feat(subagents): support per-subagent skill loading and custom subagent types (#2253)

* feat(subagents): support per-subagent skill loading and custom subagent types (#2230)

Add per-subagent skill configuration and custom subagent type registration,
aligned with Codex's role-based config layering and per-session skill injection.

Backend:
- SubagentConfig gains `skills` field (None=all, []=none, list=whitelist)
- New CustomSubagentConfig for user-defined subagent types in config.yaml
- SubagentsAppConfig gains `custom_agents` section and `get_skills_for()`
- Registry resolves custom agents with three-layer config precedence
- SubagentExecutor loads skills per-session as conversation items (Codex pattern)
- task_tool no longer appends skills to system_prompt
- Lead agent system prompt dynamically lists all registered subagent types
- setup_agent tool accepts optional skills parameter
- Gateway agents API transparently passes skills in CRUD operations

Frontend:
- Agent/CreateAgentRequest/UpdateAgentRequest types include skills field
- Agent card displays skills as badges alongside tool_groups

Config:
- config.example.yaml documents custom_agents and per-agent skills override

Tests:
- 40 new tests covering all skill config, custom agents, and registry logic
- Existing tests updated for new get_skills_prompt_section signature

Closes #2230

* fix: address review feedback on skills PR

- Remove stale get_skills_prompt_section monkeypatches from test_task_tool_core_logic.py
  (task_tool no longer imports this function after skill injection moved to executor)
- Add key prefixes (tg:/sk:) to agent-card badges to prevent React key collisions
  between tool_groups and skills

* fix(ci): resolve lint and test failures

- Format agent-card.tsx with prettier (lint-frontend)
- Remove stale "Skills Appendix" system_prompt assertion — skills are now
  loaded per-session by SubagentExecutor, not appended to system_prompt

* fix(ci): sort imports in test_subagent_skills_config.py (ruff I001)

* fix(ci): use nullish coalescing in agent-card badge condition (eslint)

* fix: address review feedback on skills PR

- Use model_fields_set in AgentUpdateRequest to distinguish "field omitted"
  from "explicitly set to null" — fixes skills=None ambiguity where None
  means "inherit all" but was treated as "don't change"
- Move lazy import of get_subagent_config outside loop in
  _build_available_subagents_description to avoid repeated import overhead

---------

Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
This commit is contained in:
Xinmin Zeng
2026-04-23 23:59:47 +08:00
committed by GitHub
parent 4e72410154
commit 30d619de08
14 changed files with 962 additions and 72 deletions
@@ -164,6 +164,36 @@ Skip simple one-off tasks.
"""
def _build_available_subagents_description(available_names: list[str], bash_available: bool) -> str:
"""Dynamically build subagent type descriptions from registry.
Mirrors Codex's pattern where agent_type_description is dynamically generated
from all registered roles, so the LLM knows about every available type.
"""
# Built-in descriptions (kept for backward compatibility with existing prompt quality)
builtin_descriptions = {
"general-purpose": "For ANY non-trivial task - web research, code exploration, file operations, analysis, etc.",
"bash": (
"For command execution (git, build, test, deploy operations)" if bash_available else "Not available in the current sandbox configuration. Use direct file/web tools or switch to AioSandboxProvider for isolated shell access."
),
}
# Lazy import moved outside loop to avoid repeated import overhead
from deerflow.subagents.registry import get_subagent_config
lines = []
for name in available_names:
if name in builtin_descriptions:
lines.append(f"- **{name}**: {builtin_descriptions[name]}")
else:
config = get_subagent_config(name)
if config is not None:
desc = config.description.split("\n")[0].strip() # First line only for brevity
lines.append(f"- **{name}**: {desc}")
return "\n".join(lines)
def _build_subagent_section(max_concurrent: int) -> str:
"""Build the subagent system prompt section with dynamic concurrency limit.
@@ -174,13 +204,12 @@ def _build_subagent_section(max_concurrent: int) -> str:
Formatted subagent section string.
"""
n = max_concurrent
bash_available = "bash" in get_available_subagent_names()
available_subagents = (
"- **general-purpose**: For ANY non-trivial task - web research, code exploration, file operations, analysis, etc.\n- **bash**: For command execution (git, build, test, deploy operations)"
if bash_available
else "- **general-purpose**: For ANY non-trivial task - web research, code exploration, file operations, analysis, etc.\n"
"- **bash**: Not available in the current sandbox configuration. Use direct file/web tools or switch to AioSandboxProvider for isolated shell access."
)
available_names = get_available_subagent_names()
bash_available = "bash" in available_names
# Dynamically build subagent type descriptions from registry (aligned with Codex's
# agent_type_description pattern where all registered roles are listed in the tool spec).
available_subagents = _build_available_subagents_description(available_names, bash_available)
direct_tool_examples = "bash, ls, read_file, web_search, etc." if bash_available else "ls, read_file, web_search, etc."
direct_execution_example = (
'# User asks: "Run the tests"\n# Thinking: Cannot decompose into parallel sub-tasks\n# → Execute directly\n\nbash("npm test") # Direct execution, not task()'
@@ -25,6 +25,47 @@ class SubagentOverrideConfig(BaseModel):
min_length=1,
description="Model name for this subagent (None = inherit from parent agent)",
)
skills: list[str] | None = Field(
default=None,
description="Skill names whitelist for this subagent (None = inherit all enabled skills, [] = no skills)",
)
class CustomSubagentConfig(BaseModel):
"""User-defined subagent type declared in config.yaml."""
description: str = Field(
description="When the lead agent should delegate to this subagent",
)
system_prompt: str = Field(
description="System prompt that guides the subagent's behavior",
)
tools: list[str] | None = Field(
default=None,
description="Tool names whitelist (None = inherit all tools from parent)",
)
disallowed_tools: list[str] | None = Field(
default_factory=lambda: ["task", "ask_clarification", "present_files"],
description="Tool names to deny",
)
skills: list[str] | None = Field(
default=None,
description="Skill names whitelist (None = inherit all enabled skills, [] = no skills)",
)
model: str = Field(
default="inherit",
description="Model to use - 'inherit' uses parent's model",
)
max_turns: int = Field(
default=50,
ge=1,
description="Maximum number of agent turns before stopping",
)
timeout_seconds: int = Field(
default=900,
ge=1,
description="Maximum execution time in seconds",
)
class SubagentsAppConfig(BaseModel):
@@ -44,6 +85,10 @@ class SubagentsAppConfig(BaseModel):
default_factory=dict,
description="Per-agent configuration overrides keyed by agent name",
)
custom_agents: dict[str, CustomSubagentConfig] = Field(
default_factory=dict,
description="User-defined subagent types keyed by agent name",
)
def get_timeout_for(self, agent_name: str) -> int:
"""Get the effective timeout for a specific agent.
@@ -82,6 +127,20 @@ class SubagentsAppConfig(BaseModel):
return self.max_turns
return builtin_default
def get_skills_for(self, agent_name: str) -> list[str] | None:
"""Get the skills override for a specific agent.
Args:
agent_name: The name of the subagent.
Returns:
Skill names whitelist if overridden, None otherwise (subagent will inherit all enabled skills).
"""
override = self.agents.get(agent_name)
if override is not None and override.skills is not None:
return override.skills
return None
_subagents_config: SubagentsAppConfig = SubagentsAppConfig()
@@ -105,15 +164,20 @@ def load_subagents_config_from_dict(config_dict: dict) -> None:
parts.append(f"max_turns={override.max_turns}")
if override.model is not None:
parts.append(f"model={override.model}")
if override.skills is not None:
parts.append(f"skills={override.skills}")
if parts:
overrides_summary[name] = ", ".join(parts)
if overrides_summary:
custom_agents_names = list(_subagents_config.custom_agents.keys())
if overrides_summary or custom_agents_names:
logger.info(
"Subagents config loaded: default timeout=%ss, default max_turns=%s, per-agent overrides=%s",
"Subagents config loaded: default timeout=%ss, default max_turns=%s, per-agent overrides=%s, custom_agents=%s",
_subagents_config.timeout_seconds,
_subagents_config.max_turns,
overrides_summary,
overrides_summary or "none",
custom_agents_names or "none",
)
else:
logger.info(
@@ -13,6 +13,8 @@ class SubagentConfig:
system_prompt: The system prompt that guides the subagent's behavior.
tools: Optional list of tool names to allow. If None, inherits all tools.
disallowed_tools: Optional list of tool names to deny.
skills: Optional list of skill names to load. If None, inherits all enabled skills.
If an empty list, no skills are loaded.
model: Model to use - 'inherit' uses parent's model.
max_turns: Maximum number of agent turns before stopping.
timeout_seconds: Maximum execution time in seconds (default: 900 = 15 minutes).
@@ -23,6 +25,7 @@ class SubagentConfig:
system_prompt: str
tools: list[str] | None = None
disallowed_tools: list[str] | None = field(default_factory=lambda: ["task"])
skills: list[str] | None = None
model: str = "inherit"
max_turns: int = 50
timeout_seconds: int = 900
@@ -13,7 +13,7 @@ from typing import Any
from langchain.agents import create_agent
from langchain.tools import BaseTool
from langchain_core.messages import AIMessage, HumanMessage
from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
from langchain_core.runnables import RunnableConfig
from deerflow.agents.thread_state import SandboxState, ThreadDataState, ThreadState
@@ -184,7 +184,63 @@ class SubagentExecutor:
state_schema=ThreadState,
)
def _build_initial_state(self, task: str) -> dict[str, Any]:
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.
"""
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 []
try:
from deerflow.skills.loader import load_skills
# Use asyncio.to_thread to avoid blocking the event loop (LangGraph ASGI requirement)
all_skills = await asyncio.to_thread(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 []
if not all_skills:
logger.info(f"[trace={self.trace_id}] Subagent {self.config.name} no enabled skills found")
return []
# 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
if not skills:
return []
# Read each skill's SKILL.md content and create conversation items
messages = []
for skill in skills:
try:
content = await asyncio.to_thread(skill.skill_file.read_text, encoding="utf-8")
content = content.strip()
if content:
messages.append(SystemMessage(content=f'<skill name="{skill.name}">\n{content}\n</skill>'))
logger.info(f"[trace={self.trace_id}] Subagent {self.config.name} loaded skill: {skill.name}")
except Exception:
logger.debug(f"[trace={self.trace_id}] Failed to read skill {skill.name}", exc_info=True)
return messages
async def _build_initial_state(self, task: str) -> dict[str, Any]:
"""Build the initial state for agent execution.
Args:
@@ -193,8 +249,17 @@ class SubagentExecutor:
Returns:
Initial state dictionary.
"""
# Load skills as conversation items (Codex pattern)
skill_messages = await self._load_skill_messages()
messages: list = []
# Skill content injected as developer/system messages before the task
messages.extend(skill_messages)
# Then the actual task
messages.append(HumanMessage(content=task))
state: dict[str, Any] = {
"messages": [HumanMessage(content=task)],
"messages": messages,
}
# Pass through sandbox and thread data from parent
@@ -230,7 +295,7 @@ class SubagentExecutor:
try:
agent = self._create_agent()
state = self._build_initial_state(task)
state = await self._build_initial_state(task)
# Build config with thread_id for sandbox access and recursion limit
run_config: RunnableConfig = {
@@ -10,53 +10,100 @@ from deerflow.subagents.config import SubagentConfig
logger = logging.getLogger(__name__)
def _build_custom_subagent_config(name: str) -> SubagentConfig | None:
"""Build a SubagentConfig from config.yaml custom_agents section.
Args:
name: The name of the custom subagent.
Returns:
SubagentConfig if found in custom_agents, None otherwise.
"""
from deerflow.config.subagents_config import get_subagents_app_config
app_config = get_subagents_app_config()
custom = app_config.custom_agents.get(name)
if custom is None:
return None
return SubagentConfig(
name=name,
description=custom.description,
system_prompt=custom.system_prompt,
tools=custom.tools,
disallowed_tools=custom.disallowed_tools,
skills=custom.skills,
model=custom.model,
max_turns=custom.max_turns,
timeout_seconds=custom.timeout_seconds,
)
def get_subagent_config(name: str) -> SubagentConfig | None:
"""Get a subagent configuration by name, with config.yaml overrides applied.
Resolution order (mirrors Codex's config layering):
1. Built-in subagents (general-purpose, bash)
2. Custom subagents from config.yaml custom_agents section
3. Per-agent overrides from config.yaml agents section (timeout, max_turns, model, skills)
Args:
name: The name of the subagent.
Returns:
SubagentConfig if found (with any config.yaml overrides applied), None otherwise.
"""
# Step 1: Look up built-in, then fall back to custom_agents
config = BUILTIN_SUBAGENTS.get(name)
if config is None:
config = _build_custom_subagent_config(name)
if config is None:
return None
# Apply runtime overrides (timeout, max_turns, model) from config.yaml
# Step 2: Apply per-agent overrides from config.yaml agents section.
# Only explicit per-agent overrides are applied here. Global defaults
# (timeout_seconds, max_turns at the top level) apply to built-in agents
# but must NOT override custom agents' own values — custom agents define
# their own defaults in the custom_agents section.
# Lazy import to avoid circular deps.
from deerflow.config.subagents_config import get_subagents_app_config
app_config = get_subagents_app_config()
effective_timeout = app_config.get_timeout_for(name)
effective_max_turns = app_config.get_max_turns_for(name, config.max_turns)
is_builtin = name in BUILTIN_SUBAGENTS
agent_override = app_config.agents.get(name)
overrides = {}
if effective_timeout != config.timeout_seconds:
logger.debug(
"Subagent '%s': timeout overridden by config.yaml (%ss -> %ss)",
name,
config.timeout_seconds,
effective_timeout,
)
overrides["timeout_seconds"] = effective_timeout
if effective_max_turns != config.max_turns:
logger.debug(
"Subagent '%s': max_turns overridden by config.yaml (%s -> %s)",
name,
config.max_turns,
effective_max_turns,
)
overrides["max_turns"] = effective_max_turns
# Timeout: per-agent override > global default (builtins only) > config's own value
if agent_override is not None and agent_override.timeout_seconds is not None:
if agent_override.timeout_seconds != config.timeout_seconds:
logger.debug("Subagent '%s': timeout overridden (%ss -> %ss)", name, config.timeout_seconds, agent_override.timeout_seconds)
overrides["timeout_seconds"] = agent_override.timeout_seconds
elif is_builtin and app_config.timeout_seconds != config.timeout_seconds:
logger.debug("Subagent '%s': timeout from global default (%ss -> %ss)", name, config.timeout_seconds, app_config.timeout_seconds)
overrides["timeout_seconds"] = app_config.timeout_seconds
# Max turns: per-agent override > global default (builtins only) > config's own value
if agent_override is not None and agent_override.max_turns is not None:
if agent_override.max_turns != config.max_turns:
logger.debug("Subagent '%s': max_turns overridden (%s -> %s)", name, config.max_turns, agent_override.max_turns)
overrides["max_turns"] = agent_override.max_turns
elif is_builtin and app_config.max_turns is not None and app_config.max_turns != config.max_turns:
logger.debug("Subagent '%s': max_turns from global default (%s -> %s)", name, config.max_turns, app_config.max_turns)
overrides["max_turns"] = app_config.max_turns
# Model: per-agent override only (no global default for model)
effective_model = app_config.get_model_for(name)
if effective_model is not None and effective_model != config.model:
logger.debug(
"Subagent '%s': model overridden by config.yaml (%s -> %s)",
name,
config.model,
effective_model,
)
logger.debug("Subagent '%s': model overridden (%s -> %s)", name, config.model, effective_model)
overrides["model"] = effective_model
# Skills: per-agent override only (no global default for skills)
effective_skills = app_config.get_skills_for(name)
if effective_skills is not None and effective_skills != config.skills:
logger.debug("Subagent '%s': skills overridden (%s -> %s)", name, config.skills, effective_skills)
overrides["skills"] = effective_skills
if overrides:
config = replace(config, **overrides)
@@ -67,18 +114,33 @@ def list_subagents() -> list[SubagentConfig]:
"""List all available subagent configurations (with config.yaml overrides applied).
Returns:
List of all registered SubagentConfig instances.
List of all registered SubagentConfig instances (built-in + custom).
"""
return [get_subagent_config(name) for name in BUILTIN_SUBAGENTS]
configs = []
for name in get_subagent_names():
config = get_subagent_config(name)
if config is not None:
configs.append(config)
return configs
def get_subagent_names() -> list[str]:
"""Get all available subagent names.
"""Get all available subagent names (built-in + custom).
Returns:
List of subagent names.
"""
return list(BUILTIN_SUBAGENTS.keys())
names = list(BUILTIN_SUBAGENTS.keys())
# Merge custom_agents from config.yaml
from deerflow.config.subagents_config import get_subagents_app_config
app_config = get_subagents_app_config()
for custom_name in app_config.custom_agents:
if custom_name not in names:
names.append(custom_name)
return names
def get_available_subagent_names() -> list[str]:
@@ -87,11 +149,11 @@ def get_available_subagent_names() -> list[str]:
Returns:
List of subagent names visible to the current sandbox configuration.
"""
names = list(BUILTIN_SUBAGENTS.keys())
names = get_subagent_names()
try:
host_bash_allowed = is_host_bash_allowed()
except Exception:
logger.debug("Could not determine host bash availability; exposing all built-in subagents")
logger.debug("Could not determine host bash availability; exposing all subagents")
return names
if not host_bash_allowed:
@@ -17,12 +17,14 @@ def setup_agent(
soul: str,
description: str,
runtime: ToolRuntime,
skills: list[str] | None = None,
) -> Command:
"""Setup the custom DeerFlow agent.
Args:
soul: Full SOUL.md content defining the agent's personality and behavior.
description: One-line description of what the agent does.
skills: Optional list of skill names this agent should use. None means use all enabled skills, empty list means no skills.
"""
agent_name: str | None = runtime.context.get("agent_name") if runtime.context else None
@@ -41,6 +43,8 @@ def setup_agent(
config_data: dict = {"name": agent_name}
if description:
config_data["description"] = description
if skills is not None:
config_data["skills"] = skills
config_file = agent_dir / "config.yaml"
with open(config_file, "w", encoding="utf-8") as f:
@@ -10,7 +10,6 @@ from langchain.tools import InjectedToolCallId, ToolRuntime, tool
from langgraph.config import get_stream_writer
from langgraph.typing import ContextT
from deerflow.agents.lead_agent.prompt import get_skills_prompt_section
from deerflow.agents.thread_state import ThreadState
from deerflow.sandbox.security import LOCAL_BASH_SUBAGENT_DISABLED_MESSAGE, is_host_bash_allowed
from deerflow.subagents import SubagentExecutor, get_available_subagent_names, get_subagent_config
@@ -35,7 +34,7 @@ async def task_tool(
- Handle complex multi-step tasks autonomously
- Execute commands or operations in isolated contexts
Available subagent types depend on the active sandbox configuration:
Built-in subagent types:
- **general-purpose**: A capable agent for complex, multi-step tasks that require
both exploration and action. Use when the task requires complex reasoning,
multiple dependent steps, or would benefit from isolated context.
@@ -43,6 +42,11 @@ async def task_tool(
available when host bash is explicitly allowed or when using an isolated shell
sandbox such as `AioSandboxProvider`.
Additional custom subagent types may be defined in config.yaml under
`subagents.custom_agents`. Each custom type can have its own system prompt,
tools, skills, model, and timeout configuration. If an unknown subagent_type
is provided, the error message will list all available types.
When to use this tool:
- Complex tasks requiring multiple steps or tools
- Tasks that produce verbose output
@@ -72,9 +76,9 @@ async def task_tool(
# Build config overrides
overrides: dict = {}
skills_section = get_skills_prompt_section()
if skills_section:
overrides["system_prompt"] = config.system_prompt + "\n\n" + skills_section
# Skills are loaded by SubagentExecutor per-session (aligned with Codex's pattern:
# each subagent loads its own skills based on config, injected as conversation items).
# No longer appended to system_prompt here.
if max_turns is not None:
overrides["max_turns"] = max_turns