fix(skills): enforce allowed-tools metadata (#2626)

* fix(skills): parse allowed-tools frontmatter

* fix(skills): validate allowed-tools metadata

* fix(skills): add shared allowed-tools policy

* fix(subagents): enforce skill allowed-tools

* fix(agent): enforce skill allowed-tools

* refactor(skills): dedupe TypeVar and reuse cached enabled skills

- Drop redundant module-level TypeVar in tool_policy; rely on PEP 695 syntax.
- Expose get_cached_enabled_skills() and have the lead agent reuse it
  instead of synchronously rescanning skills on every request.

* fix(agent): expose config-scoped skill cache

* fix(subagents): pass filtered tools explicitly

* fix(skills): clean allowed-tools policy feedback
This commit is contained in:
AochenShen99
2026-05-07 08:34:43 +08:00
committed by GitHub
parent 2b0e62f679
commit cef4224381
12 changed files with 553 additions and 55 deletions
@@ -23,6 +23,8 @@ from deerflow.agents.thread_state import SandboxState, ThreadDataState, ThreadSt
from deerflow.config import get_app_config
from deerflow.config.app_config import AppConfig
from deerflow.models import create_chat_model
from deerflow.skills.tool_policy import filter_tools_by_skill_allowed_tools
from deerflow.skills.types import Skill
from deerflow.subagents.config import SubagentConfig, resolve_subagent_model_name
logger = logging.getLogger(__name__)
@@ -260,16 +262,16 @@ class SubagentExecutor:
# Generate trace_id if not provided (for top-level calls)
self.trace_id = trace_id or str(uuid.uuid4())[:8]
# Filter tools based on config
self.tools = _filter_tools(
self._base_tools = _filter_tools(
tools,
config.tools,
config.disallowed_tools,
)
self.tools = self._base_tools
logger.info(f"[trace={self.trace_id}] SubagentExecutor initialized: {config.name} with {len(self.tools)} tools")
def _create_agent(self):
def _create_agent(self, tools: list[BaseTool] | None = None):
"""Create the agent instance."""
app_config = self.app_config or get_app_config()
if self.model_name is None:
@@ -283,26 +285,14 @@ class SubagentExecutor:
return create_agent(
model=model,
tools=self.tools,
tools=tools if tools is not None else self.tools,
middleware=middlewares,
system_prompt=self.config.system_prompt,
state_schema=ThreadState,
)
async def _load_skill_messages(self) -> list[SystemMessage]:
"""Load skill content as conversation items based on config.skills.
Aligned with Codex's pattern: each subagent loads its own skills
per-session and injects them as conversation items (developer messages),
not as system prompt text. The config.skills whitelist controls which
skills are loaded:
- None: load all enabled skills
- []: no skills
- ["skill-a", "skill-b"]: only these skills
Returns:
List of SystemMessages containing skill content.
"""
async def _load_skills(self) -> list[Skill]:
"""Load enabled skill metadata based on config.skills."""
if self.config.skills is not None and len(self.config.skills) == 0:
logger.info(f"[trace={self.trace_id}] Subagent {self.config.name} skills=[] — skipping skill loading")
return []
@@ -316,8 +306,8 @@ class SubagentExecutor:
all_skills = await asyncio.to_thread(storage.load_skills, enabled_only=True)
logger.info(f"[trace={self.trace_id}] Subagent {self.config.name} loaded {len(all_skills)} enabled skills from disk")
except Exception:
logger.warning(f"[trace={self.trace_id}] Failed to load skills for subagent {self.config.name}", exc_info=True)
return []
logger.exception(f"[trace={self.trace_id}] Failed to load skills for subagent {self.config.name}")
raise
if not all_skills:
logger.info(f"[trace={self.trace_id}] Subagent {self.config.name} no enabled skills found")
@@ -326,10 +316,26 @@ class SubagentExecutor:
# Filter by config.skills whitelist
if self.config.skills is not None:
allowed = set(self.config.skills)
skills = [s for s in all_skills if s.name in allowed]
else:
skills = all_skills
return [s for s in all_skills if s.name in allowed]
return all_skills
def _apply_skill_allowed_tools(self, skills: list[Skill]) -> list[BaseTool]:
return filter_tools_by_skill_allowed_tools(self._base_tools, skills)
async def _load_skill_messages(self, skills: list[Skill]) -> list[SystemMessage]:
"""Load skill content as conversation items based on config.skills.
Aligned with Codex's pattern: each subagent loads its own skills
per-session and injects them as conversation items (developer messages),
not as system prompt text. The config.skills whitelist controls which
skills are loaded:
- None: load all enabled skills
- []: no skills
- ["skill-a", "skill-b"]: only these skills
Returns:
List of SystemMessages containing skill content.
"""
if not skills:
return []
@@ -347,19 +353,21 @@ class SubagentExecutor:
return messages
async def _build_initial_state(self, task: str) -> dict[str, Any]:
async def _build_initial_state(self, task: str) -> tuple[dict[str, Any], list[BaseTool]]:
"""Build the initial state for agent execution.
Args:
task: The task description.
Returns:
Initial state dictionary.
Initial state dictionary and tools filtered by loaded skill metadata.
"""
# Load skills as conversation items (Codex pattern)
skill_messages = await self._load_skill_messages()
skills = await self._load_skills()
filtered_tools = self._apply_skill_allowed_tools(skills)
skill_messages = await self._load_skill_messages(skills)
messages: list = []
messages: list[Any] = []
# Skill content injected as developer/system messages before the task
messages.extend(skill_messages)
# Then the actual task
@@ -375,7 +383,7 @@ class SubagentExecutor:
if self.thread_data is not None:
state["thread_data"] = self.thread_data
return state
return state, filtered_tools
async def _aexecute(self, task: str, result_holder: SubagentResult | None = None) -> SubagentResult:
"""Execute a task asynchronously.
@@ -405,8 +413,8 @@ class SubagentExecutor:
result.ai_messages = ai_messages
try:
agent = self._create_agent()
state = await self._build_initial_state(task)
state, filtered_tools = await self._build_initial_state(task)
agent = self._create_agent(filtered_tools)
# Build config with thread_id for sandbox access and recursion limit
run_config: RunnableConfig = {