AgentSkillsCN

unit-executor

从 `UNITS.csv` 中精确执行一条可运行的单元（优先执行依赖项已全部完成的 TODO 任务），随后更新单元状态与相关产出文件。 **触发条件**：单元执行师、逐单元执行、下一单元、逐步流程、逐条执行、UNITS.csv。 **适用场景**：当您需要严格遵循“一次只执行一个单元”（可审计、可中断），并遵守检查点与人工审核的逻辑时。 **跳过条件**：若您希望全流程自动化运行（使用 `research-pipeline-runner`）或工作空间尚未创建。 **网络需求**：无。 **约束规则**：每次仅执行一个单元；只有在满足验收标准且产出文件已成功生成时，方可标记为“已完成”；每遇人工检查点，必须暂停等待。

SKILL.md
--- frontmatter
name: unit-executor
description: |
  Execute exactly one runnable unit from `UNITS.csv` (first TODO whose dependencies are DONE), then update unit status and artifacts.
  **Trigger**: unit executor, run one unit, next unit, step-by-step pipeline, 逐条执行, UNITS.csv.
  **Use when**: 需要严格“一次只做一个 unit”（可审计、可中断），并遵守 checkpoints/HUMAN 阻塞逻辑。
  **Skip if**: 要端到端自动跑（用 `research-pipeline-runner`）或 workspace 不存在。
  **Network**: none.
  **Guardrail**: 只执行一个 unit；满足验收且输出存在才可标 `DONE`；遇到 HUMAN checkpoint 必须停下。

Skill: unit-executor

Goal

  • •Execute one unit end-to-end and leave the workspace in a consistent state.

Inputs

  • •UNITS.csv
  • •Unit inputs listed in the row (files)

Outputs

  • •Unit outputs listed in the row (files)
  • •Updated UNITS.csv status (TODO → DOING → DONE/BLOCKED)
  • •Optional: STATUS.md updated

Procedure (MUST FOLLOW)

  1. •Load UNITS.csv and find the first unit with:
    • •status=TODO
    • •all depends_on units are DONE
  2. •Set its status to DOING and persist UNITS.csv.
  3. •Run the referenced skill (by following that skill’s SKILL.md).
  4. •Check the unit’s acceptance against produced artifacts.
  5. •If acceptance passes, set status to DONE; otherwise set to BLOCKED with a short note in STATUS.md.
  6. •Stop after one unit (do not start the next unit automatically).

Acceptance criteria (MUST CHECK)

  • • Exactly one unit changes from TODO to DONE/BLOCKED (via DOING).
  • • Output files exist (or acceptance explicitly allows otherwise).

Side effects

  • •Allowed: edit workspace artifacts (UNITS.csv, STATUS.md, unit outputs).
  • •Not allowed: modify .codex/skills/ content.

Script

Quick Start

  • •python .codex/skills/unit-executor/scripts/run.py --help
  • •python .codex/skills/unit-executor/scripts/run.py --workspace <workspace_dir>

All Options

  • •--strict: enable quality gate (blocks on scaffolds; writes output/QUALITY_GATE.md)

Examples

  • •Run exactly one unit (strict):
    • •python .codex/skills/unit-executor/scripts/run.py --workspace <ws> --strict
  • •Equivalent repo wrapper:
    • •python scripts/pipeline.py run-one --workspace <ws> --strict

Notes

  • •Returns 0 on DONE/IDLE, 2 on BLOCKED/ERROR (useful for automation).

Troubleshooting

Issue: no runnable unit is found

Fix:

  • •Check UNITS.csv for unmet dependencies, missing outputs, or owner=HUMAN units waiting on approvals in DECISIONS.md.

Issue: a unit is marked DONE but outputs are missing

Fix:

  • •Fix the status to TODO/BLOCKED and re-run; only mark DONE when acceptance criteria and outputs are satisfied.