude does the legwork; the user doesn't adjudicate every file. The final call is theirs. Only ask the user if an attempted push fails (e.g., rejected by remote, no remote configured).

User config (~/.claude)

Check for uncommitted changes:

git -C ~/.claude status --short

If there are changes (skills, CLAUDE.md, etc.):

• Show what changed
• Ask if user wants to commit. Push automatically after — only ask if push fails
• If yes, run these as separate sequential Bash calls (do NOT chain with &&):
  1. git -C ~/.claude add -A
  2. git -C ~/.claude commit -m "Title" -m "Co-Authored-By: Claude <noreply@anthropic.com>" (use multiple -m flags, NOT heredocs)
  3. git -C ~/.claude push

Skill registration check: If new skills were created in ~/.claude/skills/ this session, verify each appears in the Available Skills table in ~/.claude/CLAUDE.md. If any are missing, add them before committing.

Sync reminder: After committing, check if ~/.claude/.sync-canonical/ exists (indicates a two-repo sync setup). If it does, compare the cluster repo HEAD against the canonical staging HEAD. If they differ:

"You have changes in ~/.claude that haven't been synced to canonical. Run /sync-cluster now to review and push them." If on the cluster, offer to run it immediately. If on macOS (canonical machine), remind the user to run /sync-cluster on the cluster next time they're there — macOS mode only pulls, it can't push cluster changes to canonical. Do NOT run /sync-cluster automatically — it's an interactive skill that requires decisions about what to push/modify/skip.

Conditional: Conda environment drift check [Data Science only]

Only if conda or pip packages were installed/updated this session. Do NOT blindly overwrite environment.yml (it's hand-curated and grouped) — detect drift and reconcile.

1. List what's actually in the env — both conda and pip:

   conda env export --from-history                          # conda (explicit installs)
   conda list | awk 'NR>3 && $NF=="pypi" {print $1"=="$2}'  # pip-installed (channel = pypi)

   --from-history does not include pip packages — that's why the second command is required. Skipping it silently drops pip installs from the spec.
2. Compare the union against environment.yml (its dependencies: list and its pip: subsection). Identify packages in the env but not in the spec (and vice versa).
3. If drift exists, show the diff and propose the exact lines to add, then ask: "environment.yml is missing these packages — add them?"
   • conda packages → add under dependencies:
   • pip-only packages → add under a pip: subsection (ensure pip is a conda dependency)
4. Apply the agreed edits to the hand-curated file (preserve grouping/comments). Keep out: the prefix: line and defaults channel; verify conda-forge is present.
5. Include the updated environment.yml in the commit.

Reconcile the curated file rather than replacing it with a full conda env export, which records pinned, platform-specific build strings that aren't portable.

Conditional: renv snapshot [Data Science only]

Only if R packages were installed or updated during this session:

Rscript -e "renv::status()" 2>/dev/null

If out of sync, ask about renv::snapshot(). Include updated renv.lock in the commit.

Conditional: Quarto publish

Check if this is a publishable Quarto project (all three must be true):

1. _quarto.yml exists
2. It's a book or website (type: book or type: website)
3. gh-pages branch exists

If yes, ask: "This is a Quarto book/website with GitHub Pages. Publish now?"

Do NOT offer for regular .qmd analysis scripts, projects without gh-pages, or type: default.

5. Final Summary

Brief "Session complete" message listing:

• Files created/modified
• Commits made
• Planning documents updated
• Any follow-up items for next session