Tmux Remote Worker Setup

Use this skill when the user points at an existing local tmux session and wants the matching remote worker configured for Codex.

What this skill does

• Reads memory/optimized-workflow.md and recent run logs before acting so the latest proven workflow is reused.
• Reads references/worker_setup_memory.md before acting so prior pitfalls and proven fixes are reused.
• Reads tmux pane metadata and scrollback to resolve the worker replica tied to the named session.
• Uses local /kubebrain/brainctl as the primary control path; the tmux session is only used to discover the worker.
• Builds a local offline payload for node, npm, npx, codex, rg, and local Codex skills, then installs them onto the remote worker.
• Copies ~/.codex/.env, ~/.codex/AGENTS.md, and ~/.codex/feishu_notify.py through the local payload so the worker gets the same Codex config without manual paste.
• Verifies the final remote environment and checks that python /home/i-huangsiming/work/tools/gpu_util.py is still running.

Important rules

• Use the exact tmux session name the user gave you. If they did not provide one, ask for it.
• Do not guess the replica. The helper script extracts the latest replica= / pod= / JOB_ID= evidence from tmux history. If that evidence is missing, stop and ask the user for a session with preserved scrollback or for the explicit replica name.
• Always use the absolute brainctl path /kubebrain/brainctl.
• Keep Python execution inside /mnt/step3-abla/siming/.venv/bin/python.
• Do not interrupt the tmux pane or kill gpu_util.py unless the user explicitly asks. This workflow configures the worker through local brainctl exec, so the foreground process in tmux can stay untouched.

Self-evolution loop

Before running the helper:

1. Read memory/optimized-workflow.md.
2. Read references/worker_setup_memory.md.
3. Read the newest 3 logs in memory/runs/ if they exist.
4. Reuse any existing workaround or validation step that matches the current failure mode or environment.

After each configure or repair run:

1. Inspect the current run artifacts in the workspace, especially:
   • .../session_context.json
   • .../access_check.txt
   • .../remote_setup.stdout.txt
   • .../remote_setup.stderr.txt
   • .../verification.txt
   • .../verification.stderr.txt
2. Decide whether the run produced a new durable lesson:
   • a non-obvious failure signature
   • a more reliable fallback
   • a hidden dependency
   • a better verification technique
3. Record the run with python3 scripts/log_run.py ....
4. Rebuild memory/optimized-workflow.md with python3 scripts/update_workflow.py.
5. If the run produced a new durable lesson, add a concise entry to references/worker_setup_memory.md.
6. If the run only repeated known behavior, keep references/worker_setup_memory.md unchanged.

When updating memory, prefer this shape:

• Signal: what was observed
• Lesson: the reusable decision or workaround
• Why it matters: how it improves future robustness

The goal is to make future worker setup more reliable, not to create a run log.

Default workflow

1. Run the helper wrapper from this skill directory:

scripts/setup_remote_worker_from_tmux.sh configure <tmux-session-name> \
  --workspace-dir workspace

2. Read these generated files before replying:

• memory/optimized-workflow.md
• references/worker_setup_memory.md
• workspace/.../session_context.json
• workspace/.../access_check.txt
• workspace/.../verification.txt
• workspace/.../summary.json

3. Report back:

• resolved tmux session -> replica mapping
• whether setup finished successfully
• versions found for codex, node, npm, rg, and nvitop
• confirmation that notify and trusted-project config were written
• confirmation that gpu_util.py is still running

Run log shape

Use scripts/log_run.py with fields such as:

• request
• status
• session_name
• replica
• issues
• lessons
• actions
• outputs
• notes

Keep logs concise and durable. Avoid secrets and transient noise.

If proxy/bootstrap fails

The helper script first runs networked commands with:

eval $(curl -s http://deploy.i.shaipower.com/httpproxy)

If that attempt fails, it retries once after:

unset https_proxy http_proxy all_proxy

If both attempts fail, surface the error instead of inventing a workaround.

Example prompts that should trigger this skill

• "帮我把 tmux 里的 gpu2 对应远端机器配成能跑 codex 的环境"
• "tmux session 名叫 gpu，帮我修一下对应 worker 的 codex / node / nvitop 配置"
• "我只知道 tmux session 是 gpu2，你直接把那台远端服务器环境配置好"