Two user-level Codex skills for carrying project context across separate coding sessions.
The pair implements a small handoff ritual:
end-this-session: close the current session, summarize progress, and write project-local handoff artifacts.start-previous-session: begin a later session by reading those artifacts and producing a concise briefing before any implementation work starts.
The skills themselves live at the user level, while the handoff artifacts they read and write live inside the active project.
Long-running work often spans multiple assistant sessions. Chat history alone is fragile: the next session may not know which task is active, what failed, what is blocked, or which notes are durable project history versus temporary TODO state.
This repository separates those concerns into three state classes:
- Task protocol: volatile current task state, such as active goal, completed work, blockers, next steps, and constraints.
- Runtime / cognitive state: machine-observed JSON state, such as failed commands, recovery mode, and attempted approaches.
- Session logs: historical Markdown records that can safely become durable memory after a session ends.
The central rule is simple: current task state and runtime JSON are not long-term semantic memory. Session logs and crystallized decisions may be.
skills/
├── end-this-session/
│ ├── SKILL.md
│ └── agents/openai.yaml
└── start-previous-session/
├── SKILL.md
└── agents/openai.yaml
SKILL.md contains the actual workflow instructions. agents/openai.yaml contains UI metadata such as display name, short description, and default prompt.
Copy both skill directories into the user-level Codex skills folder:
mkdir -p ~/.codex/skills
cp -R skills/end-this-session ~/.codex/skills/
cp -R skills/start-previous-session ~/.codex/skills/After installation, use them from any project session.
End a session:
[$end-this-session](~/.codex/skills/end-this-session/SKILL.md)
/end-this-session
Natural-language variants also work, for example:
[$end-this-session](~/.codex/skills/end-this-session/SKILL.md)
今天到这，帮我交接
Resume a session:
[$start-previous-session](~/.codex/skills/start-previous-session/SKILL.md)
/start-previous-session
Or:
[$start-previous-session](~/.codex/skills/start-previous-session/SKILL.md)
继续上次任务，先给我 briefing
When a project has no existing convention, the skills use or look for paths like:
docs/session/task-protocol.yaml
docs/session/runtime-state.json
docs/session/logs/YYYY-MM-DD-HHMM-session-log.md
docs/session/index.md
docs/session/wisdom-log.md
docs/session/error-journal.md
docs/checkpoints/
Existing project conventions should take priority over these defaults.
- Do not index or embed task protocol files.
- Do not index or embed runtime/cognitive JSON.
- Do not treat current TODOs as durable project memory.
- Do not copy secret-like values into logs or memory.
- Do not commit or push during handoff unless explicitly requested.
- Prefer concise handoff artifacts over large, noisy logs.
The design and creation of these skills referenced this Zhihu article: