Improve data skill routing and tool organization

src/agent_prompting.py

@@ -6,6 +6,7 @@ from pathlib import Path
 from .agent_context import build_context_snapshot
 from .agent_tools import AgentTool
 from .agent_types import AgentRuntimeConfig, ModelConfig
+from .bundled_skills import format_skills_for_system_prompt
 from .builtin_agents import AgentDefinition, format_agent_listing
 
 SYSTEM_PROMPT_DYNAMIC_BOUNDARY = '__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__'
@@ -98,6 +99,7 @@ def build_system_prompt_parts(
         get_doing_tasks_section(),
         get_actions_section(),
         get_using_your_tools_section(enabled_tool_names),
+        get_skill_guidance_section(prompt_context, enabled_tool_names),
         get_agent_guidance_section(enabled_tool_names, available_agents),
         get_plugin_guidance_section(prompt_context),
         get_mcp_guidance_section(prompt_context),
@@ -236,6 +238,22 @@ def get_using_your_tools_section(enabled_tool_names: set[str]) -> str:
     return '\n'.join(['# 使用工具', *prepend_bullets(items)])
 
 
+def get_skill_guidance_section(
+    prompt_context: PromptContext,
+    enabled_tool_names: set[str],
+) -> str:
+    if 'Skill' not in enabled_tool_names:
+        return ''
+    skill_listing = format_skills_for_system_prompt(cwd=prompt_context.cwd)
+    items = [
+        '当用户请求符合某个 skill 的适用范围时，优先调用 Skill 工具进入该 skill，不要先用 Agent 子任务或通用搜索绕路。',
+        '数据生成、标签边界、产品定义、示例 query 到数据集这类任务，优先使用 product-data skill。',
+        '线上 badcase 挖掘、router session 检索、候选转样本这类任务，优先使用 online-mining skill。',
+        '调用 skill 后遵守 skill 内部的人类 review 门禁和允许工具列表。',
+    ]
+    return '\n'.join(['# Skills', *prepend_bullets(items), '', skill_listing])
+
+
 def get_tone_and_style_section() -> str:
     items = [
         '回复保持简洁直接。',

src/agent_runtime.py

@@ -2012,78 +2012,96 @@ class LocalCodingAgent:
         return fallback
 
     def _format_generation_goal_review(self, goal: dict[str, object]) -> str:
+        target_summary = self._format_review_targets(goal)
         lines = [
-            '已生成待 review 的数据生成目标，本轮已暂停。',
-            '',
-            f"- goal_id: `{goal.get('goal_id', '')}`",
-            f"- revision: `{goal.get('revision', '')}`",
-            f"- dataset_label: {goal.get('dataset_label', '')}",
-            f"- 目标摘要: {goal.get('goal_summary', '')}",
+            '我先把生成目标整理好了，先确认边界，暂时不生成数据。',
+            f"- 数据集：{goal.get('dataset_label', '')}",
         ]
-        targets = goal.get('target_definitions')
-        if isinstance(targets, list) and targets:
-            lines.append('- target_definitions:')
-            for item in targets:
-                if isinstance(item, dict):
-                    lines.append(
-                        f"  - {item.get('name', '')}: `{item.get('target', '')}`；规则：{item.get('rule', '')}"
-                    )
-        elif goal.get('target'):
-            lines.append(f"- target: `{goal.get('target')}`")
-        if goal.get('plan_hint'):
-            lines.append(f"- 计划提示: {goal.get('plan_hint')}")
+        if target_summary:
+            lines.append(f'- 标签：{target_summary}')
+        if goal.get('goal_summary'):
+            lines.append(f"- 边界：{self._short_review_text(goal.get('goal_summary'))}")
+        scope_parts = []
+        if goal.get('coverage'):
+            scope_parts.append(f"覆盖：{self._short_review_text(goal.get('coverage'), limit=70)}")
+        if goal.get('exclusions'):
+            scope_parts.append(f"排除：{self._short_review_text(goal.get('exclusions'), limit=50)}")
+        if scope_parts:
+            lines.append('- 范围：' + '；'.join(scope_parts))
         open_questions = goal.get('open_questions')
         if isinstance(open_questions, list) and open_questions:
-            lines.append('- 待确认问题:')
-            for item in open_questions:
-                if isinstance(item, str) and item:
-                    lines.append(f'  - {item}')
-        source_refs = goal.get('source_refs')
-        if isinstance(source_refs, list) and source_refs:
-            lines.append('- 来源:')
-            for item in source_refs:
-                if isinstance(item, str) and item:
-                    lines.append(f'  - {item}')
+            questions = '；'.join(str(item).strip() for item in open_questions if isinstance(item, str) and item.strip())
+            if questions:
+                lines.append(f'- 待确认：{self._short_review_text(questions)}')
+        elif goal.get('plan_hint'):
+            lines.append(f"- 下一步：{self._short_review_text(goal.get('plan_hint'))}")
+        lines.append(f"- 内部：goal_id `{goal.get('goal_id', '')}`，revision `{goal.get('revision', '')}`")
+        lines.append('')
+        lines.append('你看这个目标是否准确？没问题就回“确认目标”；想改的话直接说哪里不对。')
+        return '\n'.join(lines)
+
+    def _format_generation_plan_review(self, plan: dict[str, object]) -> str:
+        if not plan.get('confirmed_goal_id'):
+            return self._format_direct_generation_plan_review(plan)
+        lines = [
+            '目标已确认。现在只补充生成参数，标签边界沿用上一步。',
+            f"- 数量：{plan.get('total_count', '')} 条",
+            f"- 轮次：{plan.get('turn_mix', '')}",
+            f"- 覆盖补充：{self._short_review_text(plan.get('coverage'))}",
+            f"- 排除：{self._short_review_text(plan.get('exclusions'))}",
+            f"- 输出：{plan.get('output_path', '')}",
+            f"- 内部：plan_id `{plan.get('plan_id', '')}`，revision `{plan.get('revision', '')}`",
+        ]
+        lines.append('')
+        lines.append('如果这个数量和路径可以，就回“确认，开始生成”；想调整就直接说，比如“改成 20 条，全单轮”。')
+        return '\n'.join(lines)
+
+    def _format_direct_generation_plan_review(self, plan: dict[str, object]) -> str:
+        target_summary = self._format_review_targets(plan)
+        lines = [
+            '我把目标和生成参数合成一次确认，确认后就开始生成。',
+            f"- 数据集：{plan.get('dataset_label', '')}",
+        ]
+        if target_summary:
+            lines.append(f'- 标签：{target_summary}')
         lines.extend(
             [
-                f"- 覆盖范围: {goal.get('coverage', '')}",
-                f"- 排除项: {goal.get('exclusions', '')}",
+                f"- 数量：{plan.get('total_count', '')} 条；轮次：{plan.get('turn_mix', '')}",
+                (
+                    '- 范围：'
+                    f"覆盖：{self._short_review_text(plan.get('coverage'), limit=70)}；"
+                    f"排除：{self._short_review_text(plan.get('exclusions'), limit=50)}"
+                ),
+                f"- 输出：{plan.get('output_path', '')}",
+                f"- 内部：plan_id `{plan.get('plan_id', '')}`，revision `{plan.get('revision', '')}`",
                 '',
-                '请 review 这个生成目标：需要修改就直接回复修改意见；认可的话回复“确认目标”。',
+                '如果目标、数量和路径都可以，就回“确认，开始生成”；想调整就直接说哪里改。',
             ]
         )
         return '\n'.join(lines)
 
-    def _format_generation_plan_review(self, plan: dict[str, object]) -> str:
-        lines = [
-            '已生成待 review 的数据生成计划，本轮已暂停。',
-            '',
-            f"- plan_id: `{plan.get('plan_id', '')}`",
-            f"- revision: `{plan.get('revision', '')}`",
-            f"- dataset_label: {plan.get('dataset_label', '')}",
-        ]
-        targets = plan.get('target_definitions')
+    def _format_review_targets(self, payload: dict[str, object]) -> str:
+        targets = payload.get('target_definitions')
         if isinstance(targets, list) and targets:
-            lines.append('- target_definitions:')
+            parts: list[str] = []
             for item in targets:
-                if isinstance(item, dict):
-                    lines.append(
-                        f"  - {item.get('name', '')}: `{item.get('target', '')}`；规则：{item.get('rule', '')}"
-                    )
-        elif plan.get('target'):
-            lines.append(f"- target: `{plan.get('target')}`")
-        lines.extend(
-            [
-                f"- 生成数量: {plan.get('total_count', '')}",
-                f"- 单轮/多轮: {plan.get('turn_mix', '')}",
-                f"- 覆盖范围: {plan.get('coverage', '')}",
-                f"- 排除项: {plan.get('exclusions', '')}",
-                f"- 落盘路径: {plan.get('output_path', '')}",
-                '',
-                '请 review 这份计划：需要修改就直接回复修改意见；认可的话回复“确认，开始生成”。',
-            ]
-        )
-        return '\n'.join(lines)
+                if not isinstance(item, dict):
+                    continue
+                name = str(item.get('name') or '').strip()
+                target = str(item.get('target') or '').strip()
+                if name and target:
+                    parts.append(f'{name} -> `{target}`')
+                elif target:
+                    parts.append(f'`{target}`')
+            return '；'.join(parts)
+        target = payload.get('target')
+        return f'`{target}`' if isinstance(target, str) and target.strip() else ''
+
+    def _short_review_text(self, value: object, *, limit: int = 120) -> str:
+        text = ' '.join(str(value or '').split())
+        if len(text) <= limit:
+            return text
+        return text[: limit - 1].rstrip() + '…'
 
     def _compact_prefix_count(self, session: AgentSessionState) -> int:
         prefix_count = 0
@@ -2388,14 +2406,29 @@ class LocalCodingAgent:
 
         # Explicit model param in arguments takes priority
         if isinstance(model_override, str) and model_override.strip():
-            return replace(self.model_config, model=model_override.strip())
+            resolved_model = self._resolve_child_model_name(model_override.strip())
+            return replace(self.model_config, model=resolved_model)
 
         # Agent definition model
         if agent_model and agent_model != 'inherit':
-            return replace(self.model_config, model=agent_model)
+            resolved_model = self._resolve_child_model_name(agent_model)
+            return replace(self.model_config, model=resolved_model)
 
         return self.model_config
 
+    def _resolve_child_model_name(self, requested_model: str) -> str:
+        """Map Claude-style child model aliases only when the parent model is Claude-like."""
+
+        normalized = requested_model.strip()
+        if normalized in {'haiku', 'sonnet', 'opus'} and not self._is_claude_model(self.model_config.model):
+            return self.model_config.model
+        return normalized
+
+    @staticmethod
+    def _is_claude_model(model: str) -> bool:
+        lowered = model.lower()
+        return lowered.startswith('claude') or '/claude' in lowered or 'anthropic' in lowered
+
     def _filter_tools_for_agent(
         self,
         agent_def: AgentDefinition,

src/agent_tool_core.py

@@ -0,0 +1,173 @@
+from __future__ import annotations
+
+import subprocess
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Callable, Union
+
+from .agent_types import AgentPermissions, AgentRuntimeConfig, ToolExecutionResult
+
+if TYPE_CHECKING:
+    from .account_runtime import AccountRuntime
+    from .ask_user_runtime import AskUserRuntime
+    from .config_runtime import ConfigRuntime
+    from .lsp_runtime import LSPRuntime
+    from .mcp_runtime import MCPRuntime
+    from .plan_runtime import PlanRuntime
+    from .remote_runtime import RemoteRuntime
+    from .remote_trigger_runtime import RemoteTriggerRuntime
+    from .search_runtime import SearchRuntime
+    from .task_runtime import TaskRuntime
+    from .team_runtime import TeamRuntime
+    from .workflow_runtime import WorkflowRuntime
+    from .worktree_runtime import WorktreeRuntime
+
+
+class ToolPermissionError(RuntimeError):
+    """Raised when the runtime configuration does not allow a tool action."""
+
+
+class ToolExecutionError(RuntimeError):
+    """Raised when a tool cannot complete because of invalid input or state."""
+
+
+@dataclass(frozen=True)
+class ToolExecutionContext:
+    root: Path
+    command_timeout_seconds: float
+    max_output_chars: int
+    permissions: AgentPermissions
+    scratchpad_directory: Path | None = None
+    python_env_dir: Path | None = None
+    extra_env: dict[str, str] = field(default_factory=dict)
+    tool_registry: dict[str, 'AgentTool'] | None = None
+    search_runtime: 'SearchRuntime | None' = None
+    account_runtime: 'AccountRuntime | None' = None
+    ask_user_runtime: 'AskUserRuntime | None' = None
+    config_runtime: 'ConfigRuntime | None' = None
+    lsp_runtime: 'LSPRuntime | None' = None
+    mcp_runtime: 'MCPRuntime | None' = None
+    remote_runtime: 'RemoteRuntime | None' = None
+    remote_trigger_runtime: 'RemoteTriggerRuntime | None' = None
+    plan_runtime: 'PlanRuntime | None' = None
+    task_runtime: 'TaskRuntime | None' = None
+    team_runtime: 'TeamRuntime | None' = None
+    workflow_runtime: 'WorkflowRuntime | None' = None
+    worktree_runtime: 'WorktreeRuntime | None' = None
+
+
+ToolHandler = Callable[
+    [dict[str, Any], ToolExecutionContext],
+    Union[str, tuple[str, dict[str, Any]]],
+]
+
+
+@dataclass(frozen=True)
+class AgentTool:
+    name: str
+    description: str
+    parameters: dict[str, Any]
+    handler: ToolHandler
+
+    def to_openai_tool(self) -> dict[str, object]:
+        return {
+            'type': 'function',
+            'function': {
+                'name': self.name,
+                'description': self.description,
+                'parameters': self.parameters,
+            },
+        }
+
+    def execute(
+        self,
+        arguments: dict[str, Any],
+        context: ToolExecutionContext,
+    ) -> ToolExecutionResult:
+        try:
+            result = self.handler(arguments, context)
+            if isinstance(result, tuple):
+                content, metadata = result
+            else:
+                content, metadata = result, {}
+            return ToolExecutionResult(
+                name=self.name,
+                ok=True,
+                content=content,
+                metadata=metadata,
+            )
+        except ToolPermissionError as exc:
+            return ToolExecutionResult(
+                name=self.name,
+                ok=False,
+                content=str(exc),
+                metadata={'error_kind': 'permission_denied'},
+            )
+        except (ToolExecutionError, OSError, subprocess.SubprocessError) as exc:
+            return ToolExecutionResult(
+                name=self.name,
+                ok=False,
+                content=str(exc),
+                metadata={'error_kind': 'tool_execution_error'},
+            )
+
+
+@dataclass(frozen=True)
+class ToolStreamUpdate:
+    kind: str
+    content: str = ''
+    stream: str | None = None
+    result: ToolExecutionResult | None = None
+    metadata: dict[str, Any] = field(default_factory=dict)
+
+
+def build_tool_context(
+    config: AgentRuntimeConfig,
+    *,
+    scratchpad_directory: Path | None = None,
+    python_env_dir: Path | None = None,
+    extra_env: dict[str, str] | None = None,
+    tool_registry: dict[str, AgentTool] | None = None,
+    search_runtime: 'SearchRuntime | None' = None,
+    account_runtime: 'AccountRuntime | None' = None,
+    ask_user_runtime: 'AskUserRuntime | None' = None,
+    config_runtime: 'ConfigRuntime | None' = None,
+    lsp_runtime: 'LSPRuntime | None' = None,
+    mcp_runtime: 'MCPRuntime | None' = None,
+    remote_runtime: 'RemoteRuntime | None' = None,
+    remote_trigger_runtime: 'RemoteTriggerRuntime | None' = None,
+    plan_runtime: 'PlanRuntime | None' = None,
+    task_runtime: 'TaskRuntime | None' = None,
+    team_runtime: 'TeamRuntime | None' = None,
+    workflow_runtime: 'WorkflowRuntime | None' = None,
+    worktree_runtime: 'WorktreeRuntime | None' = None,
+) -> ToolExecutionContext:
+    return ToolExecutionContext(
+        root=config.cwd.resolve(),
+        command_timeout_seconds=config.command_timeout_seconds,
+        max_output_chars=config.max_output_chars,
+        permissions=config.permissions,
+        scratchpad_directory=scratchpad_directory.resolve() if scratchpad_directory else None,
+        python_env_dir=(
+            python_env_dir.resolve()
+            if python_env_dir
+            else config.python_env_dir.resolve()
+            if config.python_env_dir
+            else None
+        ),
+        extra_env=dict(extra_env or {}),
+        tool_registry=tool_registry,
+        search_runtime=search_runtime,
+        account_runtime=account_runtime,
+        ask_user_runtime=ask_user_runtime,
+        config_runtime=config_runtime,
+        lsp_runtime=lsp_runtime,
+        mcp_runtime=mcp_runtime,
+        remote_runtime=remote_runtime,
+        remote_trigger_runtime=remote_trigger_runtime,
+        plan_runtime=plan_runtime,
+        task_runtime=task_runtime,
+        team_runtime=team_runtime,
+        workflow_runtime=workflow_runtime,
+        worktree_runtime=worktree_runtime,
+    )

src/agent_tool_specs/__init__.py

@@ -0,0 +1,5 @@
+"""工具声明目录。
+
+每个子模块负责一类工具的名称、描述和 JSON Schema；handler 由运行层注入。
+"""
+

src/agent_tool_specs/_builder.py

@@ -0,0 +1,15 @@
+from __future__ import annotations
+
+from collections.abc import Mapping
+
+from ..agent_tool_core import ToolHandler
+
+
+def resolve_handler(handlers: Mapping[str, ToolHandler], name: str, group: str) -> ToolHandler:
+    """从工具组 handler 映射中取出指定工具的执行函数。"""
+
+    try:
+        return handlers[name]
+    except KeyError as exc:
+        raise KeyError(f'Missing {group} tool handler: {name}') from exc
+

src/agent_tool_specs/data_agent.py

@@ -0,0 +1,523 @@
+from __future__ import annotations
+
+from collections.abc import Mapping
+
+from ..agent_tool_core import AgentTool, ToolHandler
+from ._builder import resolve_handler
+
+
+def build_data_agent_tools(handlers: Mapping[str, ToolHandler]) -> list[AgentTool]:
+    """构建数据 Agent 专用工具声明。
+
+    这里仅保存工具名称、描述和 JSON Schema，具体执行逻辑由调用方传入的 handler 提供。
+    这样新增数据 Agent 工具时，可以优先在本文件补充声明，再在实现模块中补充 handler。
+    """
+    return [
+        AgentTool(
+            name='data_agent_prepare_generation_goal',
+            description=(
+                'Create a pending data-agent generation goal from source evidence or user rules. '
+                'Use this before data_agent_prepare_generation_plan; show the returned goal to the user and wait for confirmation.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'dataset_label': {'type': 'string'},
+                    'goal_summary': {'type': 'string'},
+                    'target': {
+                        'type': 'string',
+                        'description': 'Single full target expression, e.g. Agent(tag="餐饮服务") or a function call. For multi-target boundary tasks, use target_definitions.',
+                    },
+                    'target_definitions': {
+                        'type': 'array',
+                        'items': {
+                            'type': 'object',
+                            'properties': {
+                                'name': {'type': 'string'},
+                                'target': {
+                                    'type': 'string',
+                                    'description': 'Full target expression, e.g. Agent(tag="餐饮服务"); do not shorten to a display name.',
+                                },
+                                'rule': {'type': 'string'},
+                            },
+                            'required': ['target'],
+                        },
+                    },
+                    'plan_hint': {
+                        'type': 'string',
+                        'description': (
+                            'Optional plain-language hint for the later plan, such as '
+                            '"建议先生成 50 条单轮，输出到 tasks/.../records.jsonl". '
+                            'Structured total_count, turn_mix, and output_path belong to data_agent_prepare_generation_plan.'
+                        ),
+                    },
+                    'coverage': {'type': 'string'},
+                    'exclusions': {'type': 'string'},
+                    'open_questions': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                    },
+                    'source_refs': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                    },
+                    'notes': {'type': 'string'},
+                },
+                'required': ['dataset_label', 'goal_summary', 'coverage', 'exclusions'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_prepare_generation_goal', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_confirm_generation_goal',
+            description=(
+                'Mark a pending data-agent generation goal as confirmed after the user explicitly approves it. '
+                'The returned confirmed_goal_id is required by data_agent_prepare_generation_plan.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'goal_id': {'type': 'string'},
+                    'confirmation': {
+                        'type': 'string',
+                        'description': 'The user approval text, such as 确认, 开始生成, or approve.',
+                    },
+                    'reviewed_revision': {
+                        'type': 'integer',
+                        'description': 'The goal revision shown to the user before confirmation.',
+                    },
+                },
+                'required': ['goal_id', 'confirmation'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_confirm_generation_goal', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_prepare_generation_plan',
+            description=(
+                'Create a pending data-agent generation plan. Use this before generating dataset draft text; '
+                'requires confirmed_goal_id from data_agent_confirm_generation_goal; show the returned plan to the user and wait for confirmation.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'confirmed_goal_id': {
+                        'type': 'string',
+                        'description': 'Returned by data_agent_confirm_generation_goal after the user reviews a separate generation goal. Omit only when direct_review is true.',
+                    },
+                    'direct_review': {
+                        'type': 'boolean',
+                        'description': (
+                            'Set true for simple hand-written rules where the user already supplied target labels '
+                            'and wants generation; this creates one combined goal+plan review instead of a separate goal review.'
+                        ),
+                    },
+                    'dataset_label': {'type': 'string'},
+                    'target': {
+                        'type': 'string',
+                        'description': 'Single full target expression. For multi-target boundary tasks, omit this and fill target_definitions.',
+                    },
+                    'target_definitions': {
+                        'type': 'array',
+                        'description': 'Optional target labels for multi-class boundary data.',
+                        'items': {
+                            'type': 'object',
+                            'properties': {
+                                'name': {'type': 'string'},
+                                'target': {
+                                    'type': 'string',
+                                    'description': 'Full target expression inherited from the confirmed goal, e.g. Agent(tag="餐饮服务").',
+                                },
+                                'rule': {'type': 'string'},
+                            },
+                            'required': ['target'],
+                        },
+                    },
+                    'total_count': {'type': 'integer', 'minimum': 1},
+                    'turn_mix': {'type': 'string', 'description': 'Single-turn and multi-turn count or ratio.'},
+                    'coverage': {'type': 'string', 'description': 'Query types, intent boundaries, or error types to cover.'},
+                    'exclusions': {'type': 'string', 'description': 'Negative examples or boundaries to avoid.'},
+                    'output_path': {'type': 'string', 'description': 'Where draft, records, and validation files should be written.'},
+                    'notes': {'type': 'string'},
+                },
+                'required': [
+                    'dataset_label',
+                    'total_count',
+                    'turn_mix',
+                    'coverage',
+                    'exclusions',
+                    'output_path',
+                ],
+            },
+            handler=resolve_handler(handlers, 'data_agent_prepare_generation_plan', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_load_input_sources',
+            description=(
+                'Load data-agent input files or directories and extract structured paragraphs and table previews '
+                'from xlsx, csv, docx, pdf, txt, md, json, or jsonl sources.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'paths': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': 'Workspace-relative files or directories to load.',
+                    },
+                    'max_files': {'type': 'integer', 'minimum': 1, 'maximum': 100},
+                    'max_paragraphs_per_file': {'type': 'integer', 'minimum': 1, 'maximum': 300},
+                    'max_tables_per_file': {'type': 'integer', 'minimum': 1, 'maximum': 100},
+                    'max_rows_per_table': {'type': 'integer', 'minimum': 1, 'maximum': 200},
+                    'max_cell_chars': {'type': 'integer', 'minimum': 20, 'maximum': 2000},
+                },
+                'required': ['paths'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_load_input_sources', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_extract_case_evidence',
+            description=(
+                'Extract query/badcase evidence from loaded input sources or source paths. '
+                'Profiles table columns first and returns required questions when query or expected label columns are ambiguous.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'paths': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': 'Workspace-relative files or directories. Omit when loaded_sources is provided.',
+                    },
+                    'loaded_sources': {
+                        'type': 'object',
+                        'description': 'Output from data_agent_load_input_sources.',
+                    },
+                    'field_mapping': {
+                        'type': 'object',
+                        'description': 'Optional role-to-column mapping, e.g. {"query":"query","expected_label":"预期domain"}.',
+                    },
+                    'max_cases': {'type': 'integer', 'minimum': 1, 'maximum': 1000},
+                },
+            },
+            handler=resolve_handler(handlers, 'data_agent_extract_case_evidence', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_render_source_context',
+            description=(
+                'Render loaded data-agent sources into LLM-readable evidence text with source refs. '
+                'Use this after data_agent_load_input_sources before asking the model to structure product semantics.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'paths': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': 'Workspace-relative files or directories. Omit when loaded_sources is provided.',
+                    },
+                    'loaded_sources': {
+                        'type': 'object',
+                        'description': 'Output from data_agent_load_input_sources.',
+                    },
+                    'max_chars': {'type': 'integer', 'minimum': 1000, 'maximum': 200000},
+                    'max_tables_per_source': {'type': 'integer', 'minimum': 1, 'maximum': 100},
+                    'max_rows_per_table': {'type': 'integer', 'minimum': 1, 'maximum': 200},
+                    'focus_keywords': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': 'Optional keywords used to keep only matching paragraphs/table rows.',
+                    },
+                },
+            },
+            handler=resolve_handler(handlers, 'data_agent_render_source_context', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_profile_router_sessions',
+            description=(
+                'Profile local router_session_parquet date partitions before mining online sessions. '
+                'Returns schema, sampled row count, distributions, and example sessions.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'dates': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': 'Date partitions such as ["20260428"]. Resolved under router_session_parquet/date=YYYYMMDD.',
+                    },
+                    'paths': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': 'Workspace-relative parquet files or date directories. Use when dates is not enough.',
+                    },
+                    'max_files': {'type': 'integer', 'minimum': 1, 'maximum': 200},
+                    'max_rows_per_file': {'type': 'integer', 'minimum': 1, 'maximum': 100000},
+                    'examples_limit': {'type': 'integer', 'minimum': 0, 'maximum': 20},
+                },
+            },
+            handler=resolve_handler(handlers, 'data_agent_profile_router_sessions', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_search_router_sessions',
+            description=(
+                'Search local router_session_parquet data with reviewable filters such as device, domain, intent, func, '
+                'query keywords, regex, turn count, and match scope. Returns turn-level candidates with prev turns.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'dates': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': 'Date partitions such as ["20260428"]. Resolved under router_session_parquet/date=YYYYMMDD.',
+                    },
+                    'paths': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': 'Workspace-relative parquet files or date directories. Use when dates is not enough.',
+                    },
+                    'devices': {'type': 'array', 'items': {'type': 'string'}},
+                    'domains': {'type': 'array', 'items': {'type': 'string'}},
+                    'intents': {'type': 'array', 'items': {'type': 'string'}},
+                    'funcs': {'type': 'array', 'items': {'type': 'string'}},
+                    'query_keywords': {'type': 'array', 'items': {'type': 'string'}},
+                    'keyword_match_mode': {'type': 'string', 'enum': ['any', 'all']},
+                    'query_regex': {'type': 'string'},
+                    'match_scope': {'type': 'string', 'enum': ['any_turn', 'last_turn']},
+                    'turn_count_min': {'type': 'integer', 'minimum': 1},
+                    'turn_count_max': {'type': 'integer', 'minimum': 1},
+                    'max_files': {'type': 'integer', 'minimum': 1, 'maximum': 200},
+                    'max_rows_per_file': {'type': 'integer', 'minimum': 1, 'maximum': 100000},
+                    'max_candidates': {'type': 'integer', 'minimum': 1, 'maximum': 5000},
+                },
+            },
+            handler=resolve_handler(handlers, 'data_agent_search_router_sessions', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_sample_router_candidates',
+            description=(
+                'Sample router session candidates for human review and return candidate-level summary statistics. '
+                'Use after data_agent_search_router_sessions.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'candidates': {
+                        'oneOf': [
+                            {'type': 'array'},
+                            {'type': 'object'},
+                            {'type': 'string'},
+                        ],
+                        'description': 'Candidates array, search result object, or JSON string containing candidates.',
+                    },
+                    'sample_size': {'type': 'integer', 'minimum': 1, 'maximum': 1000},
+                    'strategy': {'type': 'string', 'enum': ['random', 'first', 'stride']},
+                    'seed': {'type': 'integer'},
+                },
+                'required': ['candidates'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_sample_router_candidates', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_convert_router_candidates_to_records',
+            description=(
+                'Convert reviewed online router session candidates directly into canonical data-agent records. '
+                'Use this when the user wants to keep mined online data as samples; do not use generation-plan tools for this path.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'candidates': {
+                        'oneOf': [
+                            {'type': 'array'},
+                            {'type': 'object'},
+                            {'type': 'string'},
+                        ],
+                        'description': 'Candidates array, search/sample result object, or JSON string containing candidates.',
+                    },
+                    'dataset_label': {'type': 'string'},
+                    'default_target': {
+                        'type': 'string',
+                        'description': 'Target label used for included candidates unless review_decisions provides a target.',
+                    },
+                    'review_decisions': {
+                        'type': 'array',
+                        'description': 'Optional include/exclude/uncertain decisions keyed by semantic_session_id or req_id.',
+                        'items': {
+                            'type': 'object',
+                            'properties': {
+                                'semantic_session_id': {'type': 'string'},
+                                'req_id': {'type': 'string'},
+                                'matched_turn_index': {'type': 'integer'},
+                                'decision': {'type': 'string', 'enum': ['include', 'exclude', 'uncertain']},
+                                'target': {'type': 'string'},
+                                'notes': {'type': 'string'},
+                            },
+                        },
+                    },
+                    'batch_id': {'type': 'string'},
+                    'include_uncertain': {'type': 'boolean'},
+                },
+                'required': ['candidates', 'dataset_label'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_convert_router_candidates_to_records', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_show_generation_plan',
+            description='Show a data-agent generation plan for human review.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'plan_id': {'type': 'string'},
+                },
+                'required': ['plan_id'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_show_generation_plan', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_update_generation_plan',
+            description=(
+                'Update a pending data-agent generation plan from human review feedback, then show it again for review.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'plan_id': {'type': 'string'},
+                    'review_feedback': {
+                        'type': 'string',
+                        'description': 'Human review feedback that explains why the plan is changing.',
+                    },
+                    'updates': {
+                        'type': 'object',
+                        'description': 'Fields to update on the plan.',
+                    },
+                },
+                'required': ['plan_id', 'review_feedback', 'updates'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_update_generation_plan', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_confirm_generation_plan',
+            description=(
+                'Mark a pending data-agent generation plan as confirmed after the user explicitly approves it.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'plan_id': {'type': 'string'},
+                    'confirmation': {
+                        'type': 'string',
+                        'description': 'The user approval text, such as 确认, 开始生成, or approve.',
+                    },
+                    'reviewed_revision': {
+                        'type': 'integer',
+                        'description': 'The plan revision shown to the user before confirmation.',
+                    },
+                },
+                'required': ['plan_id', 'confirmation'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_confirm_generation_plan', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_normalize_dataset_draft',
+            description=(
+                'Parse dataset draft text written with 用户/小爱/target blocks and build canonical '
+                'data-agent records with generated record IDs, timestamps, source metadata, context, and labels. '
+                'Generated data requires a confirmed_plan_id from data_agent_confirm_generation_plan.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'draft_text': {
+                        'type': 'string',
+                        'description': 'Dataset draft text in dataset draft text v1 format.',
+                    },
+                    'batch_id': {
+                        'type': 'string',
+                        'description': 'Batch identifier used in generated record_id values.',
+                    },
+                    'source_type': {
+                        'type': 'string',
+                        'enum': ['generated', 'online', 'manual', 'mixed'],
+                    },
+                    'base_timestamp': {
+                        'type': 'integer',
+                        'description': 'Optional millisecond timestamp for deterministic generated records.',
+                    },
+                    'timestamp_step_ms': {
+                        'type': 'integer',
+                        'minimum': 1,
+                        'description': 'Millisecond gap between adjacent turns. Defaults to 60000.',
+                    },
+                    'default_request_id': {
+                        'type': 'string',
+                        'description': 'Request ID to use when a case does not provide a real request_id.',
+                    },
+                    'confirmed_plan_id': {
+                        'type': 'string',
+                        'description': 'Required for generated data; returned by data_agent_confirm_generation_plan.',
+                    },
+                },
+                'required': ['draft_text'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_normalize_dataset_draft', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_validate_dataset_records',
+            description=(
+                'Validate canonical data-agent records for required fields, labels, source metadata, '
+                'prev_session structure, timestamp order, and 5-minute prompt stitching constraints.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'records': {
+                        'oneOf': [
+                            {'type': 'array'},
+                            {'type': 'string'},
+                        ],
+                        'description': 'Canonical records as an array, or a JSON string containing the array.',
+                    },
+                },
+                'required': ['records'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_validate_dataset_records', 'data-agent'),
+        ),
+        AgentTool(
+            name='data_agent_export_dataset_records',
+            description=(
+                'Validate canonical data-agent records and write them to a workspace file. '
+                'Defaults to compact JSONL, one canonical record per line, so the model does not hand-write JSON output.'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'records': {
+                        'oneOf': [
+                            {'type': 'array'},
+                            {'type': 'string'},
+                        ],
+                        'description': 'Canonical records as an array, or a JSON string containing the array.',
+                    },
+                    'output_path': {
+                        'type': 'string',
+                        'description': 'Workspace-relative path to write, usually ending in .jsonl.',
+                    },
+                    'output_format': {
+                        'type': 'string',
+                        'enum': ['jsonl', 'json'],
+                        'description': 'Defaults to jsonl. json writes a compact JSON array.',
+                    },
+                    'require_validation_ok': {
+                        'type': 'boolean',
+                        'description': 'When true, refuse to write records with validation errors.',
+                    },
+                    'overwrite': {
+                        'type': 'boolean',
+                        'description': 'When false, refuse to overwrite an existing file.',
+                    },
+                },
+                'required': ['records', 'output_path'],
+            },
+            handler=resolve_handler(handlers, 'data_agent_export_dataset_records', 'data-agent'),
+        ),
+    ]

src/agent_tool_specs/execution.py

@@ -0,0 +1,123 @@
+from __future__ import annotations
+
+from collections.abc import Mapping
+
+from ..agent_tool_core import AgentTool, ToolHandler
+from ._builder import resolve_handler
+
+
+def build_execution_tools(handlers: Mapping[str, ToolHandler]) -> list[AgentTool]:
+    """构建本地命令、Python 执行与短等待工具声明。"""
+
+    return [
+        AgentTool(
+            name='python_exec',
+            description=(
+                '优先用本工具执行小型 Python 代码或项目内 Python 脚本，用于结构化文件分析、'
+                'JSON/JSONL 处理、批量校验、数据抽样和快速计算。默认使用当前用户独立 Python venv；'
+                '不要用 bash 运行 python/python3/.venv/bin/python。不要用本工具安装依赖。'
+                '缺包时应向用户确认后再处理依赖。一次性分析优先传 code；如需写临时文件，'
+                '必须写入环境变量 PYTHON_EXEC_SCRATCHPAD 指向的会话隔离目录，'
+                '不要在项目根目录创建临时 .py 脚本。'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'code': {
+                        'type': 'string',
+                        'description': '要通过 python -c 执行的 Python 代码。code 和 script_path 必须二选一；一次性分析优先使用 code。',
+                    },
+                    'script_path': {
+                        'type': 'string',
+                        'description': '工作区内已有、需要长期复用的 Python 脚本路径。code 和 script_path 必须二选一；不要为一次性分析在项目根目录新建脚本。',
+                    },
+                    'args': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': '传给脚本的命令行参数。仅在 script_path 模式下使用。',
+                    },
+                    'stdin': {
+                        'type': 'string',
+                        'description': '可选标准输入内容。',
+                    },
+                    'timeout_seconds': {
+                        'type': 'number',
+                        'minimum': 1,
+                        'description': '可选超时时间，默认使用当前会话命令超时。',
+                    },
+                    'max_output_chars': {
+                        'type': 'integer',
+                        'minimum': 100,
+                        'description': '可选输出截断长度，默认使用当前会话输出限制。',
+                    },
+                },
+            },
+            handler=resolve_handler(handlers, 'python_exec', 'execution'),
+        ),
+        AgentTool(
+            name='python_package',
+            description=(
+                '在当前用户独立 Python venv 中检查或安装 Python 包。'
+                '用于 python_exec 缺少 pandas、pyarrow、openpyxl 等分析依赖时；'
+                '不要通过 bash 执行 pip，也不要安装到项目 .venv。'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'action': {
+                        'type': 'string',
+                        'enum': ['show', 'install'],
+                        'description': 'show 查看当前 Python 环境和 pip；install 安装 packages。',
+                    },
+                    'packages': {
+                        'type': 'array',
+                        'items': {'type': 'string'},
+                        'description': '要安装的包名或 pip requirement spec，例如 pandas、pyarrow==15.0.0。',
+                    },
+                    'timeout_seconds': {
+                        'type': 'number',
+                        'minimum': 1,
+                        'maximum': 600,
+                        'description': '安装超时时间，默认 120 秒。',
+                    },
+                    'max_output_chars': {
+                        'type': 'integer',
+                        'minimum': 1000,
+                        'maximum': 50000,
+                    },
+                },
+                'required': ['action'],
+            },
+            handler=resolve_handler(handlers, 'python_package', 'execution'),
+        ),
+        AgentTool(
+            name='bash',
+            description=(
+                '运行真实 shell 命令，例如系统命令、进程控制、git 只读检查或用户已确认的依赖安装。'
+                '不要用 bash 执行 python、python3、pip 或 .venv/bin/python；结构化文件分析、'
+                'JSON/JSONL 处理、批量校验、数据抽样和快速计算必须优先使用 python_exec。'
+                'Python 包安装必须优先使用 python_package。'
+            ),
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'command': {'type': 'string'},
+                },
+                'required': ['command'],
+            },
+            handler=resolve_handler(handlers, 'bash', 'execution'),
+        ),
+        AgentTool(
+            name='sleep',
+            description='Pause execution briefly for bounded local wait flows.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'seconds': {'type': 'number', 'minimum': 0.0, 'maximum': 5.0},
+                },
+                'required': ['seconds'],
+            },
+            handler=resolve_handler(handlers, 'sleep', 'execution'),
+        ),
+    ]
+

src/agent_tool_specs/files.py

@@ -0,0 +1,111 @@
+from __future__ import annotations
+
+from collections.abc import Mapping
+
+from ..agent_tool_core import AgentTool, ToolHandler
+from ._builder import resolve_handler
+
+
+def build_file_tools(handlers: Mapping[str, ToolHandler]) -> list[AgentTool]:
+    """构建工作区文件读写与文本检索工具声明。"""
+
+    return [
+        AgentTool(
+            name='list_dir',
+            description='List files and directories under a workspace path.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'path': {'type': 'string', 'description': 'Relative path from workspace root.'},
+                    'max_entries': {'type': 'integer', 'minimum': 1, 'maximum': 500},
+                },
+            },
+            handler=resolve_handler(handlers, 'list_dir', 'file'),
+        ),
+        AgentTool(
+            name='read_file',
+            description='Read the contents of a UTF-8 text file inside the workspace.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'path': {'type': 'string', 'description': 'Relative file path from workspace root.'},
+                    'start_line': {'type': 'integer', 'minimum': 1},
+                    'end_line': {'type': 'integer', 'minimum': 1},
+                },
+                'required': ['path'],
+            },
+            handler=resolve_handler(handlers, 'read_file', 'file'),
+        ),
+        AgentTool(
+            name='write_file',
+            description='Write a complete file inside the workspace. Creates parent directories when needed.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'path': {'type': 'string'},
+                    'content': {'type': 'string'},
+                },
+                'required': ['path', 'content'],
+            },
+            handler=resolve_handler(handlers, 'write_file', 'file'),
+        ),
+        AgentTool(
+            name='edit_file',
+            description='Replace text inside a workspace file using exact string matching.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'path': {'type': 'string'},
+                    'old_text': {'type': 'string'},
+                    'new_text': {'type': 'string'},
+                    'replace_all': {'type': 'boolean'},
+                },
+                'required': ['path', 'old_text', 'new_text'],
+            },
+            handler=resolve_handler(handlers, 'edit_file', 'file'),
+        ),
+        AgentTool(
+            name='notebook_edit',
+            description='Edit a Jupyter notebook cell by replacing or appending source in a .ipynb file.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'path': {'type': 'string'},
+                    'cell_index': {'type': 'integer', 'minimum': 0},
+                    'source': {'type': 'string'},
+                    'cell_type': {'type': 'string'},
+                    'create_cell': {'type': 'boolean'},
+                },
+                'required': ['path', 'cell_index', 'source'],
+            },
+            handler=resolve_handler(handlers, 'notebook_edit', 'file'),
+        ),
+        AgentTool(
+            name='glob_search',
+            description='Find files matching a glob pattern inside the workspace.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'pattern': {'type': 'string'},
+                },
+                'required': ['pattern'],
+            },
+            handler=resolve_handler(handlers, 'glob_search', 'file'),
+        ),
+        AgentTool(
+            name='grep_search',
+            description='Search for a string or regular expression inside workspace files.',
+            parameters={
+                'type': 'object',
+                'properties': {
+                    'pattern': {'type': 'string'},
+                    'path': {'type': 'string'},
+                    'literal': {'type': 'boolean'},
+                    'max_matches': {'type': 'integer', 'minimum': 1, 'maximum': 500},
+                },
+                'required': ['pattern'],
+            },
+            handler=resolve_handler(handlers, 'grep_search', 'file'),
+        ),
+    ]
+

src/data_agent_records.py

@@ -49,6 +49,8 @@ def prepare_generation_goal(
     if missing:
         raise DataRecordError('missing required goal fields: ' + ', '.join(missing))
     normalized_targets = _normalize_target_definitions(target_definitions)
+    if not target.strip() and not normalized_targets:
+        raise DataRecordError('generation goal must include target or target_definitions')
     state = _load_goal_state(root)
     sequence = int(state.get('next_sequence') or 1)
     goal_id = f'data_goal_{sequence:06d}'
@@ -186,6 +188,12 @@ def prepare_generation_plan(
         }
     )
     normalized_targets = _normalize_target_definitions(target_definitions)
+    if confirmed_goal is not None and not target.strip() and not normalized_targets:
+        normalized_targets = _normalize_target_definitions(
+            confirmed_goal.get('target_definitions') if isinstance(confirmed_goal, dict) else None
+        )
+        if not normalized_targets:
+            target = str(confirmed_goal.get('target') or '').strip()
     if not target.strip() and not normalized_targets:
         missing.append('target')
     if missing:

src/skills/bundled/online-mining/SKILL.md

@@ -8,6 +8,8 @@ allowed_tools: read_file, write_file, edit_file, grep_search, glob_search, ask_u
 
 使用这个 skill 处理“badcase 或标签定义 -> 挖掘策略 -> 候选召回 -> 抽样 review -> 策略迭代 -> mined dataset”的工作流。
 
+所有线上挖掘产物必须放在当前用户当前会话的 output 目录下。不要把 records 或 review 结果写到项目根目录的 `output/`、`tasks/` 或源码目录。工具会把相对 `output_path` 自动路由到会话 output 目录；展示给用户时以工具返回的实际路径为准。
+
 ## 两条分支
 
 线上挖掘后必须先判断用户要的是哪条分支。

src/skills/bundled/product-data/SKILL.md

@@ -50,6 +50,13 @@ source_refs:
 
 默认不要一步到位生成数据。除非用户已经明确说“开始生成”“确认计划”“按这个计划生成”或同义表达，否则只能做目标对齐、计划草案和问题确认。
 
+为了减少重复确认，优先按下面两种门禁模式选择：
+
+- **一次确认模式**：用户直接给出手写规则、完整 target 表达，并且明确希望生成数据时，直接调用 `data_agent_prepare_generation_plan`，传 `direct_review=true` 和 `target_definitions`。这一次 review 同时确认目标、数量、轮次和路径；用户回复“确认，开始生成”后即可调用 `data_agent_confirm_generation_plan`。
+- **两段确认模式**：用户提供文件、表格、badcase、长文档，或者标签/边界/字段含义有歧义时，先用 `data_agent_prepare_generation_goal` 做目标 review；目标确认后再做 plan review。
+
+所有数据产物必须放在当前用户当前会话的 output 目录下。不要把 records、draft 或 validation 写到项目根目录的 `output/`、`tasks/` 或其他源码目录。工具会把相对 `output_path` 自动路由到会话 output 目录；展示给用户时以工具返回的实际路径为准。
+
 开始生成前必须确认这些信息：
 
 - `dataset_label`：数据集或专题名称。
@@ -61,12 +68,31 @@ source_refs:
 
 如果任一信息缺失，不要生成数据，不要调用 `data_agent_prepare_generation_plan`，不要调用 `data_agent_normalize_dataset_draft`，不要调用 `data_agent_validate_dataset_records`，只向用户提出需要确认的问题。
 
-信息完整后，调用 `data_agent_prepare_generation_goal` 创建 pending goal。这个工具会暂停本轮，必须把返回的 `generation_goal` 展示给用户 review。用户确认 goal 之后，调用 `data_agent_confirm_generation_goal` 获取 `confirmed_goal_id`，再调用 `data_agent_prepare_generation_plan` 创建 pending plan。创建 plan 后也会暂停本轮，必须等待用户 review。
+两段确认模式下，信息完整后，调用 `data_agent_prepare_generation_goal` 创建 pending goal。这个工具会暂停本轮，必须把返回的 `generation_goal` 展示给用户 review。用户确认 goal 之后，调用 `data_agent_confirm_generation_goal` 获取 `confirmed_goal_id`，再调用 `data_agent_prepare_generation_plan` 创建 pending plan。创建 plan 后也会暂停本轮，必须等待用户 review。
+
+一次确认模式下，不要先创建 generation goal；直接创建 pending plan，并在 plan 里包含 `target_definitions`、`total_count`、`turn_mix`、`coverage`、`exclusions`、`output_path`。不要让用户先确认目标再确认计划。
 
 用户确认后，拿到 `confirmed_plan_id`，才能生成 dataset draft text，并继续调用工具。
 
 `data_agent_normalize_dataset_draft` 对生成数据有代码级门禁：没有 `confirmed_plan_id`，或者计划未确认，会拒绝执行。
 
+如果用户在原始需求里已经写出 `Agent(tag="xxx")`、function 调用或其他完整标签表达，`target` 必须原样保留这个完整表达，不要简化成纯标签名。例如用户说 `Agent(tag="餐饮服务")`，则 `target_definitions[*].target` 和后续 draft 的 `target:` 都必须写 `Agent(tag="餐饮服务")`，不要写成 `餐饮服务`。
+
+review 展示必须简短清晰，不要重复解释工具和流程。每次 review 最多展示 6 行，格式优先如下：
+
+```text
+我先把生成目标整理好了，先确认边界，暂时不生成数据。
+- 数据集：xxx
+- 标签：A -> Agent(tag="A")；B -> Agent(tag="B")
+- 边界：一句话说明核心判定规则
+- 覆盖：一句话说明主要 case 类型
+- 内部：goal_id `...`，revision `...`
+
+你看这个目标是否准确？没问题就回“确认目标”；想改的话直接说哪里不对。
+```
+
+计划 review 也最多展示 6 行，只展示数量、轮次、覆盖、输出路径和确认口令。不要把 goal 的完整内容再次复制到 plan review 中，开头必须说明“目标已确认，现在只补充生成参数，标签边界沿用上一步”。确认口令可以自然一点，例如“如果这个数量和路径可以，就回‘确认，开始生成’；想调整就直接说，比如‘改成 20 条，全单轮’。”
+
 多标签边界数据不要拆成多个互不相关的单标签计划。应该创建一个计划，并在 `target_definitions` 中列出所有候选标签。例如：
 
 ```json
@@ -96,9 +122,9 @@ source_refs:
 2. 用 `data_agent_render_source_context` 把输入渲染成大模型可读 evidence text。
 3. 大模型只基于 evidence text 抽取 `generation_goal`，包括 `dataset_label`、`target_definitions`、`plan_hint`、`coverage`、`exclusions`、`open_questions`、`source_refs`。
 4. 如果目标标签、边界或字段含义不清楚，先用普通回复向用户提问并停止。
-5. 信息足够时，调用 `data_agent_prepare_generation_goal` 创建 pending goal，并停止等待用户 review。
-6. 用户确认 generation goal 后，调用 `data_agent_confirm_generation_goal` 获取 `confirmed_goal_id`。
-7. 调用 `data_agent_prepare_generation_plan` 创建 pending 计划，必须传入 `confirmed_goal_id`。
+5. 如果是手写规则且信息完整，调用 `data_agent_prepare_generation_plan`，设置 `direct_review=true`，创建一次确认的 pending 计划，并停止等待用户 review。
+6. 如果是文件/示例归纳/歧义场景，调用 `data_agent_prepare_generation_goal` 创建 pending goal，并停止等待用户 review。
+7. 用户确认 generation goal 后，调用 `data_agent_confirm_generation_goal` 获取 `confirmed_goal_id`，再调用 `data_agent_prepare_generation_plan` 创建 pending 计划。
 8. 展示计划后停止本轮，等待用户 review。
 9. 用户提出修改意见时，调用 `data_agent_update_generation_plan`，再展示计划。
 10. 用户明确确认当前计划版本后，调用 `data_agent_confirm_generation_plan`。
@@ -201,9 +227,9 @@ canonical records 落盘必须使用 `data_agent_export_dataset_records`，默
 1. 直接从用户描述里抽取标签边界。
 2. 多标签边界必须使用 `target_definitions`，不要拆成多个无关单标签计划。
 3. 识别规则中的冲突词、优先级和反例。
-4. 整理为 `generation_goal` 并调用 `data_agent_prepare_generation_goal`。
-5. 对不明确的 target、数量、单轮/多轮比例、输出路径提出问题。
-6. 用户确认 generation goal 后，调用 `data_agent_confirm_generation_goal`，再调用 `data_agent_prepare_generation_plan`。
+4. 如果用户已经给出完整 target 表达，直接写入 `target_definitions`；如果 target 表达不明确，先提出问题。
+5. 如果数量、轮次或输出路径未指定，可以由模型给出保守建议，走一次确认模式；不要为了这些默认参数单独多问一轮。
+6. 调用 `data_agent_prepare_generation_plan`，传入 `direct_review=true`，让用户一次确认目标和生成参数。
 
 ### 示例归纳型