Remote Experiment Execution

You can use the compute-helper CLI to autonomously execute code and commands on a remote server.

The compute-helper path and server information are given in the <compute_node> block of the system prompt. If not, look in sidecar/bin/compute-helper.mjs .

Command Quick Reference

Command	Purpose	When to Use
`node <helper> ssh "<cmd>"`	Execute commands remotely only	When checking the environment, viewing files, and no code changes are involved
`node <helper> sync up --cwd <root>`	Sync local code to the server	Manual sync (usually not needed,`run` syncs automatically)
`node <helper> run "<cmd>" --cwd <root>`	Sync code + execute remotely	When code changes need to be run on the server
`node <helper> sync down --cwd <root> --files "logs/ results/"`	Pull files back from the server	When you need to view result files
`node <helper> info`	View node configuration	When confirming connection information

Core Behavior Rules

Rule 1: Act proactively, don't wait to be prompted

After modifying code, you must immediately run verify by remote execution; do not stop and wait for user confirmation after making changes.

❌ 错误: "我已经修改了代码，你可以运行看看。"
✅ 正确: 修改代码 → 立即 run → 分析输出 → 汇报结果

Rule 2: After failure, analyze and fix autonomously

When remote execution reports an error,immediately analyze the error output → modify the code → run again, forming a self-repair loop. Only ask the user for help when encountering a problem you cannot judge.

❌ 错误: "执行出错了，错误信息如下：..."（等用户处理）
✅ 正确: 分析错误 → 修改代码 → 再次 run → 如果还是失败 → 换策略再试

Rule 3: Probe the environment first

When operating on a remote server for the first time, first use ssh to check the environment:

which python / python3 --version — whether Python is available
nvidia-smi — GPU status (if needed)
pip list | grep <package> — whether dependencies are installed
ls <workdir> — working directory status

Rule 4: Distinguish `run` vs `ssh`

local code was changed → use run(automatic sync + execution)
only want to execute a command on the server(view processes, install packages, check logs) → use ssh
only need to sync code without executing → use sync up

Rule 5: Single-variable changes

Modify only one variable or one aspect per iteration to make it easier to locate problems. When multiple places are changed at once and an error occurs, it is impossible to tell which change caused it.

Iteration Workflow

When the user asks to run/test/experiment on the server:

1. ssh 检查环境 ──→ 缺依赖？安装
                        │
2. 修改本地代码 ◄──────┘
        │
3. run 远程执行 ──→ 成功？
        │              ├─ 是 → 汇报结果，问下一步
        │              └─ 否 → 分析错误 → 回到步骤 2
        │
4. (可选) sync down 拉回结果文件

Progress Reporting

After each execution, briefly explain:

what changes were made and why
execution result (success/failure + key output)
next-step plan

📊 执行结果
━━━━━━━━━━
改动: 将 batch_size 从 32 改为 64
命令: python train.py --batch_size 64
结果: ✅ 训练完成，loss 从 0.45 降到 0.32
下一步: 尝试增大学习率到 3e-4