research-pipeline-runner
端到端运行本项目的Units+Checkpoints研究流程(survey/综述/review/调研/教程/系统综述/审稿),工作区+checkpoints。**触发**:运行pipeline、kickoff、继续执行、自动跑、写文章、survey/综述/review/调研/教程/系统综述/审稿。**使用时机**:用户希望终点跑流程(创建`workspaces/<name>/`、生成/执行) `UNITS.csv`、遇到 HUMAN 检查点暂停等待)。**如果**:用户明确要手工逐条执行(用 `unit-executor`),或者你不能自动推进到散文阶段。**网络**:取决于所选管道(arXiv/PDF/引文验证可能需要网络;支持离线导入)。**Guardrail**:必须尊重检查点(无批准不写散文);遇到 HUMAN 单元必须停止等待;禁止在repo root 创建工作空间工件。
📖 中文导读:端到端运行本项目的Units+Checkpoints研究流程(survey/综述/review/调研/教程/系统综述/审稿),工作区+checkpoints。**触发**:运行pipeline、kickoff、继续执行、自动跑、写文章、survey/综述/review/调研/教程/系统综述/审稿。**使用时机**:用户希望终点跑流程(创建`workspaces/<name>/`、生成/执行) `UNITS.csv`、遇到 HUMAN 检查点暂停等待)。**如果**:用户明确要手工逐条执行(用 `unit-executor`),或者你不能自动推进到散文阶段。**网络**:取决于所选管道(arXiv/PDF/引文验证可能需要网络;支持离线导入)。**Guardrail**:必须尊重检查点(无批准不写散文);遇到 HUMAN 单元必须停止等待;禁止在repo root 创建工作空间工件。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。
name: research-pipeline-runner
description: |
Run this repo’s Units+Checkpoints research pipelines end-to-end (survey/综述/review/调研/教程/系统综述/审稿), with workspaces + checkpoints.
Trigger: run pipeline, kickoff, 继续执行, 自动跑, 写一篇, survey/综述/review/调研/教程/系统综述/审稿.
Use when: 用户希望端到端跑流程(创建 workspaces/<name>/、生成/执行 UNITS.csv、遇到 HUMAN checkpoint 停下等待)。
Skip if: 用户明确要手工逐条执行(用 unit-executor),或你不应自动推进到 prose 阶段。
Network: depends on selected pipeline (arXiv/PDF/citation verification may need network; offline import supported where available).
Guardrail: 必须尊重 checkpoints(无 Approve 不写 prose);遇到 HUMAN 单元必须停下等待;禁止在 repo root 创建 workspace 工件。
Research Pipeline Runner
Goal: let a user trigger a full pipeline with one natural-language request, while keeping the run auditable (Units + artifacts + checkpoints).
This skill is coordination:
- semantic work is done by the relevant skills’
SKILL.md - scripts are deterministic helpers (scaffold/validate/compile), not the author
Inputs
- User goal (one sentence is enough), e.g.:
- “给我写一个 agent 的 latex-survey”
- Optional:
- explicit pipeline path (e.g.,
pipelines/arxiv-survey-latex.pipeline.md) - constraints (time window, language: EN/中文, evidence_mode: abstract/fulltext)
- explicit pipeline path (e.g.,
Outputs
- A workspace under
workspaces/<name>/containing:STATUS.md,GOAL.md,PIPELINE.lock.md,UNITS.csv,CHECKPOINTS.md,DECISIONS.md- pipeline-specific artifacts (papers/outline/sections/output/latex)
Non-negotiables
- Use
UNITS.csvas the execution contract; one unit at a time. - Respect checkpoints (
CHECKPOINTS.md): no long prose until required approvals are recorded inDECISIONS.md(survey default:C2). - Stop at HUMAN checkpoints and wait for explicit sign-off.
- Never create workspace artifacts in the repo root; always use
workspaces/<name>/.
Decision tree: pick a pipeline
User goal → choose:
- Survey/综述/调研 + Markdown draft →
pipelines/arxiv-survey.pipeline.md - Survey/综述/调研 + PDF output →
pipelines/arxiv-survey-latex.pipeline.md - Snapshot/速览 →
pipelines/lit-snapshot.pipeline.md - Tutorial/教程 →
pipelines/tutorial.pipeline.md - Systematic review/系统综述 →
pipelines/systematic-review.pipeline.md - Peer review/审稿 →
pipelines/peer-review.pipeline.md
Recommended run loop (skills-first)
- Initialize workspace (C0):
- create
workspaces/<name>/ - write
GOAL.md, lock pipeline (PIPELINE.lock.md), seedqueries.md
- Execute units sequentially:
- follow each unit’s
SKILL.mdto produce the declared outputs - only mark
DONEwhen acceptance criteria are satisfied and outputs exist
- Stop at HUMAN checkpoints:
- default survey checkpoint is
C2(scope + outline) - write a concise approval request in
DECISIONS.mdand wait
- Writing-stage self-loop (when drafts look thin/template-y):
- prefer local fixes over rewriting everything:
writer-context-pack(C4→C5 bridge) makes packs debuggablesubsection-writerwrites per-file unitswriter-selfloopfixes only failingsections/*.mddraft-polisherremoves generator voice without changing citation keys
- Finish:
- merge → audit → (optional) LaTeX scaffold/compile
Practical commands (optional helpers)
- Kickoff + run (recommended):
python scripts/pipeline.py kickoff --topic "<topic>" --pipeline <pipeline-name> --run --strict - Resume:
python scripts/pipeline.py run --workspace <ws> --strict - Approve checkpoint:
python scripts/pipeline.py approve --workspace <ws> --checkpoint C2 - Mark refined unit:
python scripts/pipeline.py mark --workspace <ws> --unit-id <U###> --status DONE --note "LLM refined"
Handling common blocks
- HUMAN approval required: summarize produced artifacts, ask for approval, then record it and resume.
- Quality gate blocked (
output/QUALITY_GATE.mdexists): treat current outputs as scaffolding; refine per the unit’sSKILL.md; markDONE; resume. - No network: use offline imports (
papers/imports/orarxiv-search --input). - Weak coverage: broaden queries or reduce/merge subsections (
outline-budgeter) before writing.
Quality checklist
-
UNITS.csvstatuses reflect actual outputs (noDONEwithout outputs). - No prose is written unless
DECISIONS.mdexplicitly approves it. - The run stops at HUMAN checkpoints with clear next questions.
- In strict mode, scaffold/stub outputs do not get marked
DONEwithout refinement.