メニュー
Resume the active manifest-based Hypo-Workflow Delivery from Runtime, Continuation, and the latest valid Recovery Pack. Use for /hw:resume or when tracked Goal/Cycle work must continue after restart or compaction.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Route project work through the manifest-based Hypo-Workflow protocol. Use for /hw:* commands, repository initialization, Goal or Cycle delivery, structured experiments, focused maintenance records, final acceptance, and restart-safe Resume.
Define and run a multi-stage Hypo-Workflow delivery with ordered Milestones and one final manual acceptance. Use for /hw:cycle or work whose stages have real dependency order.
Maintain a non-linear, reproducible experiment lane with project knowledge, run identity, parameter scans, long-run supervision, scientific review, immutable Git records, and instant materialized status. Use for /hw:experiment, experiment setup or reruns, NeRF/AceSim-style comparisons, baseline changes, parameter sweeps, result triage, and questions such as "现在实验怎么样".
Define and run one bounded Hypo-Workflow delivery outcome without fake Milestones. Use for /hw:goal or when the user wants one explicit acceptance target that a capable model can execute after approval.
Inspect the repository and recommend one available Hypo-Workflow path. Use for /hw:guide or when the user is unsure whether to initialize, use Goal, use Cycle, manage experiments, maintain, resume, accept, or reject.
Build an evidence-driven Goal Design or Cycle Plan. Use for /hw:plan and planning requests that need questions, architecture choices, ordered decomposition, or durable research before execution.
| name | hypo-workflow-resume |
| description | Resume the active manifest-based Hypo-Workflow Delivery from Runtime, Continuation, and the latest valid Recovery Pack. Use for /hw:resume or when tracked Goal/Cycle work must continue after restart or compaction. |
用户可见内容遵循项目输出语言；缺失时跟随当前对话语言。内部 schema key 保持英文。
.pipeline/runtime/active.yaml and select only active.delivery.pack_status: missing, pack_ref: null, and degraded context. Do not block first-use projects merely because no compaction/checkpoint Pack has been sealed.current; otherwise report stale and replay bounded Journal context without replacing lifecycle state.next_action. Do not replay completed transitions or accept an arbitrary inactive object.If Runtime or Continuation contains worker_routing, validate and reuse that exact semantic decision, Task Assessment, reason codes, failure count, and policy version. Recovery Capsule Worker projections may restore the same bounded fields, but they never override newer Continuation. Do not reclassify on Resume, and do not let routing metadata change role identity, evidence, acceptance, or user authority.
Use createDeliveryStore({ clock }).resume(root, {}). Never shell-execute a saved command and never read legacy state as fallback in a current workspace.
Import createDeliveryStore from the installed Skill bundle's core/src/index.js; pass the target workspace separately as root and use a zero-argument timezone-bearing Clock. Resume itself is read-only and takes no transaction { id }.
The target workspace root and installed Skill bundle root are separate. repoRoot identifies the project whose .pipeline/ is resumed; backend routing uses explicit skillRoot or the installed bundle default, never the target repository as a Skill directory.
For worker-separated work, preserve requested -> started -> completed|failed|blocked -> closed|close_failed evidence and distinct identities. When /hw:resume stops, blocks, aborts, or completes, close/release workers it opened or record close_failed; incomplete lifecycle evidence blocks worker-separated completion.
When a resumed Worker needs a different semantic handoff, start it with no-history or bounded-history context. A full-history fork inherits its parent execution context and cannot be silently switched.