Skip to content
Automation
Command

/rebuttal-start

启动 Rebuttal 对抗迭代循环

From plugin
autoresearch-sibylsystem
27115 skills3 agents15 commands
Install
$ npx -y skills add Sibyl-Research-Team/AutoResearch-SibylSystem --agent claude-code

How it fires

How this command gets triggered: by you, by Claude, or both.

  • Fires itselfClaude auto-loads it when your prompt matches the work.
  • You can call itInvoke it directly when you want it.
  • Slash command/rebuttal-start

Context preview

What this command does when you run it.

启动 Rebuttal 对抗迭代循环

Command definition

rebuttal-start.md
description: "启动 Rebuttal 对抗迭代循环"
argument-hint: "<workspace_path>"

/sibyl-research:rebuttal-start

启动 Rebuttal 对抗迭代循环。需要已初始化的工作区(通过 `/sibyl-research:rebuttal-init` 创建)。

工作目录: `$SIBYL_ROOT`

Python 环境

所有 python3 调用必须使用 `.venv/bin/python3`,不要使用裸 `python3`。

参数: `$ARGUMENTS`(workspace 路径)

步骤

0. **打印启动横幅**:

cd $SIBYL_ROOT && .venv/bin/python3 -c "from sibyl.orchestrate import cli_rebuttal_status; cli_rebuttal_status('$ARGUMENTS')"

解析输出,打印横幅:

╔═════════════════════════════════════════════════════════════════╗
║     SIBYL REBUTTAL SYSTEM  ·  Adversarial Rebuttal Engine       ║
╚═════════════════════════════════════════════════════════════════╝

  工作区:<workspace_path>
  阶段:<current_stage>
  轮次:#<round> / <max_rounds>
  Reviewers:<reviewer_ids>
  字数限制:<word_limit or "无限制">
  Codex:<enabled/disabled>

  正在启动对抗迭代循环 →

1. **创建 Task 依赖链**追踪进度:

  • 为每个剩余阶段创建 Task(从当前 stage 到 done)
  • 用 `addBlockedBy` 串联依赖

2. **进入编排循环**:

LOOP:
  1. 获取下一步操作:
     cd $SIBYL_ROOT && .venv/bin/python3 -c "from sibyl.orchestrate import cli_rebuttal_next; cli_rebuttal_next('WORKSPACE_PATH')"

  2. 导出语言环境变量:
     export SIBYL_LANGUAGE=<action.language>

  3. 按 action_type 分发执行:

     skill:
       调用 action.skills[0].name 技能,args 为 action.skills[0].args

     skills_parallel:
       并行启动所有 action.skills 中的技能,等待全部完成

     team:
       a. 使用 TeamCreate 创建团队 action.team.team_name
       b. 为每个 teammate 创建 TaskCreate,指定 owner 为 teammate.name
       c. 为每个 teammate 启动 Agent,调用其 skill,传入 args
       d. 等待所有 teammates 完成
       e. 按顺序执行 action.team.post_steps(每个 post_step 是一个 skill)

     bash:
       执行 action.bash_command

     done:
       打印完成横幅,输出最终 rebuttal 路径,退出循环

  4. 记录完成:
     cd $SIBYL_ROOT && .venv/bin/python3 -c "from sibyl.orchestrate import cli_rebuttal_record; cli_rebuttal_record('WORKSPACE_PATH', 'STAGE')"

  5. 更新对应 Task 状态为 completed

  6. 检查返回结果:
     - 如果 new_stage == "rebuttal_draft" 且 round > 1:打印迭代状态面板
       (当前轮次、分数变化、剩余 concern 数)
     - 如果 new_stage == "done":退出循环

  7. 回到 LOOP 开头

3. **完成横幅**:

╔═════════════════════════════════════════════════════════════════╗
║     REBUTTAL COMPLETE                                            ║
╚═════════════════════════════════════════════════════════════════╝

  最终分数:<avg_score>
  迭代轮次:<total_rounds>
  输出文件:
    - WORKSPACE/output/rebuttal_letter.md(完整 rebuttal 信)
    - WORKSPACE/output/per_reviewer/(逐 reviewer 回复)
    - WORKSPACE/output/experiment_plan.md(补充实验计划)
    - WORKSPACE/output/score_trajectory.json(分数轨迹)

注意事项

  • **不涉及 GPU/实验执行**:rebuttal 系统不自动执行实验,只生成实验计划
  • **不需要 Sentinel**:rebuttal 循环较短(通常 2-3 轮 × 几分钟),不需要看门狗
  • **字数限制**:team prompt 和 synthesizer 会自动感知并强制执行配置的字数限制
  • **每轮迭代**包含:Rebuttal Team 起草(8人) → Simulated Reviewers 攻击(N人) → [Codex 审查] → 评分判断
  • **停止条件**:`平均分 ≥ 阈值` 或 `轮次 ≥ 上限`,两者取先
Read more
Ships withautoresearch-sibylsystem

Fully Autonomous AI Research System with Self-Evolution, built natively on Claude Code

Get the whole plugin