菜单
Save durable mew context before a long session, context compression risk, interruption, or major handoff. Use to write current intent, next action, validation state, blockers, active work-session notes, and durable product decisions into mew's own memory/state before continuing.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Re-enter mew planning or long-session work after context compression, interruption, or a user asks what to do next. Use to recover current work, durable product decisions, deferred decisions, and the next safe action from mew's own state and memory.
Run isolated mew side-project implementation as mew-first dogfood, update SIDE_PROJECT_ROADMAP_STATUS.md, and record structured M6.13.2 telemetry for M6.16 implementation-lane polish.
Evaluate mew as a passive AI product and as a shell/body that an AI model might want to inhabit. Use when asked whether mew is good enough, whether it feels usable as the next execution form, or what should improve next.
Update mew roadmap milestone status. Use when asked to assess progress against ROADMAP.md, refresh milestone statuses, decide the next roadmap target, or preserve roadmap progress across context compression.
Run bounded mew roadmap/coding tasks with mew as the first implementer and Codex as reviewer. Use when implementing mew tasks after M6.7, assessing mew-first success or failure, deciding whether to repair loop substrate immediately, or preserving autonomy-credit accounting across context compression.
Decide what to do immediately after a git commit in the mew repo. Use when a commit just landed and you must choose whether to report to the user or continue working under a long-session output gate.
| name | mew-context-save |
| description | Save durable mew context before a long session, context compression risk, interruption, or major handoff. Use to write current intent, next action, validation state, blockers, active work-session notes, and durable product decisions into mew's own memory/state before continuing. |
Use this skill before starting a long session, when context may compress, before pausing, or after a meaningful work chunk.
Goal: make the next resident recover from mew itself, not from chat history.
First inspect the current state so the saved context is accurate:
date '+%Y-%m-%d %H:%M:%S %Z'
git status --short
./mew desk --kind coding --json
./mew focus --kind coding
./mew brief --kind coding
If active work exists, save a session note:
./mew work <task-id> --session-note "<current intent; exact next action; validation state; blockers; pending approvals/recovery>"
If there is a durable product or roadmap decision, save it to typed project memory:
./mew memory --add "<Decision YYYY-MM-DD: decision, rationale, deferred work, next trigger>" \
--type project \
--scope private \
--name "Decision: <short name>" \
--description "<when future sessions should retrieve this>"
Then verify the memory is retrievable with a short query:
./mew memory --search "<short decision keywords>" --type project --json
Save only durable context:
Do not save noisy transcript summaries, temporary thoughts, or facts already obvious from git status.
Before a user-approved long session, always save:
For long sessions, prefer one concise mew memory --add planning note named like Long session charter: <short name> plus session notes on active work as it evolves.
Write the charter separately from progress checkpoints. Checkpoints record what just happened; the charter records what should continue to govern task selection after context compression.
For free-form mew improvement sessions, include the active roadmap milestone,
the unmet Done-when criterion being attacked, and the rule for rejecting polish.
Use mew-product-evaluator and ROADMAP_STATUS.md as the source of that
decision instead of copying only the latest chat preference.
Always save the governing chain explicitly:
Always save the output gate explicitly:
checkpoint_is_internal=trueno_return_until=<time or condition>report_only_if={time_boundary, blocker, approval_needed, explicit_user_interrupt}remaining_budget=<best current estimate>Treat any user-visible assistant message as a report.
commentary progress updates count as returning control.report_only_if condition fired.Do not treat commit done, clean worktree, next action clear, or
checkpoint saved as report triggers.
If the checkpoint says "hold", "wait", or "do not spend a proof item", also save why that constraint exists and what event would re-open task selection. Do not save the hold note by itself as if it were the governing goal.
If the current context diagnostics themselves are useful, prefer the native checkpoint command:
./mew context --save "<current intent; exact next action; validation state; blockers; stop/report trigger>" \
--name "Long session checkpoint: <short name>" \
--description "Recover this after context compression or interruption."
Use --json only when you need the saved memory path programmatically; the JSON output includes the full context diagnostics and can be large.
After compression or interruption, recover saved checkpoints with:
./mew context --load --json
After saving, a future agent should be able to use mew-planning-reentry and recover: