session-end

name	session-end
version	1.0.0
lifecycle	experimental
type	workflow
category	session-management
risk_level	low
trust	supervised
description	Session wrap-up workflow — captures decisions, syncs memory, reports costs, updates tasks, and commits cleanly. Run before ending any session with meaningful work.
metadata	{"openclaw":{"emoji":"📦","os":["darwin","linux","win32"]}}
user-invocable	true

Session End

Role

You are a session wrap-up agent. Your job is to ensure nothing gets lost between sessions: decisions are logged, learnings are captured, work is committed, and the next session can pick up without re-discovery. You prevent the "what did I do last time?" problem.

When to Use

Use this skill when:

• Finishing a Claude Code session where meaningful work was done
• Switching away from a project mid-day
• Completing a multi-step task that produced decisions or learnings
• Before any extended break from a project

When NOT to Use

Do NOT use this skill when:

• The session was purely exploratory with no code changes
• Everything is already committed and pushed
• The user explicitly says they will handle wrap-up manually

Workflow Steps

Step 1: Inventory the session

• Run git status to identify all uncommitted changes
• Run git diff --stat to see the scope of changes
• Run git log --oneline origin/main..HEAD (or equivalent) to list unpushed commits
• Summarize: what was accomplished, what is pending

Step 2: Capture decisions

• Review the session for high-leverage decisions (architecture, scope, tooling, tradeoffs)
• If the /decision-log skill is available, prompt the user to log key decisions
• Decision candidates:
  • Technology or library choices
  • Scope cuts or expansions
  • Architecture changes
  • Deployment or infrastructure decisions
  • Tradeoffs made under time pressure
• Format: one-line summary per decision, with rationale

Step 2b: Capture reusable prompts

• Review the session for prompts that worked notably well — got it right first try, used a clever pattern, or solved a class of problem cleanly
• If /prompt-library is available, prompt the user to log them
• Prompt candidates:
  • Got the desired outcome in one shot with no rework
  • Demonstrated a reusable pattern (constraint sandwich, output-format spec, negative examples, step-by-step decomposition)
  • Solved a common class of task in a way worth keeping
• Skip when: prompts were one-off, needed multiple corrections, or are too context-heavy to generalize
• Format: P-YYYYMMDD-### in the catalog, with Category + Context fields