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>
2026-05-26 15:42:02 +08:00
parent a09d4afacd
commit b48032e15c
5 changed files with 119 additions and 9 deletions
@@ -2031,6 +2031,9 @@ def create_app(state: AgentState) -> FastAPI:
            session_id,
        )
        run_active = _is_run_active(state, account_id, session_id)
+        state_entries = _maybe_inject_gate_fallback(
+            state_path, state_entries, sessions_dir, session_id, run_active,
+        )
        _apply_program_state(
            payload,
            state_entries,
@@ -2104,6 +2107,9 @@ def create_app(state: AgentState) -> FastAPI:
                        session_id,
                    )
                    run_active = _is_run_active(state, account_id, session_id)
+                    state_entries = _maybe_inject_gate_fallback(
+                        state_path, state_entries, sessions_dir, session_id, run_active,
+                    )
                    _apply_program_state(
                        payload,
                        state_entries,
@@ -6606,6 +6612,85 @@ def _compute_in_flight(
    }


+_QUESTION_PATTERN = re.compile(
+    r'[？?]|拍板|选择.*哪|决定|确认.*吗|建议.*哪|要不要|是否|请.*选|你.*决定|三个选项|哪个方案|哪条路'
+)
+
+
+def _last_assistant_text(sessions_dir: Path, session_id: str | None) -> str | None:
+    if not session_id:
+        return None
+    safe_id = _safe_session_id(session_id)
+    if safe_id is None:
+        return None
+    path = sessions_dir / safe_id / 'session.json'
+    try:
+        data = json.loads(path.read_text(encoding='utf-8'))
+    except (OSError, json.JSONDecodeError):
+        return None
+    messages = data.get('messages')
+    if not isinstance(messages, list):
+        return None
+    for msg in reversed(messages):
+        if not isinstance(msg, dict):
+            continue
+        if msg.get('role') != 'assistant':
+            continue
+        content = msg.get('content')
+        if isinstance(content, str) and content.strip():
+            return content.strip()
+        if isinstance(content, list):
+            for block in content:
+                if isinstance(block, dict) and block.get('type') == 'text':
+                    text = block.get('text', '').strip()
+                    if text:
+                        return text
+    return None
+
+
+def _maybe_inject_gate_fallback(
+    state_path: Path,
+    state_entries: list[dict[str, Any]],
+    sessions_dir: Path,
+    session_id: str | None,
+    run_active: bool | None,
+) -> list[dict[str, Any]]:
+    """If agent's last message asks a question but no gate is running,
+    append a synthetic human-review entry to program-state.jsonl."""
+    if run_active is not False:
+        return state_entries
+    for entry in reversed(state_entries):
+        step = entry.get('step')
+        if step in ('human-check', 'human-review') and entry.get('status') == 'running':
+            return state_entries
+    last_text = _last_assistant_text(sessions_dir, session_id)
+    if not last_text:
+        return state_entries
+    if not _QUESTION_PATTERN.search(last_text[-500:]):
+        return state_entries
+    run_id = None
+    for entry in reversed(state_entries):
+        rid = entry.get('run_id')
+        if isinstance(rid, str) and rid.strip():
+            run_id = rid.strip()
+            break
+    gate_entry: dict[str, Any] = {
+        'step': 'human-review',
+        'status': 'running',
+        'run_id': run_id or 'R0',
+        'reason': last_text[-300:],
+        'ts': datetime_utc_iso(),
+        '_synthetic': True,
+    }
+    try:
+        with open(state_path, 'a', encoding='utf-8') as f:
+            f.write(json.dumps(gate_entry, ensure_ascii=False) + '\n')
+    except OSError:
+        pass
+    state_entries.append(gate_entry)
+    return state_entries
+
+
 def _apply_program_state(
    payload: dict[str, Any],
    state_entries: list[dict[str, Any]],
@@ -1926,6 +1926,9 @@ function SkillInsertDialog() {
 								className="min-w-0 flex-1 text-left"
 								onClick={() => {
 									insertComposerText(`Use the ${skill.name} skill.\n\n`);
+									if (skill.name === "model-iteration") {
+										dispatchSkillToggled({ skill: skill.name, enabled: true });
+									}
 									setOpen(false);
 								}}
 							>
@@ -219,8 +219,10 @@ for k in prev:

 | 字段 | 含义 |
 | --- | --- |
-| `origin_predict_base` | 旧模型原始输出。具体格式（是否含 `complex=` 前缀、tag 包装、自定义 class 名等）以当前评测产出为准，**分析前 head 一下实物** |
-| `origin_predict_dev`  | 新模型原始输出。同上，格式以当前产出为准 |
+| `origin_predict_base` | 旧模型（baseline）原始输出。具体格式（是否含 `complex=` 前缀、tag 包装、自定义 class 名等）以当前评测产出为准，**分析前 head 一下实物** |
+| `origin_predict_dev`  | 新模型（迭代模型）原始输出。同上，格式以当前产出为准 |
+| `complex_dev` | 迭代模型的 complex 标签（true/false），从 `origin_predict_dev` 解析出的复杂度判定 |
+| `complex_base` | baseline 模型的 complex 标签（true/false），从 `origin_predict_base` 解析出的复杂度判定 |
 | `cleaned_predict_*`   | 清洗后输出，用于 GSB 对比 |
 | `label` / `code_label_base` | ground truth；`label` 为空时回退解析 `code_label_base` |

@@ -19,7 +19,9 @@ SFT_RUNDIC=${1:?usage: $0 <SFT_RUNDIC> [PREV_RUNDIC]}
 PREV_RUNDIC=${2:-$((SFT_RUNDIC-1))}

 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-ROOT=${AUTORESEARCH_ROOT:-/mnt/wangsenhao/autoresearch-zk}
+# chat-isolated 时代 AUTORESEARCH_ROOT == AUTORESEARCH_CHAT_ROOT；优先用 chat
+# root，避免 agent 没 export AUTORESEARCH_ROOT 时回退到全局共享路径（已废弃）。
+ROOT=${AUTORESEARCH_ROOT:-${AUTORESEARCH_CHAT_ROOT:-$(dirname "$SCRIPT_DIR")}}
 TPL=${SFT_TRAIN_JOB_TEMPLATE:-$SCRIPT_DIR/sft_train_job.yaml.tpl}
 YAML=/tmp/sft_train_job_r${SFT_RUNDIC}.yaml

@@ -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,
        )