research-pipeline-runner（研究管道运行者）

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)

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.csv as the execution contract; one unit at a time.
• Respect checkpoints (CHECKPOINTS.md): no long prose until required approvals are recorded in DECISIONS.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
• Research brief / rapid review / 速览 → pipelines/research-brief.pipeline.md
• Paper review / paper critique / 审稿 → pipelines/paper-review.pipeline.md
• Evidence review / systematic review / 系统综述 → pipelines/evidence-review.pipeline.md
• Idea finding / 选题 / 点子 / 找方向 → pipelines/idea-brainstorm.pipeline.md
• Tutorial/教程 → pipelines/source-tutorial.pipeline.md

Recommended run loop (skills-first)

1. Initialize workspace (C0):

• create workspaces/<name>/
• write GOAL.md, lock pipeline (PIPELINE.lock.md), seed queries.md

1. Execute units sequentially:

• follow each unit’s SKILL.md to produce the declared outputs
• only mark DONE when acceptance criteria are satisfied and outputs exist

1. Stop at HUMAN checkpoints:

• default survey checkpoint is C2 (scope + outline)
• write a concise approval request in DECISIONS.md and wait

1. Writing-stage self-loop (when drafts look thin/template-y):

• prefer local fixes over rewriting everything:

- writer-context-pack (C4→C5 bridge) makes packs debuggable - subsection-writer writes per-file units - writer-selfloop fixes only failing sections/*.md - paragraph-curator / style-harmonizer / opener-variator converge structure and de-template the prose - evaluation-anchor-checker is the late section-level numeric hygiene sweep before merge - draft-polisher removes generator voice without changing citation keys

Strict-mode behavior (by design)

In --strict runs, several semantic C3/C4 artifacts are treated as *scaffolds* until explicitly marked refined. This is intentional: it prevents bootstrap JSONL from silently passing into C5 writing (a major source of hollow/templated prose).

Create these markers only after you have manually refined/spot-checked the artifacts:

• outline/subsection_briefs.refined.ok
• outline/chapter_briefs.refined.ok
• outline/evidence_bindings.refined.ok
• outline/evidence_drafts.refined.ok
• outline/anchor_sheet.refined.ok
• outline/writer_context_packs.refined.ok

The runner may BLOCK even if the JSONL exists; add the marker after refinement, then rerun/resume the unit.

1. Finish:

• merge → audit → (optional) LaTeX scaffold/compile

Optional CLI helpers (debug only)

• Kickoff + run (optional; convenient, not required): 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.md exists): treat current outputs as scaffolding; refine per the unit’s SKILL.md; mark DONE; resume.
• No network: use offline imports (papers/imports/ or arxiv-search --input).
• Weak coverage: broaden queries or reduce/merge subsections (outline-budgeter) before writing.

Quality checklist

• UNITS.csv statuses reflect actual outputs (no DONE without outputs).
• No prose is written unless DECISIONS.md explicitly approves it.
• The run stops at HUMAN checkpoints with clear next questions.
• In strict mode, scaffold/stub outputs do not get marked DONE without refinement.