session-end

name	session-end
description	Structured end-of-session ritual — task audit, retrospective, memory update, skill creation, dream compression, and next-session planning.
user-invocable	true

Session End

A 6-step closing ritual that ensures nothing gets lost between conversations.

Step 1 — Task Audit

Review every task the user explicitly requested this session. List each with status:

• Task name → ✅ Done / ❌ Not done / 🔄 Partial

If any are incomplete or partial, surface them immediately and wait for user confirmation before continuing.

Step 2 — Retrospective

Focus: find mistakes, not celebrate wins. One sentence on outcomes, then enumerate errors:

• Wrong judgment calls (acted when should have asked)
• Known pitfalls walked into anyway (documented in memory/CLAUDE.md but not anticipated)
• Roundabout tool/approach choices (took 3 steps where 1 existed)
• Agent prompt design flaws (caused repeated reports, unclear scope, etc.)
• Scripts/code without defensive design (missing precondition checks, unhandled failure states)

If no mistakes: say "No significant mistakes." — do not skip this step.

New error patterns found that haven't been seen before and could recur → append to feedback memory in Step 4.

Step 3 — Decide What to Update

Scan the conversation. For each category, decide independently:

Memory files (almost always needed)

• Project state changed? → Update corresponding project_*.md
• New feedback/lessons? → Write or update feedback_*.md
• New external resources? → Write reference_*.md
• User info changed? → Update user_profile.md

Global CLAUDE.md (only if: affects behavior across all projects)

• New rule that applies everywhere?
• User corrected a global behavior?
• New disk/tool/env conventions?
• → Yes: edit ~/.claude/CLAUDE.md

Project CLAUDE.md (only if: affects current project conventions)

• New dev norms or architecture decisions for this project?
• → Yes: edit the project's CLAUDE.md

New Skill (needs at least 2 of these):

• Used a reusable multi-step workflow this session?
• Future conversations likely need it again?
• No existing skill covers it?
• → Yes: create using the appropriate skill-authoring tool

Step 4 — Execute Updates

Execute each item from Step 3 in sequence. Brief confirmation after each.

Step 5 — Memory Compression

Call the dream skill (or equivalent memory compression tool) to deduplicate and compress memory files.

Note: dream is a Claude Code–specific skill. On other agents, skip this step or replace with your own memory compression workflow.

Step 6 — Next Session Plan

List the top todos for next session (max 5, only genuinely important ones).