Gonglitian

大改动 (>3 文件) 之前先画依赖图 + 改动计划，确认后再动手。**所有外部 ref（line# / API / partition / env var）必须配 _verify 守门函数 runtime fail-loud**，不能盲信 WebFetch 结果。Use when user says "arch plan", "改动计划", "画依赖图", "重构前", "big change", "多文件改", "先规划", "monkey patch", "插入点".

Driver Claude（通常通过 /remote-drive 控制 sub-agent CC）跑实战发现新 bug / 配置问题 / 架构选择需重新拍板时，把 findings 标准化写到 .driver_findings_<round>_<context>.md，让 sub-agent 读了去 patch。比 send-keys 长 message 可靠 100 倍。Use when user says "driver findings", "findings file", "round N 反馈", "远程发现", "hpcc 验证发现", "driver patch", "sub-agent 修 bug".

用一个 Claude (driver) 通过 tmux 远程控制另一个 Claude Code session (sub-agent CC) 走完 pipeline。把 sub-agent 当 worker 用：driver brief 任务 → send-keys 输入 → capture-pane 监听 → 处理 AskUserQuestion → 等下次 idle。Use when user says "remote drive", "驱动另一个 cc", "sub-agent cc", "控制别的 claude", "tmux drive", "我已经开了 cc 你来控制".

起任何实验/sprint 前必须填 Karpathy 5-tuple: Goal / Scope / Metric / Verify / Guard。不填拒绝起训练。Stage 4 → 5 之间 RESOURCE gate 的一部分。用户说 'contract' / '5-tuple' / '签合同' / 'sprint plan' 时调用。

把本地 project 推到远程 compute host（hpcc / bcc / tasl-labserver）+ 远端 bootstrap auto-production framework + 验证 stubs import 通。给 driver 用，跟 /cross-host-sync (run sync) 互补——后者管 run state，本 skill 管 code state。Use when user says "sync to remote", "部署到 hpcc", "推到集群", "rsync project", "远程跑", "hpcc 上验证", "deploy 上去".

写完代码先用 AST parser 做静态分析（import 完整性、签名匹配、未定义符号），过了才执行。防 "跑了 30 秒才发现 ImportError"。Use when user says "ast validate", "静态分析", "先检查", "语法检查", "import 验证", "static check".

标准化失败 retrain 循环：scancel → audit (sub-agent) → patch → 重起。这次 hf-jax 做了 v3/v4/v5/v6 4 次的模板化。Use when user says "audit retrain", "重训", "训挂了", "scancel audit patch", "diagnose retrain", "为什么 SR 低".

决策为 REFINE 或 PIVOT 时，自动 snapshot stage-N_v{K}/ 目录（保留旧版本），新 run 用 3-part prompt: CONTINUATION (上次进度) + PRIOR PLAN (上次方案) + DELTA (这次改动)。Use when user says "auto version", "snapshot", "备份旧版本", "v3 v4 v5", "version run", "保留旧 run".

巡检本期 SSH 可连的 5 台远程服务器（HPCC / BCC / TASL-LabServer / TASL-7 / Brev）的 GPU 情况——每卡显存与利用率、Slurm 集群的账号配额/排队作业/全节点空闲 GPU，并算出"我实际还能新申请多少张"。当用户说"检查 GPU""看下 GPU 显存""GPU 情况""哪台服务器有空卡""查一下算力""where can I run""check GPU status""GPU 够不够""有没有空闲显卡"等，或在准备起训练/挑服务器时触发。

统一绘图 skill——文生图 / 图生图 / 多图合成 / 编辑 / 论文与科普配图。两个后端由用户指定：nano(默认，Google Gemini "Nano Banana"，快 3-10s、便宜、原生中英文文字渲染、可 1K/2K/4K，含 lite/nb2/pro 三档) 或 gpt(速创API GPT-Image-2，真实感/精细文字更强，异步稍慢 60-90s)。四个专用风格预设也由用户按名指定，两后端通用：UniVLA(pastel macaron 论文方法图)、Physical Intelligence(π/pi-zero retro-computing monospace 图)、Fast-WAM(CVPR/ICLR 矢量图+attention-mask 矩阵)、蓝皮书 BlueBook(蓝白科技科普/咨询长图)。Use this skill for ANY image generation or editing request. 触发词：「画一张」「生成一张图」「出张图」「文生图」「图生图」「改这张图」「编辑图片」「把这张图改成…」「合成/融合这几张图」「做张海报」「配图」「方法图」「架构图」「用 nano/gemini 画」「用 GPT 画」「generate/edit/compose an image」「make a poster」「render a figure」，以及风格名「UniVLA 风格」「Physical Intelligence / PI / π 风格」「Fast-WAM 风格」「蓝皮书 / BlueBook 风格」。默认走 nano；用户点名「用 GPT」时走 gpt。

Write structured experiment reports from experiment-run output. Consumes results/<run_id>/ + EXPERIMENT_TRACKER.md. Reports are scannable: TL;DR + summary table first. Trigger when user says "写个报告", "总结实验结果", "实验报告", "experiment report", or after experiment-run completes.

Adversarial paper review + improvement loop. Sends paper PDF/manuscript to Codex GPT-5.5 as a senior ML reviewer, scores 7 dimensions, then iteratively fixes weaknesses until score ≥ 8. Use after paper-write, or when user says "improve paper", "审稿", "review my paper", "polish paper", "paper review", "提升论文质量", "review loop".

Write a structured rebuttal using Concede–Clarify–Commit structure. Parses reviewer comments, identifies overlapping concerns across reviewers, builds consolidated Table R1, produces venue-formatted rebuttal. Use when user says "写 rebuttal", "rebuttal", "回复审稿意见", "response to reviewers", or receives paper reviews and needs to respond.

Execute experiments from EXPERIMENT_TRACKER.md produced by experiment-plan. For each TODO run: launch training/eval, monitor via gpu-train-monitor, auto-diagnose failures via train-debug, collect results, update tracker. Bridges the gap between planning (experiment-plan) and reporting (experiment_report). Use when user says "跑实验", "run experiments", "执行实验", "按照 plan 跑", or has an EXPERIMENT_TRACKER.md ready to execute.

Automated full-spectrum research survey. Uses litian-academic-search (7-source) for academic papers + WebSearch site: filters for social media. Ingests all found papers via paper-read --light. Produces structured FINAL_REPORT.md per topic + concept glossary with pseudocode. Use when user wants "全面调研", "综合调研", "survey this topic", "调研一下", "帮我全面了解一下这个方向".

Turn a refined research proposal into a claim-driven experiment roadmap with progressive 4-stage execution order. Reads from research-refine output (FINAL_PROPOSAL.md). Produces EXPERIMENT_PLAN.md + EXPERIMENT_TRACKER.md. Use after research-refine, or when user says "实验计划", "experiment plan", "设计实验", "ablation matrix", "evaluation protocol", "run order".

Deploy and manage the GPU monitor data collector on Slurm clusters. Use this skill whenever the user asks to set up, deploy, configure, troubleshoot, or manage GPU monitoring on any cluster — including phrases like "deploy on BCC", "setup monitor", "monitor not working", "add a new cluster", "GPU data not updating", or "配置监控", "部署到新集群". Also use when the user mentions wanting to check why data stopped updating, restart the collector, or add a new cluster to the dashboard.