Source Code

Context Checkpoint Skill

Purpose: Save conversation state before context compression kills it.

The Problem

Context compression is unpredictable. One moment you're mid-conversation, the next you wake up with amnesia. Important decisions, open threads, and working context — gone.

The Solution

Proactive checkpointing. Save state regularly so when compression hits, you have something to reload.

Usage

Manual Checkpoint

When you're in a conversation and want to preserve state:

# Save current state
./skills/context-checkpoint/checkpoint.sh "Brief description of what we're doing"

Heartbeat Integration

Add to HEARTBEAT.md:

### Context Checkpoint
- If conversation has important open threads, run checkpoint
- Check memory/checkpoints/ for stale checkpoints (>24h old, can clean up)

On Session Start

Read the latest checkpoint:

cat memory/checkpoints/latest.md

What Gets Saved

The checkpoint creates a markdown file with:

Timestamp
Description (what you were doing)
Open threads / active tasks
Key decisions made
Important context to remember

File Structure

memory/checkpoints/
├── latest.md           # Symlink to most recent
├── 2025-01-30_1530.md  # Timestamped checkpoints
├── 2025-01-30_1745.md
└── ...

Security Considerations

Risk: Low. Only writes to local workspace.
No credentials: Doesn't touch external services.
No exec: Just file operations.
Blast radius: Worst case, fills up disk with checkpoints. Mitigated by cleanup routine.

context-checkpoint