audit-driven-retrain

name	audit-driven-retrain
description	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".
argument-hint	[--current-run <id>] [--hypothesis '...']
allowed-tools	Bash(*), Read, Write, Edit, Agent, Skill

/audit-driven-retrain — scancel → audit → patch → retry

借鉴 user 在 hf-jax v3→v6 实战 pattern。

Overview

4 阶段闭环：

1. scancel 现 run（保留 ckpt + log）
2. audit sub-agent (/spawn-task --type audit)：读 log + 最后 ckpt metric + 用户给的 hypothesis，定位 root cause
3. patch 用 Edit 改文件（小 diff）+ commit
4. retry 走 /smoke-test → /run-experiment

每轮自动 /auto-version snapshot 旧 run dir。

Workflow

RUN=${RUN:-$(/status --format json | jq -r '.runs[0].id')}
# Phase 1: scancel
ssh $HOST scancel <jobid>
# Phase 2: audit
Skill(/spawn-task --type audit --task "diagnose run $RUN, hypothesis: $HYP")
wait
# Phase 3: patch
# (user or agent reads audit result, makes Edit calls, commits)
# Phase 4: retry
Skill(/smoke-test)
Skill(/auto-version)
ssh $HOST sbatch <next.sbatch>

Output

• .auto-production/retrain/<round>/audit.md
• .auto-production/retrain/<round>/patch.diff
• 自动 commit fix(retrain-v<N>): <root-cause>

Composition

• 触发 conditions（自动）：训练 SR 卡住 / job 反复 FAILED / /pivot 决策 REFINE
• 后置 /run-experiment 起新 run