面向 GPT-5.4 等强模型和熟练用户的轻量 AI Agent Harness / checkpoint-driven coding skill。默认用户已经把任务切到基本可执行的最小混沌单元；模型自行分解、探索与推进，人类通过最终目标、最小 spec、复述、checkpoint、证据验证与回写来低干扰控盘。
将 SDD-RIPER 方法论落地为严格可执行流程的重型 Harness 技能。用于辅助用户澄清最终目标、生成 codemap/context、拆分最小混沌单元、维护完整 spec、执行 RIPER 阶段门禁、阻塞高风险动作和沉淀可 new chat 恢复的本地任务轨迹。
Prepare seamless new-chat handoff packs when a user wants to start a fresh chat, continue elsewhere, pause a long task, recover from context decay, recover a lost conversation from local Codex or Claude Code logs, preserve reusable project knowledge in Markdown, or hand work to another agent. Generates a durable handoff document, optional project-level Markdown updates such as AGENTS.md/README.md/PROJECT_SPEC.md/PROJECT_MEMORY.md, and a paste-ready next-chat prompt grounded in current task state, local conversation records, source files, specs, codemaps, validation evidence, open risks, and constraints.
Generate, update, or drift-check agent-facing CodeMaps as progressive code terrain indexes for projects, features, capabilities, functions, modules, or bug chains. Use when Codex needs to inspect an unfamiliar codebase, map a feature/capability from entry to effect, create or update `mydocs/codemap/*`, answer `create_codemap`, `MAP`, `PROJECT MAP`, code terrain, impact-map, "where should the agent look first", check whether existing CodeMaps are stale after a diff, or prepare SDD-RIPER Research without loading the whole repository into context.