这是一篇由原始材料转换而来的阅读页，保留了源文件的主要结构，并补充了可追溯的来源说明与链接。

Framework 2 — Long-Running Multi-Session Engineering

Use this when a task will span many agent sessions or many context windows.

Intent

Preserve continuity without relying on memory.

This framework combines

• Anthropic-style handoff artifacts (init.sh, progress.md, feature_list.json)
• OpenAI-style docs map, plan files, and mechanized invariants

Required artifacts

• progress.md append-only handoff log
• feature_list.json with explicit end-to-end checks
• plans/active/<task>.md for complex tasks
• docs/index.md pointing to architecture and runbooks
• init.sh smoke baseline

Typical loop

1. Re-read the map and recent history
2. Re-establish a green baseline
3. Choose exactly one increment
4. Implement and verify
5. Leave crisp artifacts for the next shift

Best fit

• 6h+ tasks
• overnight runs
• large refactors
• work likely to be resumed by a different agent or model