Add run-scoped skill and prompt asset management

backend/agents/analyst.py

@@ -11,9 +11,7 @@ from agentscope.message import Msg
 
 from ..config.constants import ANALYST_TYPES
 from ..utils.progress import progress
-from .prompt_loader import PromptLoader
-
-_prompt_loader = PromptLoader()
+from .prompt_factory import build_agent_system_prompt, clear_prompt_factory_cache
 
 
 class AnalystAgent(ReActAgent):
@@ -55,9 +53,10 @@ class AnalystAgent(ReActAgent):
 
         if agent_id is None:
             agent_id = analyst_type
+        self.agent_id = agent_id
 
         self.config = config or {}
-
+        self.toolkit = toolkit
         sys_prompt = self._load_system_prompt()
 
         kwargs = {
@@ -77,27 +76,11 @@ class AnalystAgent(ReActAgent):
 
     def _load_system_prompt(self) -> str:
         """Load system prompt for analyst"""
-        personas_config = _prompt_loader.load_yaml_config(
-            "analyst",
-            "personas",
-        )
-        persona = personas_config.get(self.analyst_type_key, {})
-
-        # Get focus items and format as bullet points
-        focus_items = persona.get("focus", [])
-        focus_text = "\n".join(f"- {item}" for item in focus_items)
-
-        # Get description
-        description = persona.get("description", "").strip()
-
-        return _prompt_loader.load_prompt(
-            "analyst",
-            "system",
-            variables={
-                "analyst_type": self.analyst_persona,
-                "focus": focus_text,
-                "description": description,
-            },
+        return build_agent_system_prompt(
+            agent_id=self.agent_id,
+            config_name=self.config.get("config_name", "default"),
+            toolkit=self.toolkit,
+            analyst_type=self.analyst_type_key,
         )
 
     async def reply(self, x: Msg = None) -> Msg:
@@ -131,3 +114,15 @@ class AnalystAgent(ReActAgent):
             )
 
         return result
+
+    def reload_runtime_assets(self, active_skill_dirs: Optional[list] = None) -> None:
+        """Reload toolkit and system prompt from current run assets."""
+        from .toolkit_factory import create_agent_toolkit
+
+        clear_prompt_factory_cache()
+        self.toolkit = create_agent_toolkit(
+            self.agent_id,
+            self.config.get("config_name", "default"),
+            active_skill_dirs=active_skill_dirs,
+        )
+        self.sys_prompt = self._load_system_prompt()

backend/agents/portfolio_manager.py

@@ -12,9 +12,7 @@ from agentscope.message import Msg, TextBlock
 from agentscope.tool import Toolkit, ToolResponse
 
 from ..utils.progress import progress
-from .prompt_loader import PromptLoader
-
-_prompt_loader = PromptLoader()
+from .prompt_factory import build_agent_system_prompt, clear_prompt_factory_cache
 
 
 class PMAgent(ReActAgent):
@@ -36,6 +34,9 @@ class PMAgent(ReActAgent):
         margin_requirement: float = 0.25,
         config: Optional[Dict[str, Any]] = None,
         long_term_memory: Optional[LongTermMemoryBase] = None,
+        toolkit_factory: Any = None,
+        toolkit_factory_kwargs: Optional[Dict[str, Any]] = None,
+        toolkit: Optional[Toolkit] = None,
     ):
         self.config = config or {}
 
@@ -49,11 +50,28 @@ class PMAgent(ReActAgent):
 
         # Decisions made in current cycle
         self._decisions: Dict[str, Dict] = {}
+        toolkit_factory_kwargs = toolkit_factory_kwargs or {}
+        self._toolkit_factory = toolkit_factory
+        self._toolkit_factory_kwargs = toolkit_factory_kwargs
 
-        # Create toolkit
-        toolkit = self._create_toolkit()
+        # Create toolkit after local state is ready so bound tool methods can be registered.
+        if toolkit is None:
+            if toolkit_factory is not None:
+                toolkit = toolkit_factory(
+                    name,
+                    self.config.get("config_name", "default"),
+                    owner=self,
+                    **toolkit_factory_kwargs,
+                )
+            else:
+                toolkit = self._create_toolkit()
+        self.toolkit = toolkit
 
-        sys_prompt = _prompt_loader.load_prompt("portfolio_manager", "system")
+        sys_prompt = build_agent_system_prompt(
+            agent_id=name,
+            config_name=self.config.get("config_name", "default"),
+            toolkit=self.toolkit,
+        )
 
         kwargs = {
             "name": name,
@@ -186,3 +204,25 @@ class PMAgent(ReActAgent):
     def update_portfolio(self, portfolio: Dict[str, Any]):
         """Update portfolio after external execution"""
         self.portfolio.update(portfolio)
+
+    def reload_runtime_assets(self, active_skill_dirs: Optional[list] = None) -> None:
+        """Reload toolkit and system prompt from current run assets."""
+        from .toolkit_factory import create_agent_toolkit
+
+        clear_prompt_factory_cache()
+        toolkit_factory = self._toolkit_factory or create_agent_toolkit
+        toolkit_kwargs = dict(self._toolkit_factory_kwargs)
+        if active_skill_dirs is not None:
+            toolkit_kwargs["active_skill_dirs"] = active_skill_dirs
+
+        self.toolkit = toolkit_factory(
+            self.name,
+            self.config.get("config_name", "default"),
+            owner=self,
+            **toolkit_kwargs,
+        )
+        self.sys_prompt = build_agent_system_prompt(
+            agent_id=self.name,
+            config_name=self.config.get("config_name", "default"),
+            toolkit=self.toolkit,
+        )

backend/agents/prompt_factory.py

@@ -0,0 +1,112 @@
+# -*- coding: utf-8 -*-
+"""Assemble system prompts from base prompts, run assets, and toolkit context."""
+
+from pathlib import Path
+from typing import Any, Optional
+
+from backend.config.bootstrap_config import get_bootstrap_config_for_run
+from .prompt_loader import PromptLoader
+from .skills_manager import SkillsManager
+
+_prompt_loader = PromptLoader()
+
+
+def _read_file_if_exists(path: Path) -> str:
+    if not path.exists() or not path.is_file():
+        return ""
+    return path.read_text(encoding="utf-8").strip()
+
+
+def _append_section(parts: list[str], title: str, content: str) -> None:
+    content = content.strip()
+    if content:
+        parts.append(f"## {title}\n{content}")
+
+
+def build_agent_system_prompt(
+    agent_id: str,
+    config_name: str,
+    toolkit: Any,
+    analyst_type: Optional[str] = None,
+) -> str:
+    """Build the final system prompt for an agent."""
+    sections: list[str] = []
+
+    if analyst_type:
+        personas_config = _prompt_loader.load_yaml_config(
+            "analyst",
+            "personas",
+        )
+        persona = personas_config.get(analyst_type, {})
+        focus_text = "\n".join(
+            f"- {item}" for item in persona.get("focus", [])
+        )
+        description = persona.get("description", "").strip()
+        base_prompt = _prompt_loader.load_prompt(
+            "analyst",
+            "system",
+            variables={
+                "analyst_type": persona.get("name", analyst_type),
+                "focus": focus_text,
+                "description": description,
+            },
+        )
+    elif agent_id == "portfolio_manager":
+        base_prompt = _prompt_loader.load_prompt(
+            "portfolio_manager",
+            "system",
+        )
+    elif agent_id == "risk_manager":
+        base_prompt = _prompt_loader.load_prompt(
+            "risk_manager",
+            "system",
+        )
+    else:
+        raise ValueError(f"Unsupported agent prompt build for: {agent_id}")
+
+    sections.append(base_prompt.strip())
+
+    skills_manager = SkillsManager()
+    asset_dir = skills_manager.get_agent_asset_dir(config_name, agent_id)
+    asset_dir.mkdir(parents=True, exist_ok=True)
+    bootstrap_config = get_bootstrap_config_for_run(
+        skills_manager.project_root,
+        config_name,
+    )
+
+    _append_section(
+        sections,
+        "Bootstrap",
+        bootstrap_config.prompt_body,
+    )
+
+    _append_section(
+        sections,
+        "Role",
+        _read_file_if_exists(asset_dir / "ROLE.md"),
+    )
+    _append_section(
+        sections,
+        "Style",
+        _read_file_if_exists(asset_dir / "STYLE.md"),
+    )
+    _append_section(
+        sections,
+        "Policy",
+        _read_file_if_exists(asset_dir / "POLICY.md"),
+    )
+
+    skill_prompt = toolkit.get_agent_skill_prompt()
+    if skill_prompt:
+        _append_section(sections, "Skills", str(skill_prompt))
+
+    activated_notes = toolkit.get_activated_notes()
+    if activated_notes:
+        _append_section(sections, "Tool Usage Notes", str(activated_notes))
+
+    return "\n\n".join(section for section in sections if section.strip())
+
+
+def clear_prompt_factory_cache() -> None:
+    """Clear cached prompt and YAML templates before hot reload."""
+    _prompt_loader.clear_cache()

backend/agents/prompts/analyst/personas.yaml

@@ -8,12 +8,6 @@ fundamentals_analyst:
     - "管理层质量和公司治理"
     - "行业地位和市场份额"
     - "长期投资价值评估"
-  tools:
-    - "analyze_profitability"
-    - "analyze_growth"
-    - "analyze_financial_health"
-    - "analyze_valuation_ratios"
-    - "analyze_efficiency_ratios"
   description: |
     作为基本面分析师，你专注于：
     - 公司财务健康状况和盈利能力
@@ -39,11 +33,6 @@ technical_analyst:
     - 支撑/阻力位和关键价格点
     - 中短期交易机会
     你倾向于选择能够捕捉价格动态和市场趋势的工具，更偏好技术分析类工具。
-  tools:
-    - "analyze_trend_following"
-    - "analyze_momentum"
-    - "analyze_mean_reversion"
-    - "analyze_volatility"
 
 sentiment_analyst:
   name: "情绪分析师"
@@ -61,10 +50,6 @@ sentiment_analyst:
     - 投资者恐慌和贪婪情绪
     - 市场预期和心理因素
     你倾向于选择能够反映市场情绪和投资者行为的工具，更偏好情绪和行为类工具。
-  tools:
-    - "analyze_news_sentiment"
-    - "analyze_insider_trading"
-
 valuation_analyst:
   name: "估值分析师"
   focus:
@@ -81,12 +66,6 @@ valuation_analyst:
     - 相对估值和绝对估值
     - 投资安全边际评估
     你倾向于选择能够准确计算公司价值的工具，更偏好估值模型和基本面工具。
-  tools:
-    - "dcf_valuation_analysis"
-    - "owner_earnings_valuation_analysis"
-    - "ev_ebitda_valuation_analysis"
-    - "residual_income_valuation_analysis"
-
 comprehensive_analyst:
   name: "综合分析师"
   focus:
@@ -103,15 +82,3 @@ comprehensive_analyst:
     - 提供全面的投资建议
     - 适应不同市场环境
     你会根据具体情况灵活选择各类工具，追求分析的全面性和准确性。
-  tools:
-    - "analyze_profitability"
-    - "analyze_growth"
-    - "analyze_financial_health"
-    - "analyze_valuation_ratios"
-    - "analyze_efficiency_ratios"
-    - "analyze_trend_following"
-    - "analyze_momentum"
-    - "analyze_mean_reversion"
-    - "analyze_volatility"
-    - "analyze_news_sentiment"
-    - "analyze_insider_trading"

backend/agents/prompts/risk_manager/system.md

@@ -8,6 +8,8 @@
 5. 基于市场条件建议仓位限制
 
 你的决策流程：
+1. 优先使用可用的风险工具量化集中度、波动率和保证金压力
+2. 结合工具结果与当前市场上下文做判断
 3. 生成可操作的风险警告和仓位限制建议
 4. 为你的风险评估提供清晰的理由
 

backend/agents/risk_manager.py

@@ -11,9 +11,7 @@ from agentscope.message import Msg
 from agentscope.tool import Toolkit
 
 from ..utils.progress import progress
-from .prompt_loader import PromptLoader
-
-_prompt_loader = PromptLoader()
+from .prompt_factory import build_agent_system_prompt, clear_prompt_factory_cache
 
 
 class RiskAgent(ReActAgent):
@@ -29,6 +27,7 @@ class RiskAgent(ReActAgent):
         name: str = "risk_manager",
         config: Optional[Dict[str, Any]] = None,
         long_term_memory: Optional[LongTermMemoryBase] = None,
+        toolkit: Optional[Toolkit] = None,
     ):
         """
         Initialize Risk Manager Agent
@@ -41,12 +40,14 @@ class RiskAgent(ReActAgent):
             long_term_memory: Optional ReMeTaskLongTermMemory instance
         """
         self.config = config or {}
+        self.agent_id = name
+
+        if toolkit is None:
+            toolkit = Toolkit()
+        self.toolkit = toolkit
 
         sys_prompt = self._load_system_prompt()
 
-        # Create dedicated toolkit for this agent
-        toolkit = Toolkit()
-
         kwargs = {
             "name": name,
             "sys_prompt": sys_prompt,
@@ -64,9 +65,10 @@ class RiskAgent(ReActAgent):
 
     def _load_system_prompt(self) -> str:
         """Load system prompt for risk manager"""
-        return _prompt_loader.load_prompt(
-            "risk_manager",
-            "system",
+        return build_agent_system_prompt(
+            agent_id=self.agent_id,
+            config_name=self.config.get("config_name", "default"),
+            toolkit=self.toolkit,
         )
 
     async def reply(self, x: Msg = None) -> Msg:
@@ -86,3 +88,15 @@ class RiskAgent(ReActAgent):
         progress.update_status(self.name, None, "Risk assessment completed")
 
         return result
+
+    def reload_runtime_assets(self, active_skill_dirs: Optional[list] = None) -> None:
+        """Reload toolkit and system prompt from current run assets."""
+        from .toolkit_factory import create_agent_toolkit
+
+        clear_prompt_factory_cache()
+        self.toolkit = create_agent_toolkit(
+            self.agent_id,
+            self.config.get("config_name", "default"),
+            active_skill_dirs=active_skill_dirs,
+        )
+        self.sys_prompt = self._load_system_prompt()

backend/agents/skills_manager.py

@@ -0,0 +1,144 @@
+# -*- coding: utf-8 -*-
+"""Manage builtin/customized/active skill directories for each run."""
+
+from pathlib import Path
+import shutil
+from typing import Dict, Iterable, List
+
+import yaml
+
+from backend.config.bootstrap_config import get_bootstrap_config_for_run
+
+
+class SkillsManager:
+    """Sync named skills into a run-scoped active skills workspace."""
+
+    def __init__(self, project_root: Path | None = None):
+        self.project_root = (
+            project_root or Path(__file__).resolve().parents[2]
+        )
+        self.builtin_root = self.project_root / "backend" / "skills" / "builtin"
+        self.customized_root = (
+            self.project_root / "backend" / "skills" / "customized"
+        )
+        self.runs_root = self.project_root / "runs"
+
+    def get_active_root(self, config_name: str) -> Path:
+        return self.runs_root / config_name / "skills" / "active"
+
+    def get_activation_manifest_path(self, config_name: str) -> Path:
+        return self.runs_root / config_name / "skills" / "activation.yaml"
+
+    def get_agent_asset_dir(self, config_name: str, agent_id: str) -> Path:
+        return self.runs_root / config_name / "agents" / agent_id
+
+    def ensure_activation_manifest(self, config_name: str) -> Path:
+        manifest_path = self.get_activation_manifest_path(config_name)
+        manifest_path.parent.mkdir(parents=True, exist_ok=True)
+        if not manifest_path.exists():
+            manifest_path.write_text(
+                "global_enabled_skills: []\n"
+                "global_disabled_skills: []\n"
+                "agent_enabled_skills: {}\n"
+                "agent_disabled_skills: {}\n",
+                encoding="utf-8",
+            )
+        return manifest_path
+
+    def load_activation_manifest(self, config_name: str) -> Dict[str, object]:
+        manifest_path = self.ensure_activation_manifest(config_name)
+        with open(manifest_path, "r", encoding="utf-8") as file:
+            parsed = yaml.safe_load(file) or {}
+        return parsed if isinstance(parsed, dict) else {}
+
+    def _resolve_source_dir(self, skill_name: str) -> Path:
+        customized_dir = self.customized_root / skill_name
+        if customized_dir.exists():
+            return customized_dir
+
+        builtin_dir = self.builtin_root / skill_name
+        if builtin_dir.exists():
+            return builtin_dir
+
+        raise FileNotFoundError(f"Unknown skill: {skill_name}")
+
+    def resolve_agent_skill_names(
+        self,
+        config_name: str,
+        agent_id: str,
+        default_skills: Iterable[str],
+    ) -> List[str]:
+        """Resolve final skill names after bootstrap and activation overlays."""
+        bootstrap = get_bootstrap_config_for_run(self.project_root, config_name)
+        override = bootstrap.agent_override(agent_id)
+        skills = list(override.get("skills", list(default_skills)))
+
+        manifest = self.load_activation_manifest(config_name)
+        for skill_name in manifest.get("global_enabled_skills", []):
+            if skill_name not in skills:
+                skills.append(skill_name)
+
+        for skill_name in manifest.get("agent_enabled_skills", {}).get(agent_id, []):
+            if skill_name not in skills:
+                skills.append(skill_name)
+
+        disabled = set(manifest.get("global_disabled_skills", []))
+        disabled.update(
+            manifest.get("agent_disabled_skills", {}).get(agent_id, []),
+        )
+
+        return [skill for skill in skills if skill not in disabled]
+
+    def sync_active_skills(
+        self,
+        config_name: str,
+        skill_names: Iterable[str],
+    ) -> List[Path]:
+        """Sync selected skills into the run workspace and return their paths."""
+        active_root = self.get_active_root(config_name)
+        active_root.mkdir(parents=True, exist_ok=True)
+
+        synced_paths: List[Path] = []
+        wanted = set(skill_names)
+
+        for existing in active_root.iterdir():
+            if existing.is_dir() and existing.name not in wanted:
+                shutil.rmtree(existing)
+
+        for skill_name in skill_names:
+            source_dir = self._resolve_source_dir(skill_name)
+            target_dir = active_root / skill_name
+            if target_dir.exists():
+                shutil.rmtree(target_dir)
+            shutil.copytree(source_dir, target_dir)
+            synced_paths.append(target_dir)
+
+        return synced_paths
+
+    def prepare_active_skills(
+        self,
+        config_name: str,
+        agent_defaults: Dict[str, Iterable[str]],
+    ) -> Dict[str, List[Path]]:
+        """Resolve all agent skills, sync the union once, and map paths per agent."""
+        resolved: Dict[str, List[str]] = {}
+        union: List[str] = []
+
+        for agent_id, default_skills in agent_defaults.items():
+            resolved_skills = self.resolve_agent_skill_names(
+                config_name=config_name,
+                agent_id=agent_id,
+                default_skills=default_skills,
+            )
+            resolved[agent_id] = resolved_skills
+            for skill_name in resolved_skills:
+                if skill_name not in union:
+                    union.append(skill_name)
+
+        self.sync_active_skills(config_name=config_name, skill_names=union)
+        active_root = self.get_active_root(config_name)
+
+        return {
+            agent_id: [active_root / skill_name for skill_name in skill_names]
+            for agent_id, skill_names in resolved.items()
+        }

backend/agents/toolkit_factory.py

@@ -0,0 +1,197 @@
+# -*- coding: utf-8 -*-
+"""Toolkit factory following AgentScope's skill + tool group practices."""
+
+from typing import Any, Dict, Iterable
+
+from backend.config.bootstrap_config import get_bootstrap_config_for_run
+import yaml
+
+from .skills_manager import SkillsManager
+
+
+def load_agent_profiles() -> Dict[str, Dict[str, Any]]:
+    config_path = SkillsManager().project_root / "backend" / "config" / "agent_profiles.yaml"
+    with open(config_path, "r", encoding="utf-8") as file:
+        return yaml.safe_load(file) or {}
+
+
+def _register_analysis_tool_groups(toolkit: Any) -> None:
+    from backend.tools.analysis_tools import TOOL_REGISTRY
+
+    tool_groups = {
+        "fundamentals": {
+            "description": "Financial health, profitability, growth, and efficiency analysis tools.",
+            "active": False,
+            "notes": (
+                "Use these tools to validate business quality, financial resilience, "
+                "and earnings durability before making directional conclusions."
+            ),
+            "tools": [
+                "analyze_profitability",
+                "analyze_growth",
+                "analyze_financial_health",
+                "analyze_efficiency_ratios",
+                "analyze_valuation_ratios",
+                "get_financial_metrics_tool",
+            ],
+        },
+        "technical": {
+            "description": "Trend, momentum, mean reversion, and volatility analysis tools.",
+            "active": False,
+            "notes": (
+                "Use these tools to assess timing, price structure, and risk-reward in "
+                "the current market regime."
+            ),
+            "tools": [
+                "analyze_trend_following",
+                "analyze_momentum",
+                "analyze_mean_reversion",
+                "analyze_volatility",
+            ],
+        },
+        "sentiment": {
+            "description": "News sentiment and insider activity analysis tools.",
+            "active": False,
+            "notes": (
+                "Use these tools to capture short-horizon catalysts, sentiment shifts, "
+                "and behavioral signals around each ticker."
+            ),
+            "tools": [
+                "analyze_news_sentiment",
+                "analyze_insider_trading",
+            ],
+        },
+        "valuation": {
+            "description": "Intrinsic value and relative valuation analysis tools.",
+            "active": False,
+            "notes": (
+                "Use these tools when the task requires fair value estimation, margin of "
+                "safety analysis, or valuation scenario comparison."
+            ),
+            "tools": [
+                "dcf_valuation_analysis",
+                "owner_earnings_valuation_analysis",
+                "ev_ebitda_valuation_analysis",
+                "residual_income_valuation_analysis",
+            ],
+        },
+    }
+
+    for group_name, group_config in tool_groups.items():
+        toolkit.create_tool_group(
+            group_name=group_name,
+            description=group_config["description"],
+            active=group_config["active"],
+            notes=group_config["notes"],
+        )
+        for tool_name in group_config["tools"]:
+            tool_func = TOOL_REGISTRY.get(tool_name)
+            if tool_func:
+                toolkit.register_tool_function(
+                    tool_func,
+                    group_name=group_name,
+                )
+
+
+def _register_portfolio_tool_groups(toolkit: Any, pm_agent: Any) -> None:
+    toolkit.create_tool_group(
+        group_name="portfolio_ops",
+        description="Portfolio decision recording tools.",
+        active=False,
+        notes=(
+            "Use portfolio tools only after synthesizing analyst and risk inputs. "
+            "Record one explicit decision per ticker."
+        ),
+    )
+    toolkit.register_tool_function(
+        pm_agent._make_decision,
+        group_name="portfolio_ops",
+    )
+
+
+def _register_risk_tool_groups(toolkit: Any) -> None:
+    from backend.tools.risk_tools import (
+        assess_margin_and_liquidity,
+        assess_position_concentration,
+        assess_volatility_exposure,
+    )
+
+    toolkit.create_tool_group(
+        group_name="risk_ops",
+        description="Risk diagnostics for concentration, leverage, and volatility.",
+        active=False,
+        notes=(
+            "Use risk tools to quantify concentration, margin pressure, and volatility "
+            "before writing the final risk memo."
+        ),
+    )
+    toolkit.register_tool_function(
+        assess_position_concentration,
+        group_name="risk_ops",
+    )
+    toolkit.register_tool_function(
+        assess_margin_and_liquidity,
+        group_name="risk_ops",
+    )
+    toolkit.register_tool_function(
+        assess_volatility_exposure,
+        group_name="risk_ops",
+    )
+
+
+def create_agent_toolkit(
+    agent_id: str,
+    config_name: str,
+    owner: Any = None,
+    active_skill_dirs: Iterable[str] | None = None,
+) -> Any:
+    """Create a Toolkit with agent skills and grouped tools."""
+    from agentscope.tool import Toolkit
+
+    profiles = load_agent_profiles()
+    profile = profiles.get(agent_id, {})
+    skills_manager = SkillsManager()
+    bootstrap_config = get_bootstrap_config_for_run(
+        skills_manager.project_root,
+        config_name,
+    )
+    override = bootstrap_config.agent_override(agent_id)
+    active_groups = override.get(
+        "active_tool_groups",
+        profile.get("active_tool_groups", []),
+    )
+
+    toolkit = Toolkit(
+        agent_skill_instruction=(
+            "<system-info>You have access to project skills. Each skill lives in a "
+            "directory and is described by SKILL.md. Follow the skill instructions "
+            "when they are relevant to the current task.</system-info>"
+        ),
+        agent_skill_template="- {name} (dir: {dir}): {description}",
+    )
+
+    if agent_id.endswith("_analyst"):
+        _register_analysis_tool_groups(toolkit)
+    elif agent_id == "portfolio_manager" and owner is not None:
+        _register_portfolio_tool_groups(toolkit, owner)
+    elif agent_id == "risk_manager":
+        _register_risk_tool_groups(toolkit)
+
+    if active_skill_dirs is None:
+        skill_names = skills_manager.resolve_agent_skill_names(
+            config_name=config_name,
+            agent_id=agent_id,
+            default_skills=profile.get("skills", []),
+        )
+        active_skill_dirs = [
+            skills_manager.get_active_root(config_name) / skill_name
+            for skill_name in skill_names
+        ]
+
+    for skill_dir in active_skill_dirs:
+        toolkit.register_agent_skill(str(skill_dir))
+
+    if active_groups:
+        toolkit.update_tool_groups(group_names=active_groups, active=True)
+
+    return toolkit

backend/agents/workspace_manager.py

@@ -0,0 +1,140 @@
+# -*- coding: utf-8 -*-
+"""Initialize run-scoped agent workspace assets."""
+
+from pathlib import Path
+from typing import Dict, Iterable, Optional
+
+from .skills_manager import SkillsManager
+
+
+class WorkspaceManager:
+    """Create and maintain run-level prompt asset files for each agent."""
+
+    def __init__(self, project_root: Optional[Path] = None):
+        self.skills_manager = SkillsManager(project_root=project_root)
+        self.project_root = self.skills_manager.project_root
+
+    def get_run_dir(self, config_name: str) -> Path:
+        return self.project_root / "runs" / config_name
+
+    def ensure_run_workspace(self, config_name: str) -> Path:
+        run_dir = self.get_run_dir(config_name)
+        run_dir.mkdir(parents=True, exist_ok=True)
+        self.skills_manager.ensure_activation_manifest(config_name)
+        bootstrap_path = run_dir / "BOOTSTRAP.md"
+        if not bootstrap_path.exists():
+            bootstrap_path.write_text(
+                "---\n"
+                "tickers:\n"
+                "  - AAPL\n"
+                "  - MSFT\n"
+                "initial_cash: 100000\n"
+                "margin_requirement: 0.0\n"
+                "enable_memory: false\n"
+                "max_comm_cycles: 2\n"
+                "agent_overrides: {}\n"
+                "---\n\n"
+                "# Bootstrap\n\n"
+                "Use this file to describe run-specific setup notes, preferred tickers,\n"
+                "risk bounds, or strategy constraints before the first execution.\n\n"
+                "The YAML front matter above is machine-readable runtime configuration.\n"
+                "The markdown body below is injected into agent prompts as run context.\n",
+                encoding="utf-8",
+            )
+        return run_dir
+
+    def bootstrap_path(self, config_name: str) -> Path:
+        return self.get_run_dir(config_name) / "BOOTSTRAP.md"
+
+    def ensure_agent_assets(
+        self,
+        config_name: str,
+        agent_id: str,
+        role_seed: str = "",
+        style_seed: str = "",
+        policy_seed: str = "",
+    ) -> Path:
+        asset_dir = self.skills_manager.get_agent_asset_dir(
+            config_name,
+            agent_id,
+        )
+        asset_dir.mkdir(parents=True, exist_ok=True)
+
+        self._ensure_file(
+            asset_dir / "ROLE.md",
+            "# Role\n\n"
+            "Optional run-scoped role override.\n\n"
+            f"{role_seed}".strip()
+            + "\n",
+        )
+        self._ensure_file(
+            asset_dir / "STYLE.md",
+            "# Style\n\n"
+            "Optional run-scoped communication or reasoning style.\n\n"
+            f"{style_seed}".strip()
+            + "\n",
+        )
+        self._ensure_file(
+            asset_dir / "POLICY.md",
+            "# Policy\n\n"
+            "Optional run-scoped constraints, limits, or strategy policy.\n\n"
+            f"{policy_seed}".strip()
+            + "\n",
+        )
+        return asset_dir
+
+    def initialize_default_assets(
+        self,
+        config_name: str,
+        agent_ids: Iterable[str],
+        analyst_personas: Optional[Dict[str, Dict]] = None,
+    ) -> None:
+        self.ensure_run_workspace(config_name)
+        analyst_personas = analyst_personas or {}
+
+        for agent_id in agent_ids:
+            if agent_id.endswith("_analyst"):
+                persona = analyst_personas.get(agent_id, {})
+                role_seed = persona.get("description", "").strip()
+                focus_items = persona.get("focus", [])
+                style_seed = "\n".join(f"- {item}" for item in focus_items)
+                policy_seed = (
+                    "State a clear signal, confidence, and the conditions that would invalidate the thesis."
+                )
+            elif agent_id == "portfolio_manager":
+                role_seed = (
+                    "Synthesize analyst and risk inputs into explicit portfolio decisions."
+                )
+                style_seed = (
+                    "Be concise, capital-aware, and explicit about sizing rationale."
+                )
+                policy_seed = (
+                    "Respect cash, margin, and portfolio concentration constraints before recording decisions."
+                )
+            elif agent_id == "risk_manager":
+                role_seed = (
+                    "Quantify concentration, leverage, liquidity, and volatility risk before trade execution."
+                )
+                style_seed = (
+                    "Prioritize the highest-severity risk first and state concrete limits."
+                )
+                policy_seed = (
+                    "Use available risk tools before issuing the final risk memo."
+                )
+            else:
+                role_seed = ""
+                style_seed = ""
+                policy_seed = ""
+
+            self.ensure_agent_assets(
+                config_name=config_name,
+                agent_id=agent_id,
+                role_seed=role_seed,
+                style_seed=style_seed,
+                policy_seed=policy_seed,
+            )
+
+    @staticmethod
+    def _ensure_file(path: Path, content: str) -> None:
+        if not path.exists():
+            path.write_text(content, encoding="utf-8")