feat: backend gate fallback + metric_diff docs + submit_sft fix

- Auto-inject human-review gate when agent asks question but forgets gate entry
- Add complex_dev/complex_base field docs to program.md metric_diff table
- Fix submit_sft.sh minor issues
- jupyter_runtime improvements

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
wangsenhao
2026-05-26 15:42:02 +08:00
parent a09d4afacd
commit b48032e15c
5 changed files with 119 additions and 9 deletions
+24 -6
View File
@@ -297,14 +297,32 @@ class JupyterRuntimeSession:
if not self.skills_root:
return
chat_root = self.chat_workspace_root
src = f'{self.skills_root}/model-iteration/scripts/prepare_and_train_sft.py'
dst = f'{chat_root}/scripts/prepare_and_train_sft.py'
result = self.run_command(
(
# SKILL/program §5 让 agent 直接 `bash scripts/<x>` 调下面这些脚本,
# 不同步过去 → agent 当 "脚本不存在" 处理后会肉手写 yaml,常把
# imageCommand 里的 `/scripts/prepare_and_train_sft.py` 写丢前缀,
# 导致训练 pod 报 "No such file or directory"。每次 bind 全量覆盖
# 一份,跟 prepare_and_train_sft.py 同样 "刷最新版"。
script_names = (
'prepare_and_train_sft.py',
'submit_sft.sh',
'sft_train_job.yaml.tpl',
'submit_cml_eval.sh',
'resolve_run_ids.sh',
)
src_dir = f'{self.skills_root}/model-iteration/scripts'
dst_dir = f'{chat_root}/scripts'
copy_cmds = []
for name in script_names:
src = f'{src_dir}/{name}'
dst = f'{dst_dir}/{name}'
copy_cmds.append(
f'if [ -f {shlex.quote(src)} ]; then '
f'cp {shlex.quote(src)} {shlex.quote(dst)}; '
f'cp {shlex.quote(src)} {shlex.quote(dst)} && '
f'chmod +x {shlex.quote(dst)} 2>/dev/null || true; '
f'fi'
),
)
result = self.run_command(
' && '.join(copy_cmds),
timeout_seconds=timeout_seconds,
max_output_chars=2000,
)