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。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/appautomaton/mlx-speech --skill auto-resume命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| name | auto-resume |
| description | Recover active change and next action from artifacts. Use on fresh session with existing work. |
| metadata | {"stage":"resume"} |
Session recovery. Rebuilds context from durable artifacts, not memory or guessing.
First action: run node .agent/.automaton/scripts/get-context.mjs from the project root.
auto-resume rebuilds context from durable artifacts, not from the user's description or the agent's training data. It does not modify artifacts, advance the stage, or start new work. It loads canonical artifacts in dependency order (spec first, then design, then plan) and reports what it found, what was blocked, and what comes next.
Loading discipline: start with artifacts needed for the current stage. Read project files when understanding the codebase helps rebuild accurate context for the next action. Read .agent/.automaton/references/CONTEXT-BUDGET.md when wider reads threaten context pressure.
Before producing the recovery summary:
references/quality.md when the summary becomes narrative recap.Halt and report when:
.agent/ does not exist or current.json is missing.Recommend automaton install to scaffold .agent/, then stop. Do not attempt recovery without a state file.
If work is complete or absent, read .agent/steering/ROADMAP.md only to surface pending phases as context.
Check that canonical_spec, canonical_design, and canonical_plan resolve when present. If any pointer is stale, report it plainly. Recommend auto-frame for missing SPEC.md or auto-plan for missing PLAN.md.
Treat current.json as the only source for active change, stage, and canonical artifact pointers. Load artifacts in dependency order and stop at the current stage. Read references/artifact-order.md for the full stage table.
When stage is execute or verify and the project is a git repo, read the execution ledger before summarizing: git log --oneline -15 and git status --porcelain. Per-slice commits (slice N: ..., slice N gap-fix: ...) mark verified slices; match them against PLAN.md slice evidence. A dirty tree on top of the last slice commit is in-flight work for the next slice, not noise: name the touched files. When commits and PLAN.md evidence disagree, trust the commits and report the mismatch. Also run git worktree list: a stray worktree is the fingerprint of an interrupted parallel dispatch. Report it. Do not remove it.
If current.json contains engineering_review, read the ## Review: Engineering section from the canonical plan and include it in the resume summary.
Omit any line that would report nothing. A healthy resume is five lines, not ten fields of "none".
Always:
**Active change:** [name]
**Stage:** [frame|plan|execute|verify|verified]
**Artifacts loaded:** [list]
**What was done:** [1-2 sentences]
**What comes next:** [specific next action, or "change complete"]
Add only when it carries something:
**Blocked:** [what stopped, and on what]
**Execution ledger:** [last slice commit, in-flight files]
**Review verdicts:** [engineering: X]
**Missing state:** [stale pointer or absent artifact]
**Roadmap:** [N pending]
The goal is orientation, not transcription.
Use references/recovery-scenarios.md for the full routing table. The invariant: recommend the next lifecycle skill only when recovered state is incomplete or blocked.
Resume orients and stops. It never starts the work it just found.
After the recovery summary, end the turn with **Next:** <skill>, <reason> when the recovered state has incomplete or blocked work. For a verified change, report Change status: complete and print no Next: line. When ROADMAP.md has pending items, surface them as optional future work rather than an automatic auto-frame handoff.
.agent/.automaton/state/current.json is read-only for auto-resume. Stale pointers are reported, not silently repaired