checkpointing

Checkpointing (Codex Native)

Use the local checkpoint script directly and perform analysis in the same Codex session.

Usage

• /checkpointing
• /checkpointing --since YYYY-MM-DD

Workflow

1. Run checkpoint collection

   • Command:
     • python .claude/skills/checkpointing/checkpoint.py
     • or python .claude/skills/checkpointing/checkpoint.py --since YYYY-MM-DD
   • If python is unavailable, use:
     • python3 .claude/skills/checkpointing/checkpoint.py
     • or python3 .claude/skills/checkpointing/checkpoint.py --since YYYY-MM-DD

2. Verify generated artifacts

   • Confirm latest files exist:
     • .claude/checkpoints/*.md
     • .claude/checkpoints/*.analyze-prompt.md
   • Confirm CLAUDE.md session history was updated.
   • Extract the checkpoint UTC timestamp and reuse it for every derived artifact.

3. Analyze skill patterns without Claude subagents

   • Read the generated .analyze-prompt.md in the current Codex session.
   • Produce candidate skill patterns with confidence and evidence.
   • Save analysis to:
     • .claude/docs/research/skill-patterns-{checkpoint-timestamp}.md
   • Naming convention:
     • Use the same UTC timestamp as checkpoint filename (example: 2026-02-20-225248).
     • This avoids overwriting date-only reports.

4. Enforce output quality gate

   • Artifact consistency:
     • checkpoint, analyze-prompt, and skill-patterns files all share the same timestamp.
     • New timestamped report is created (skill-patterns-{timestamp}.md), not date-only overwrite.
   • Analysis structure (required):
     • Checkpoint Stats section exists.
     • Candidate Patterns section has at least 3 patterns.
     • Each pattern includes: Description, trigger phrases (JA/EN), Confidence (0-1), Evidence, Existing skill overlap.
   • Recommendation quality:
     • Explicitly decide: "create new skill" vs "strengthen existing skill".
     • Prioritize top 1-3 actions by impact and effort.

5. Report to user

   • Summarize checkpoint stats and top pattern suggestions.
   • Include output-quality gate result (pass/fail and missing items if any).
   • Ask whether to convert high-confidence patterns into new skills now.

Reporting Template

Use this minimal structure in the user-facing report:

1. Checkpoint result: branch, commits, tasks completed, generated artifact paths.
2. Quality gate: pass/fail with concrete checks performed.
3. Top pattern suggestions: top 3 with confidence and overlap note.
4. Recommended next action: single clear recommendation.

Command Contract

• /checkpointing