Context Checkpoint

Context Checkpoint Skill

When to Use This Skill

Trigger this skill when any of the following are true:

• User says: "save context", "we're running out of tokens", "checkpoint this", "save progress", "I want to continue in a new window"
• User types a slash command: /save, /resume [code], /resume light, /bookmark "name", /list bookmarks, /recover_last, /merge, /diff, /export
• Token usage is estimated at >= 80% of the context window (auto-trigger; show dashboard first)
• The conversation has exceeded ~40 messages or ~15,000 words
• A complex task (coding, research, debugging) is mid-flight and must be resumed later

Read the relevant reference files before acting:

Core Commands

Save Pipeline (Overview)

When /save is triggered, follow these steps in order:

Step 1 — Show Health Dashboard

Before saving, always show the Context Health Dashboard (see references/dashboard.md). This lets the user decide compression level.

Step 2 — Estimate Output Size

Before compressing, tell the user:

"Saving will produce approximately N characters (~X KB). Choose compression level: low (fast, full fidelity) / medium (drop low-value turns) / high (aggressive summarisation + code hashing)."

Default to medium if the user does not respond.

Step 3 — Compression and Selective Forgetting

Read references/compression.md. Drop low-value messages. Optionally hash long code blocks. Produce compressed conversation representation.

Step 4 — Assemble Checkpoint File

Build context-checkpoint.md with this exact top-level structure:

# Context Checkpoint: [Title]

Date: [ISO date]
Model: [model name or "unknown"]
Version: [vN — increment on each save]
Checksum: [SHA-256 of Sections 1-6, computed last]
TTL: [expiry date if set, else "none"]
Summary: [one-sentence description]

## 1. Objectives
## 2. Key Decisions and Conclusions
## 3. Artifacts [full code reproduction]
## 4. Current State [what works / broken / last error]
## 5. Open Issues and Next Steps [numbered, priority order]
## 6. Continuation Prompt [paste-ready, model-agnostic]
## 7. Pending Actions [unexecuted tasks, if any]
## 8. Bookmarks Index [named checkpoints, if any]
## 9. Redundancy Block [duplicate of critical facts for self-healing]
## 10. Checksum Block [SHA-256 verification]

For format details of each section, see references/save-pipeline.md.

Step 5 — Self-Healing Integrity

Compute SHA-256 of the full file content (Sections 1-9). Write it into Section 10 as:

SHA256: <hash>
REDUNDANCY: [repeat 3-5 most critical facts verbatim]

See references/recovery.md for checksum computation and repair logic.

Step 6 — Versioning

Save as context-checkpoint_v[N].md. Maintain rolling history of last 5 versions. Update _master_context_index.md with a one-line entry for this snapshot. See references/versioning.md.

Step 7 — Encryption (if requested)

If /save encrypt was used, apply encryption before writing to disk. See references/security.md.

Step 8 — Output and Instructions

Present the file for download. Then tell the user exactly:

1. Download context-checkpoint_vN.md
2. Open a new chat window
3. Paste the Continuation Prompt (Section 6) as the first message
4. Then paste or upload the Artifacts section (Section 3)
5. If resuming on a different AI model, paste the model-specific adapter from Section 6 (see references/cross-model.md)

Quality Checklist

Before finalising the file, verify every item:

• Token dashboard was shown before saving
• Output size was estimated and compression level confirmed
• All code artifacts reproduced in full — no truncation or placeholder comments
• Continuation prompt is self-contained and model-agnostic
• All error messages and tracebacks reproduced exactly
• Pending actions (unexecuted tool calls) captured in Section 7
• Checksum written and redundancy block populated
• Version number incremented; _master_context_index.md updated
• TTL set if user requested expiry
• Encryption applied if /save encrypt was used

Adaptive Behaviour

Track the following across sessions (store in _master_context_index.md):

• How often the user actually resumes from saved context
• Average token usage at time of save
• Preferred compression level

If the user has never resumed from a saved context across 5+ saves: stop auto-triggering at 80% — only save when explicitly commanded. If the user resumes frequently and the save threshold feels late: lower auto-trigger to 70%. Log the adapted threshold in _master_context_index.md under adaptive_config.

Emergency Hard Stop

If the token limit is hit unexpectedly mid-response:

1. Immediately save whatever is in the buffer
2. Mark the file header: [INCOMPLETE — possible message loss]
3. Log which messages may be truncated in Section 4 (Current State)
4. Output recovery instructions (see references/recovery.md)

Notes

• Never store encryption keys — user provides them each time
• For voice/multimodal sessions, see references/analytics.md for audio fingerprint guidance
• The skill works on any AI model — the continuation prompt in Section 6 is always model-agnostic
• For collaborative handovers, generate a read-only summary variant (omit Section 3 raw code if confidential)