human-review-skill

name	human-review-skill
description	This skill should be used when the user wants to set up checkpoints, record approval decisions, gate execution pending human confirmation, or manage human-in-the-loop review points in a computational math workflow.

Human Review Skill

Manages lightweight human approval gates for the compact workflow.

For end-to-end computational math research-code reproduction workflows, this Skill should be selected by computational_math_reproduction_workflow_skill only when durable approval logs are needed.

Lightweight Approval Types

Four approval types, all conversation-led:

1. plan approval — before executing external code (plan.md)
2. repair approval — before modifying source, dependencies, entrypoint, or data (repair_plan.md)
3. tuning approval — before starting or expanding tuning budget (tuning/tuning_plan.md)
4. final acknowledgment — before accepting conclusions

Approval Aliases

User replies approve, ok, yes, 可以, 好的, 继续, 同意, 批准, 就这个吧, 选这个, 没问题 are all treated as approval.

Revise: revise, 修改, 改一下 Reject: reject, 拒绝, 不行 Skip: skip, 跳过, 跳过这个

Workflow

1. Present the compact artifact (plan.md, repair_plan.md, tuning/tuning_plan.md) to the user.
2. Ask for exactly one decision: approve, revise, reject, or skip.
3. Record the decision in conversation.
4. Only write approvals/approval_log.jsonl when the operation is high-risk (source changes, dependency installs, budget increases).
5. Do not write numbered checkpoint files (01_task_understanding.md, etc.) by default.

Scripts

approval_gate

Input: --run <path>, --checkpoint <name>, --risk-level <low|high> Exits 0 if allowed, exits 2 if blocked.

approval_logger

Records an approval decision to approvals/approval_log.jsonl. Only call when explicitly logging a high-risk approval.

python -m skills.human_review_skill.scripts.approval_logger --run /path/to/output --checkpoint plan --decision approve --reason "low-risk demo"

Agent Conversation Semantics

• Ask for exactly one decision: approve, revise, reject, or skip.
• If the user approves, continue to the next stage.
• If the user asks to revise, keep the same artifact pending and address the requested changes.
• If the user rejects, record the rejection and stop or reroute.
• If the user skips, continue only if the stage is optional.

Do not write numbered checkpoint files by default. Process reasoning stays in conversation.