Wrap Session

Capture session state for seamless resumption.

Announce: "Wrapping session — saving context for next time."

The Iron Law

NO SESSION END WITHOUT WRAP OFFER

When you detect a session is ending (goodbye, thanks, natural stopping point, context pressure), proactively offer to wrap.

What to Capture

1. Progress

• What was accomplished this session
• Which tasks/stories are complete
• Current branch, commit, state

2. Decisions

• Architectural choices made (with rationale)
• Approach decisions (with alternatives considered)
• User preferences discovered

3. Blockers & Open Questions

• What's stuck and why
• Questions that need answers
• Dependencies not yet resolved

4. Next Steps

• Prioritized list of what to do next
• Which skill to invoke first in next session

5. Resume Prompt

Generate a self-contained prompt the user can paste into a fresh session. Use @ references for files so Claude reads them automatically.

@CLAUDE.md  Continue [PROJECT] - [CURRENT TASK]

Branch: `branch-name`

Completed:
- [completed items]

Next Task: [description]
1. [next task with context]
2. [following task]

Context Files:
- @path/to/design-doc.md (implementation plan)
- @path/to/relevant-code

Decisions to Remember:
- [key decision]: [rationale]

Start With:
/turbocharge:[skill] [args]

5.5. Atlas Freshness

If ATLAS.md exists and significant structural changes were made this session (new modules, changed entry points, new integrations), note in the resume prompt:

• "Consider running /turbocharge:atlas to update the domain map"

6. Encode Session Learnings

Before generating the resume prompt, check: did the user correct any misunderstandings or wrong approaches during this session? If so:

• Update memory files with corrections that apply to future sessions (domain concepts, preferences, conventions)
• Update CLAUDE.md if the correction reveals a missing rule or domain term that would prevent the same mistake
• Don't save ephemeral task details — only save what future sessions need to know

Examples of things worth encoding:

• "autotuning is proactive, not reactive" → domain term in CLAUDE.md
• "always use asyncpg, never psycopg2" → infrastructure note in CLAUDE.md or memory
• "don't ask permission to run tests" → rule in CLAUDE.md

7. Write Memory With Confidence Metadata

When persisting items to ~/.claude/projects/<project>/memory/*.md, use this bullet format:

- Summary sentence _(source, YYYY-MM-DD, conf: 0.8)_

• source — where the claim came from (websearch, user, codebase, testrun)
• YYYY-MM-DD — date written
• conf — 0.1 to 1.0; 0.9+ only for verified facts, 0.5–0.8 for reasoned claims, <0.5 for speculation

Example:

• Tailwind v4 drops the JS config (websearch, 2026-04-14, conf: 0.9)

8. 200-Line Cap on MEMORY.md — prune-before-build

Before appending new entries, if ~/.claude/projects/<project>/memory/MEMORY.md is at or over 200 lines:

1. Sort existing entries by (confidence asc, date asc) — lowest confidence and oldest first
2. Drop entries until under 180 lines
3. Then append new entries

Dropped entries are not archived separately — file-based memory is lossy by design; high-value entries survive by being re-surfaced each session.

9. Session Snapshot JSON

Write .claude/turbocharge-session.json in the project root (the standard Claude Code project-state directory — same place .claude/settings.json and .claude/settings.local.json live):

{
  "date": "2026-04-14",
  "branch": "master",
  "current_task": "ATLAS reshape design",
  "blockers": [],
  "next_steps": ["Chain to /turbocharge:plan", "Break into tasks"],
  "open_files": ["skills/atlas/SKILL.md"]
}

The SessionStart hook cats this on next session so resume is zero-tool-call. Make sure .claude/turbocharge-session.json is in the project .gitignore — it is per-user session-local state (don't exclude all of .claude/ because settings.json is team-shared).

How to Deliver

Output the resume prompt directly in chat so the user can copy it. Do NOT write files, create directories, or modify .gitignore (except for CLAUDE.md and memory updates above, and the session snapshot JSON).

Agent memory (via memory: project) handles persistence automatically — no manual file saving needed for session state.

Workflow Position

[any skill] → wrap → [new session] → [resume with prompt]