auto-research-loop

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Auto Research Loop — Autonomous Iteration Engine

自动研究循环——自主迭代引擎

Combines Karpathy's autoresearch with Ralph Loop infrastructure. Modify, Verify, Keep/Discard, Repeat.

结合Karpathy's autoresearch与Ralph Loop基础设施。流程为：修改、验证、保留/舍弃、重复。

Invocation

调用方式

/auto-research-loop [PROMPT] [FLAGS]
— Run the loop:

"${CLAUDE_PLUGIN_ROOT}/scripts/setup-auto-research-loop.sh" $ARGUMENTS

Then follow the injected instructions. The stop hook auto-installs and intercepts exit to re-feed the prompt.

/auto-research-loop-plan
— Interactive planning wizard:

Don't run the setup script. Instead, read

${CLAUDE_PLUGIN_ROOT}/skills/auto-research-loop/references/plan-workflow.md

and walk the user through 7 phases to build a validated configuration:

Capture goal
Analyze codebase context
Define scope (which files to modify)
Define metric (must be mechanical — a command that outputs a number)
Define direction (higher or lower is better)
Define verify command (dry-run it to confirm it works)
Confirm and launch — output a ready-to-paste
```
/auto-research-loop
```
command

Use this wizard when the user says "help me set up", "plan a run", "what should my metric be", or invokes the plan command.

/auto-research-loop [PROMPT] [FLAGS]
—— 运行循环：

"${CLAUDE_PLUGIN_ROOT}/scripts/setup-auto-research-loop.sh" $ARGUMENTS

随后按照注入的说明操作。停止钩子会自动安装，并拦截退出操作以重新传入提示词。

/auto-research-loop-plan
—— 交互式规划向导：

无需运行设置脚本。请阅读

${CLAUDE_PLUGIN_ROOT}/skills/auto-research-loop/references/plan-workflow.md

，并引导用户完成7个阶段以构建经过验证的配置：

捕获目标
分析代码库上下文
定义范围（要修改的文件）
定义指标（必须可量化——即能输出数字的命令）
定义方向（数值越高越好还是越低越好）
定义验证命令（先试运行以确认可用）
确认并启动——输出可直接粘贴的
```
/auto-research-loop
```
命令

当用户说“帮我设置”、“规划一次运行”、“我的指标应该是什么”，或是调用规划命令时，请使用此向导。

Two Modes

两种模式

	Metric Mode	Task Mode
When	`--metric` + `--verify` provided	No metric provided
Decision	Metric improved? Keep. Worse? `git revert`	Accumulate toward completion
Exit	Max iterations or manual	Completion promise or max iterations
Journal	`autoresearch-results.tsv`	`IMPLEMENTATION_PLAN.md`

	指标模式	任务模式
适用场景	提供 `--metric` + `--verify` 参数时	未提供指标时
决策逻辑	指标是否提升？是则保留，否则 `git revert`	逐步推进直至任务完成
退出条件	达到最大迭代次数或手动停止	完成任务承诺或达到最大迭代次数
日志记录	`autoresearch-results.tsv`	`IMPLEMENTATION_PLAN.md`

The Loop

循环流程

LOOP:
  0. Scratchpad: READ .claude/auto-research-loop-scratchpad.md
  1. Review: State + git log + results/plan
  2. Ideate: Fix crashes > exploit wins > explore > simplify > radical
  3. Modify: ONE focused change
  4. Commit: git commit BEFORE verification
  5. Verify: Metric command (metric) or gate commands (task)
  6. Decide:
     Metric: IMPROVED -> keep. WORSE -> git reset --hard HEAD~1
     Task: Gates pass + promise true -> exit. Else -> continue
  7. Log: Results TSV (metric) or update plan (task)
  8. Scratchpad: UPDATE before exit
  9. Repeat

Read

${CLAUDE_PLUGIN_ROOT}/skills/auto-research-loop/references/autonomous-loop-protocol.md

for full protocol.

LOOP:
  0. 草稿本：读取.claude/auto-research-loop-scratchpad.md
  1. 回顾：状态 + git日志 + 结果/规划
  2. 构思：修复崩溃 > 利用现有成果 > 探索新方向 > 简化 > 激进改进
  3. 修改：每次仅做一个聚焦的变更
  4. 提交：验证前先执行git commit
  5. 验证：执行指标命令（指标模式）或门控命令（任务模式）
  6. 决策：
     指标模式：指标提升 -> 保留。指标下降 -> git reset --hard HEAD~1
     任务模式：通过门控且完成承诺 -> 退出。否则 -> 继续
  7. 记录：写入结果TSV文件（指标模式）或更新规划（任务模式）
  8. 草稿本：退出前更新内容
  9. 重复

完整协议请阅读

${CLAUDE_PLUGIN_ROOT}/skills/auto-research-loop/references/autonomous-loop-protocol.md

。

Critical Rules

核心规则

NEVER STOP — loop until interrupted, max iterations, or promise met
One change per iteration — atomic, attributable
Mechanical verification only — no subjective judgment
Simplicity wins — equal results + less code = KEEP
Git is memory — commit before verify, revert on failure
Scratchpad is mandatory — read at start, update before exit

永不停止——循环直至被中断、达到最大迭代次数或完成任务承诺
每次迭代一个变更——原子性、可追溯
仅机械验证——无主观判断
简洁优先——结果相同+代码更少=保留
Git作为记忆——验证前提交，失败时回滚
必须使用草稿本——开始时读取，退出前更新

Domain Adaptation

领域适配

Domain	Metric	Direction	Verify	Scope
Test coverage	%	higher	`pytest --cov \| grep TOTAL`	`src/*/.py`
Bundle size	KB	lower	`npm run build \| grep size`	`src/*/.ts`
ML training	val_bpb	lower	`uv run train.py \| grep val_bpb`	`train.py`
Performance	ms	lower	`npm run bench \| grep p95`	target files

To manually stop:

rm .claude/auto-research-loop.local.md

领域	指标	优化方向	验证方式	范围
测试覆盖率	%	越高越好	`pytest --cov \| grep TOTAL`	`src/*/.py`
包体积	KB	越小越好	`npm run build \| grep size`	`src/*/.ts`
ML训练	val_bpb	越小越好	`uv run train.py \| grep val_bpb`	`train.py`
性能	ms	越小越好	`npm run bench \| grep p95`	目标文件

手动停止方式：

rm .claude/auto-research-loop.local.md