Agent Skills: Checkpoint

Checkpoint

Save session context, print summary, update changelog. One command to wrap up a session.

What It Does

1. Save — writes session context to .agents/contexts/ (continuation prompt for LLM)
2. Summary — prints conversation recap to stdout
3. Changelog — updates CHANGELOG.md with current changes

Usage

/ag:checkpoint
/ag:checkpoint --compact    # Compact summary format

Workflow

Phase 1 — Save Context

Save current session state to .agents/contexts/ as a briefing for an LLM to continue work.

1. Analyze conversation and extract:

   • Context: What we're working on, relevant files/paths
   • Problem: What we were solving (specific, not abstract)
   • Decisions: Choices made and reasoning
   • Current State: Where we left off, what's done/not done
   • Next Steps: What to pick up, open questions

2. Generate filename

   • Format: {YYMMDDHHMMSS}-{slug}.md
   • Timestamp: current time as YYMMDDHHMMSS (no separators)
   • Slug: summary in lowercase, spaces to hyphens, max 50 chars
   • Run: fish -c "date '+%y%m%d%H%M%S'" to get timestamp

3. Check for existing contexts

   • Run: fish -c "ls .agents/contexts/*.md 2>/dev/null | tail -5"
   • If similar work exists, ask: "Update existing or create new?"

4. Write context to .agents/contexts/{timestamp}-{slug}.md

5. File format:

---
created: <ISO timestamp, e.g. 2025-01-30T14:23:15>
project: <project/repo name>
description: <one-line summary of what this context is about>
context: <free text - what this relates to: ticket, parent feature, epic, other contexts>
tags: []
---

# {Title}

## Context
{What we're working on, project, relevant files/paths with line numbers}

## Problem
{What we were solving - be specific, not abstract}

## Decisions
{Choices made and WHY - this is critical for continuity}
- Decision 1: chose X over Y because Z

## Current State
{Where we left off}
- Done: ...
- Not done: ...
- Blocked on: ... (if any)

## Next Steps
{What to pick up, in priority order}
1. First thing to do
2. Open questions to resolve

6. Confirm saved: print filename and path

Save Writing Guidelines

This context will be read by an LLM to continue work. Apply prompt engineering principles:

• Be specific: "auth module" → "src/auth/token.go:45-80, the refresh logic"
• Include reasoning: Don't just say what, say why decisions were made
• Provide context: File paths, line numbers, relevant code snippets if needed
• Clear next steps: Prioritized, actionable items
• No fluff: Every sentence should help the next session continue effectively

Phase 2 — Print Summary

Output conversation summary to stdout.

What Gets Captured

ALWAYS START WITH THE GOAL — answer "what was this session about?" first.

1. Goal/Context (REQUIRED) — what the user wanted, why the session happened
2. Work Done (REQUIRED) — what actually happened, what was built/fixed/explored
3. Outcome (REQUIRED) — current state, what's ready, what's pending
4. Key Decisions (OPTIONAL) — trade-offs, alternatives considered
5. Next Steps (OPTIONAL) — clear follow-up actions

Standard Format

## Summary

Goal: <what the user wanted to accomplish>

Work Done: <what actually happened>

Outcome: <current state>

Key Decisions: <only if meaningful>

Next Steps: <only if clear follow-ups exist>

Compact Format (--compact)

Single paragraph: goal + work + outcome.

Summary Rules

• Goal first — always answer "what was this session about?"
• Summary not changelog — focus on what user wanted, not file-by-file changes
• Why over what — explain reasoning, not just choices
• Skip minutiae — don't list every file read or test written
• Write for memory — reader should understand session purpose
• Concise — be succinct

Phase 3 — Update Changelog

Update CHANGELOG.md with current changes.

Step 1 — Determine Scope

git log --oneline --no-merges -20
git status --short
git diff --name-status
git diff --cached --name-status

• New commits since last changelog date: Add/update date entry
• No new commits, just uncommitted changes: Update today's entry

Step 2 — Gather What Changed

1. Git commits since last changelog date
2. Uncommitted changes — include in today's entry
3. Conversation context — decisions and rationale from session
4. Verify existence — confirm file/feature actually exists. Don't document ghosts.

Step 3 — Reconcile

Read existing CHANGELOG.md before writing:

1. Remove stale entries — "Added" but no longer exists = delete it
2. Remove reversed changes — added then removed = no entry
3. Update descriptions — if behavior changed, fix the description
4. Deduplicate — multiple commits for one thing = one entry
5. If today's date exists — update in place, don't duplicate

Step 4 — Write

If no CHANGELOG.md exists, ask before creating. Present proposed changes before applying.

Changelog Structure

# Changelog

## YYYY-MM-DD

### Added
- feature — brief description

### Changed
- what changed — why

### Removed
- what was removed

Changelog Rules

• Date-based — no semver, no [Unreleased], just YYYY-MM-DD
• Truthful — every entry verifiable against current codebase
• Remove when gone — deleted feature = delete its entry
• Update today, don't duplicate
• No fluff — no boilerplate, no preamble beyond # Changelog
• No trivial changes — skip typos, formatting, WIP commits
• Deduplicate — one feature = one entry
• Concise — one line per change

Output Format

The final output printed to the user should be:

---
Context saved: .agents/contexts/{filename}.md
---

## Summary

{summary content}

---

## Changelog

{changelog entries that were added/updated}

Rules

• Run all three phases in order: save → summary → changelog
• If changelog has nothing meaningful to add, skip it and note "No changelog updates"
• Save file is always written (that's the point of checkpoint)
• Summary is always printed
• Be concise across all phases — user can't stand verbose output