From trainhub
训练完成后下载产物，解析指标，归档到 weights/ 目录。接在 train-watch 检测到 COMPLETE 之后。触发词：收实验结果、下载 kernel output、归档 weights、parse results.csv。
npx claudepluginhub shallow-dusty/claude-plugins --plugin trainhubThis skill uses the workspace's default tool permissions.
| 参数 | 必填 | 含义 |
Creates isolated Git worktrees for feature branches with prioritized directory selection, gitignore safety checks, auto project setup for Node/Python/Rust/Go, and baseline verification.
Executes implementation plans in current session by dispatching fresh subagents per independent task, with two-stage reviews: spec compliance then code quality.
Dispatches parallel agents to independently tackle 2+ tasks like separate test failures or subsystems without shared state or dependencies.
| 参数 | 必填 | 含义 |
|---|---|---|
run | 是 | run_name |
platform | 否（默认读 .trainhub.json） | kaggle / ssh / colab |
kaggle kernels output <account>/<kernel_prefix>-<run_name> -p /tmp/kg-<run_name>
# 产物通常是 <run_name>.zip 里面打包了整个 results/<run_name>/
unzip -q /tmp/kg-<run_name>/<run_name>.zip -d /tmp/kg-<run_name>/extracted/
rsync -az <host>:<remote_dir>/<run_name>/results/ /tmp/ssh-<run_name>/
从 Google Drive 的 drive_folder/<run_name>/ 下载（用 rclone 或 gdown）。
.trainhub.json.metric）# 最佳 epoch 行（按 primary metric 的 best_fn=max/min）
awk -F',' -v col="metrics/mAP50(B)" '
NR==1 { for(i=1;i<=NF;i++) if($i==col) k=i; next }
{ if(!best || $k>best){ best=$k; row=$0 } }
END { print row }
' results.csv
提取：epoch, metrics/mAP50(B), metrics/mAP50-95(B), metrics/precision(B), metrics/recall(B)
.trainhub.json.archive）target="<weights_dir>/<filename_pattern>" # 如 weights/L2-P4-ema-p5-kaggle/
mkdir -p "$target"
cp /tmp/*/extracted/**/best.pt "$target/"
cp /tmp/*/extracted/**/results.csv "$target/"
若项目有 docs/CHRONICLE.md，追加一条实验条目，格式：
### <run_name>（YYYY-MM-DD 完成）
- 平台：<platform>
- 最佳 mAP50: <val> @ epoch <n>
- 对照 baseline <baseline_val> → Δ <diff> pp
- weights 归档：<target>
- 备注：<自动空或 TODO>
baseline 数值从 .trainhub.json.metric.baseline 读（没有就留 TODO）。
harvest 成功后删 /tmp/<platform>-<run_name>/ 临时目录，避免 /tmp 堆积。