菜单
Reference contract for the canonical Arnold native-first pipeline template scaffold.
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
Run the canonical Megaplan planning pipeline from its Arnold plugin home.
Launch one durable, mutation-authorized meta-fixer to repair a failed automated fixer and its missed backstop, retrigger ordinary repair, and prove the identified epic or session advances. Use when a watchdog, repair loop, meta-repair loop, or progress auditor ran or should have run but the real chain stayed stuck, repeated stale evidence, claimed false success, or failed to self-correct. Invoke with `--target "EPIC_OR_SESSION_TEXT"`.
Launch an external model as a subagent for a second opinion, adversarial review, or delegated work. Default pathway is an agentic DeepSeek / Kimi / Zhipu GLM hermes subagent (file/web/terminal tools, one process or fanned out N-wide); also Codex (GPT-5.5) and Claude via the Agent tool. Use for independent root-cause analysis, cross-checking your reasoning, judge/jury panels, or handing implementation to a different model.
Scaffold a new Arnold native-first pipeline package from the _template skeleton.
Adversarial review of prose drafts by N reviewers, then revise. Not for code.
Survey every place loose work hides in a git repo — starting with the current checkout's own uncommitted/untracked/unpushed work, then local branches, all worktrees (`.megaplan-worktrees`, `.megaplan/bakeoffs/*`, agent-tool worktrees), stashes, detached HEADs, interrupted rebases/merges, submodules, remote branches on origin and other remotes, fork PR refs, sibling repo variants, other clones of this repo elsewhere on disk, the megaplan-cloud machine (Hetzner box) and every workspace on it, and GitHub Codespaces — classify each as land-on-main / delete / parked with reasoning, and act only on explicit per-item approval. Use when the user says "clean up loose branches", "prune branches", "what branches can I delete", "clean up worktrees", "review my stashes", "what's lying around in this repo", or asks for branch / worktree / stash housekeeping.
| name | pipeline-template-reference |
| description | Reference contract for the canonical Arnold native-first pipeline template scaffold. |
The canonical scaffold is arnold_pipelines/_template/. It shows the minimum
required contract for a native-first pipeline:
build_pipeline() returns arnold.pipeline.types.Pipeline with a
non-null native_program.driver=("native", "<kind>") and
supported_modes containing "native".pipelines.py using @pipeline, @phase,
@decision, parallel, compile_pipeline, and project_graph.For native-first authoring, organise components as package-local exports:
pipelines.py owns the native declaration topology. Supporting modules
(steps.py, prompts.py, policies.py, schemas.py, subflows.py)
may be imported by the native declaration, but the declaration itself
is the single source of topology truth.
The build_pipeline() entrypoint compiles the native program and projects
it into a Pipeline shell. The shell exists for discovery and validation.
The runtime executes the native program directly.
The native_program field is a dispatch substrate, not a final
compositional surface. It proves the package is executable by the native
runtime, but it does not lock in panel synthesis, join delegation,
parallel merge strategy, subpipeline ownership, or Capsule projection.
Those concerns belong to later Megaplan layers above the dispatch
boundary.
Do not use these patterns in new packages built from the template:
arnold.workflow.Pipeline, Step, Route._legacy.py, graph fallback builders, compatibility namespaces, shim
packages, or temporary wrapper modules.--driver graph scaffolding or manual graph construction.WorkflowManifest, NativeProgram builder objects, or
_forward_m2_m3 graph objects in build_pipeline().Derived manifests, catalogs, and projection artifacts are compiler output, not package source.