Wrap Up

Wrap-Up

Use this at the end of a meaningful conversation or when the user asks to ship, close out, commit, push, or prepare a handoff.

The job is not just repo cleanup. The job is to leave the workspace in a known state, avoid stepping on running schedulers, ship only intentional changes, and turn repeated friction into improvement candidates.

Opening

Start with Workflow: wrap-up. Then run the deterministic preflight before making any changes:

python3 ~/.agents/skills/workflows/wrap-up/scripts/wrapup_preflight.py --workspace "$PWD" --due-window-minutes 90

If this skill is running from a copied surface where ~/.agents is unavailable, run the copy-local script at scripts/wrapup_preflight.py.

Core Sequence

Run these phases in order.

1. Outcome State

Summarize the session's original objective and current state:

• what got done
• what did not get done
• whether the work changed scope
• whether the handoff is Complete or Continues

Do not call the session complete if the stated objective is unfinished.

2. Scheduler-Safe Preflight

Use the preflight script output as the collision map. At minimum, account for:

• user LaunchAgents in ~/Library/LaunchAgents
• user crontab from crontab -l
• Hermes cron jobs in ~/.hermes/profiles/*/cron/jobs.json
• Codex automations in ~/.codex/automations/*/automation.toml
• running Hermes gateways, dashboards, WebUI, Dev Secrets, history-search, and

other launchd-managed services when visible

Default safety rules:

• Do not restart, unload, kickstart, repair, or create background services

during normal wrap-up.

• Do not edit LaunchAgents, Hermes config.yaml, Hermes cron/jobs.json,

secrets runtime files, scheduler scripts, or automation definitions unless the session explicitly touched scheduler work.

• Treat scheduler output, logs, state snapshots, lock files, and generated

reports as owned by running processes unless they are part of this session's deliverable.

• If a scheduled job is unhealthy, record it as an improvement candidate unless

the user explicitly asked for scheduler repair.

• If a job is due soon, avoid broad staging. Stage only files known to belong to

this session.

If the preflight cannot inspect a surface because of permissions, report that surface as unverified. Do not claim it is clean.

3. Verification

Run the narrowest meaningful checks for the work done:

• tests for code changes
• build or typecheck for frontend/app changes
• script smoke checks for wrappers and automation helpers
• syntax checks for JSON, TOML, YAML, plist, shell, or Python files
• health checks for services only when the session touched those services

If verification is skipped, state why. A skipped check is not a pass.

4. Change Review

Inspect git before staging:

git status --short
git diff --stat

Classify changes:

• session changes to stage
• user or scheduler changes to leave alone
• generated artifacts to review before staging
• secret or credential-looking files to avoid and flag
• unrelated dirty files to preserve

Never use git add . or git add -A. Stage specific files only.

5. Commit

Commit only after verification and change review.

Use a descriptive message that says what shipped, not that the session ended. Prefer:

Add scheduler-safe wrap-up workflow
Repair Pam calendar auth probe
Update Jules Live Studio turn handling

Avoid vague messages like:

Wrap-up changes
Session updates
Misc fixes

Do not use --no-verify unless the user explicitly approves bypassing hooks and the reason is stated in the final report.

6. Push

Push the current branch to its configured upstream or to the branch requested by the user.

Rules:

• Do not force push.
• If push is rejected, report the reason and stop for review unless the user

already authorized pull/rebase behavior.

• Do not pull or rebase during generic wrap-up unless that behavior was clearly

approved for this session.

• If the user said "do not pull", preserve that constraint.

7. Open Loops and Commitments

List only real open loops:

• blockers
• next actions
• user commitments to another person
• outbound drafts needing approval
• calendar, email, social, or partner follow-ups
• background jobs or health issues discovered but not repaired

For a continuing handoff, include a resume prompt specific enough for a fresh session.

8. Improvement Candidates

Look for improvements in these categories:

• Skill gap: the assistant struggled or missed existing guidance.
• Friction: a repeated manual step should become a script, check, or skill.
• Knowledge: a durable fact should be captured in the right local doc.
• Automation: a scheduled job, wrapper, or health check could be hardened.
• Quality: tests, docs, or validation were missing for behavior that matters.

Do not automatically edit shared instructions during wrap-up. Present concrete candidates with suggested files or procedures.

Hermes profiles may run their own profile-native skill review/update procedure. Codex and Claude should follow this skill's guidance and propose shared skill edits for review before changing .agents or profile-specific skills.

9. Final Report

End with a compact report:

Wrap-up complete.

Verification:
- [check]: pass/fail/skipped

Git:
- branch:
- commit:
- push:

Scheduler safety:
- checked:
- due soon:
- unverified:

Open loops:
- ...

Improvement candidates:
- ...

Handoff: Complete - [one sentence]

If the objective continues, end with:

Handoff: Continues - [what remains]
Resume prompt: [exact prompt]

On Codex, after a successful stage, commit, push, branch creation, or PR creation, emit the relevant Codex app git directives in the final answer.

10. Optional Report Artifact

The conversational final report is always required. A file artifact is conditional.

Save a wrap-up report file when at least one is true:

• code, docs, config, or assets were committed or pushed
• a durable decision was made
• the work continues and needs a fresh-session resume prompt
• a background job, scheduler, automation, auth surface, or service was touched
• there are concrete improvement candidates worth reviewing later
• Jonathan explicitly asks for a report artifact

Do not save a file for trivial Q&A, quick lookups, or conversations with no state change.

When saving, use a run artifact path rather than a durable Knowledge-OS report:

System/Runs//wrap-up/YYYY-MM-DD/HHMMSS-.md
System/Runs//wrap-up/YYYY-MM-DD/HHMMSS-.json

Use the helper to create the artifact deterministically:

python3 ~/.agents/skills/workflows/wrap-up/scripts/wrapup_report.py \
  --workspace "$PWD" \
  --surface codex \
  --status Complete \
  --focus "Add scheduler-safe wrap-up workflow" \
  --summary "Created the shared wrap-up skill and helper scripts." \
  --verification "preflight helper runs" \
  --improvement "Review surface symlink plan before installing adapters" \
  --write

Report artifacts are receipts and ingestion inputs. They do not replace transcripts, git history, history-search, Pam reports, Subconscious notes, Researcher intake, Archivist distillation, Builder's Log, or future gbrain.

For gbrain, keep the artifact short and structured: session id if known, surface, status, focus, files/commits, verification, scheduler safety, open loops, improvement candidates, and resume prompt. Do not paste raw transcript content or secret values.

Promote only durable decisions, reusable facts, or stable operating rules into Knowledge-OS or memory. Do not treat every wrap-up artifact as durable knowledge.

Surface Adapters

The canonical skill lives at:

~/.agents/skills/workflows/wrap-up

Use the read-only target planner before installing or changing surface links:

python3 ~/.agents/skills/workflows/wrap-up/scripts/wrapup_surface_targets.py

Recommended adapter behavior:

• Codex: symlink ~/.codex/skills/wrap-up to the canonical skill. Codex

adds app git directives after successful git operations.

• Claude Code: symlink ~/.claude/skills/wrap-up to the canonical skill.

Claude uses plain final text and its own slash-command behavior.

• Hermes profiles: symlink

~/.hermes/profiles//skills/workflows/wrap-up for approved workflow-owning profiles only. Current live adapters are pam and director. Do not install this skill on archivist, dreamer, researcher, or subconscious unless Jonathan explicitly re-approves those profile surfaces. Hermes may run profile-native skill review before proposing shared edits.

• Public/reference Jules repo: copy or sync a sanitized copy into the repo

rather than symlinking to a host-local path.

Do not replace existing physical skill directories with symlinks without review.

Deterministic Helpers

• scripts/wrapup_preflight.py: read-only scheduler, git, and automation

preflight. It intentionally avoids prompts and secret values.

• scripts/wrapup_surface_targets.py: read-only inventory of surface adapter

targets and whether each is missing, a symlink, or a physical directory.

• scripts/wrapup_report.py: optional writer for compact Markdown plus JSON

wrap-up artifacts under System/Runs//wrap-up/.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jonathanmalkin
• Source: jonathanmalkin/jules
• License: MIT

Install and usage instructions live in the source repository linked above.