Milestone checkpoints before agent fan-out

[dean0x]

Summary

Save structured state checkpoint before spawning parallel sub-agents, so that agent failure or session crash doesn't lose pre-fan-out context.

Key Features

• PreToolUse hook matching on Agent|Task tool calls
• Captures: current task state, key decisions, modified files, agent dispatch intent
• Cooldown + dedup logic to prevent redundant checkpoints on rapid successive spawns
• Survives compaction and session recovery

Competitive Reference

Ref: SIEVE — hooks.json:22-30, measure.py:11985-12023. Checkpoint before Agent|Task fan-out with cooldown+dedup.

Implementation Notes

• Relevant to all orchestration skills (implement:orch, review:orch, plan:orch, etc.)
• Could be a hook (automatic/transparent) or a convention in orchestration skills (explicit)
• Needs discussion: hook-based vs. skill-based approach — hook is zero-effort for skill authors but less controllable
• Related to compaction checkpoint — shares state extraction logic
• Risk: adds latency before every agent spawn if not careful

Milestone

post-v2 (for consideration)