feat(agent): add custom-agent self-updates with user isolation (#2713)

* feat(agent): add update_agent tool for in-chat custom-agent self-updates (#2616)

Custom agents had no built-in way to persist updates to their own SOUL.md /
config.yaml from a normal chat — `setup_agent` was only bound during the
bootstrap flow, so when the user asked the agent to refine its description
or personality, the agent would shell out via bash/write_file and the edits
landed in a temporary sandbox/tool workspace instead of
`{base_dir}/agents/{agent_name}/`.

Changes:
- New `update_agent` builtin tool with partial-update semantics (only the
  fields you pass are written) and atomic temp-file + os.replace writes so
  a failed update never corrupts existing SOUL.md / config.yaml.
- Lead agent now binds `update_agent` in the non-bootstrap path whenever
  `agent_name` is set in the runtime context. Default agent (no
  agent_name) and bootstrap flow are unchanged.
- New `<self_update>` system-prompt section is injected for custom agents,
  instructing them to use `update_agent` — and explicitly NOT bash /
  write_file — to persist self-updates.
- Tests: 11 new cases in `tests/test_update_agent_tool.py` covering
  validation (missing/invalid agent_name, unknown agent, no fields),
  partial updates (soul-only, description-only, skills=[] vs omitted),
  no-op detection, atomic-write safety, and AgentConfig round-tripping;
  plus 2 new cases in `tests/test_lead_agent_prompt.py` covering the
  self-update prompt section.
- Docs: updated backend/CLAUDE.md builtin tools list and tools.mdx
  (en/zh) with the new tool description.

* feat(agent): isolate custom agents per user

Store custom agent definitions under the effective user, keep legacy agents readable until migration, and cover API/tool/migration behavior with tests.

Co-authored-by: Cursor <cursoragent@cursor.com>

* feat: consistent write/delete targets & add --user-id to migration

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
yangzheli
2026-05-05 23:17:42 +08:00
committed by GitHub
parent e8675f266d
commit 59c4a3f0a4
18 changed files with 955 additions and 60 deletions
@@ -318,7 +318,7 @@ def make_lead_agent(config: RunnableConfig):
def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
# Lazy import to avoid circular dependency
from deerflow.tools import get_available_tools
from deerflow.tools.builtins import setup_agent
from deerflow.tools.builtins import setup_agent, update_agent
cfg = _get_runtime_config(config)
resolved_app_config = app_config
@@ -390,6 +390,9 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
state_schema=ThreadState,
)
# Custom agents can update their own SOUL.md / config via update_agent.
# The default agent (no agent_name) does not see this tool.
extra_tools = [update_agent] if agent_name else []
# Default lead agent (unchanged behavior)
return create_agent(
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, reasoning_effort=reasoning_effort, app_config=resolved_app_config),
@@ -398,7 +401,8 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
groups=agent_config.tool_groups if agent_config else None,
subagent_enabled=subagent_enabled,
app_config=resolved_app_config,
),
)
+ extra_tools,
middleware=_build_middlewares(config, model_name=model_name, agent_name=agent_name, app_config=resolved_app_config),
system_prompt=apply_prompt_template(
subagent_enabled=subagent_enabled,
@@ -344,6 +344,7 @@ You are {agent_name}, an open-source super agent.
</role>
{soul}
{self_update_section}
{memory_context}
<thinking_style>
@@ -643,6 +644,26 @@ def get_agent_soul(agent_name: str | None) -> str:
return ""
def _build_self_update_section(agent_name: str | None) -> str:
"""Prompt block that teaches the custom agent to persist self-updates via update_agent."""
if not agent_name:
return ""
return f"""<self_update>
You are running as the custom agent **{agent_name}** with a persisted SOUL.md and config.yaml.
When the user asks you to update your own description, personality, behaviour, skill set, tool groups, or default model,
you MUST persist the change with the `update_agent` tool. Do NOT use `bash`, `write_file`, or any sandbox tool to edit
SOUL.md or config.yaml — those write into a temporary sandbox/tool workspace and the changes will be lost on the next turn.
Rules:
- Always pass the FULL replacement text for `soul` (no patch semantics). Start from your current SOUL above and apply the user's edits.
- Only pass the fields that should change. Omit the others to preserve them.
- Pass `skills=[]` to disable all skills, or omit `skills` to keep the existing whitelist.
- After `update_agent` returns successfully, tell the user the change is persisted and will take effect on the next turn.
</self_update>
"""
def get_deferred_tools_prompt_section(*, app_config: AppConfig | None = None) -> str:
"""Generate <available-deferred-tools> block for the system prompt.
@@ -772,6 +793,7 @@ def apply_prompt_template(
prompt = SYSTEM_PROMPT_TEMPLATE.format(
agent_name=agent_name or "DeerFlow 2.0",
soul=get_agent_soul(agent_name),
self_update_section=_build_self_update_section(agent_name),
skills_section=skills_section,
deferred_tools_section=deferred_tools_section,
memory_context=memory_context,
@@ -1,13 +1,22 @@
"""Configuration and loaders for custom agents."""
"""Configuration and loaders for custom agents.
Custom agents are stored per-user under ``{base_dir}/users/{user_id}/agents/{name}/``.
A legacy shared layout at ``{base_dir}/agents/{name}/`` is still readable so that
installations that pre-date user isolation continue to work until they run the
``scripts/migrate_user_isolation.py`` migration. New writes always target the
per-user layout.
"""
import logging
import re
from pathlib import Path
from typing import Any
import yaml
from pydantic import BaseModel
from deerflow.config.paths import get_paths
from deerflow.runtime.user_context import get_effective_user_id
logger = logging.getLogger(__name__)
@@ -40,14 +49,47 @@ class AgentConfig(BaseModel):
skills: list[str] | None = None
def load_agent_config(name: str | None) -> AgentConfig | None:
def resolve_agent_dir(name: str, *, user_id: str | None = None) -> Path:
"""Return the on-disk directory for an agent, preferring the per-user layout.
Resolution order:
1. ``{base_dir}/users/{user_id}/agents/{name}/`` (per-user, current layout).
2. ``{base_dir}/agents/{name}/`` (legacy shared layout — read-only fallback).
If neither exists, the per-user path is returned so callers that intend to
create the agent write into the new layout.
Args:
name: Validated agent name.
user_id: Owner of the agent. Defaults to the effective user from the
request context (or ``"default"`` in no-auth mode).
"""
paths = get_paths()
effective_user = user_id or get_effective_user_id()
user_path = paths.user_agent_dir(effective_user, name)
if user_path.exists():
return user_path
legacy_path = paths.agent_dir(name)
if legacy_path.exists():
return legacy_path
return user_path
def load_agent_config(name: str | None, *, user_id: str | None = None) -> AgentConfig | None:
"""Load the custom or default agent's config from its directory.
Reads from the per-user layout first; falls back to the legacy shared layout
for installations that have not yet been migrated.
Args:
name: The agent name.
user_id: Owner of the agent. Defaults to the effective user from the
current request context.
Returns:
AgentConfig instance.
AgentConfig instance, or ``None`` if ``name`` is ``None``.
Raises:
FileNotFoundError: If the agent directory or config.yaml does not exist.
@@ -58,7 +100,7 @@ def load_agent_config(name: str | None) -> AgentConfig | None:
return None
name = validate_agent_name(name)
agent_dir = get_paths().agent_dir(name)
agent_dir = resolve_agent_dir(name, user_id=user_id)
config_file = agent_dir / "config.yaml"
if not agent_dir.exists():
@@ -84,7 +126,7 @@ def load_agent_config(name: str | None) -> AgentConfig | None:
return AgentConfig(**data)
def load_agent_soul(agent_name: str | None) -> str | None:
def load_agent_soul(agent_name: str | None, *, user_id: str | None = None) -> str | None:
"""Read the SOUL.md file for a custom agent, if it exists.
SOUL.md defines the agent's personality, values, and behavioral guardrails.
@@ -92,11 +134,16 @@ def load_agent_soul(agent_name: str | None) -> str | None:
Args:
agent_name: The name of the agent or None for the default agent.
user_id: Owner of the agent. Defaults to the effective user from the
current request context.
Returns:
The SOUL.md content as a string, or None if the file does not exist.
"""
agent_dir = get_paths().agent_dir(agent_name) if agent_name else get_paths().base_dir
if agent_name:
agent_dir = resolve_agent_dir(agent_name, user_id=user_id)
else:
agent_dir = get_paths().base_dir
soul_path = agent_dir / SOUL_FILENAME
if not soul_path.exists():
return None
@@ -104,32 +151,50 @@ def load_agent_soul(agent_name: str | None) -> str | None:
return content or None
def list_custom_agents() -> list[AgentConfig]:
def list_custom_agents(*, user_id: str | None = None) -> list[AgentConfig]:
"""Scan the agents directory and return all valid custom agents.
Returns the union of agents in the per-user layout and the legacy shared
layout, so that pre-migration installations remain visible until they are
migrated. Per-user entries shadow legacy entries with the same name.
Args:
user_id: Owner whose agents to list. Defaults to the effective user
from the current request context.
Returns:
List of AgentConfig for each valid agent directory found.
"""
agents_dir = get_paths().agents_dir
if not agents_dir.exists():
return []
paths = get_paths()
effective_user = user_id or get_effective_user_id()
seen: set[str] = set()
agents: list[AgentConfig] = []
for entry in sorted(agents_dir.iterdir()):
if not entry.is_dir():
user_root = paths.user_agents_dir(effective_user)
legacy_root = paths.agents_dir
for root in (user_root, legacy_root):
if not root.exists():
continue
for entry in sorted(root.iterdir()):
if not entry.is_dir():
continue
if entry.name in seen:
continue
config_file = entry / "config.yaml"
if not config_file.exists():
logger.debug(f"Skipping {entry.name}: no config.yaml")
continue
config_file = entry / "config.yaml"
if not config_file.exists():
logger.debug(f"Skipping {entry.name}: no config.yaml")
continue
try:
agent_cfg = load_agent_config(entry.name)
agents.append(agent_cfg)
except Exception as e:
logger.warning(f"Skipping agent '{entry.name}': {e}")
try:
agent_cfg = load_agent_config(entry.name, user_id=effective_user)
if agent_cfg is None:
continue
agents.append(agent_cfg)
seen.add(entry.name)
except Exception as e:
logger.warning(f"Skipping agent '{entry.name}': {e}")
agents.sort(key=lambda a: a.name)
return agents
@@ -132,15 +132,20 @@ class Paths:
@property
def agents_dir(self) -> Path:
"""Root directory for all custom agents: `{base_dir}/agents/`."""
"""Legacy root for shared (pre user-isolation) custom agents: `{base_dir}/agents/`.
New code should use :meth:`user_agents_dir` instead. This property remains
only as a read-side fallback for installations that have not yet run the
``migrate_user_isolation.py`` script.
"""
return self.base_dir / "agents"
def agent_dir(self, name: str) -> Path:
"""Directory for a specific agent: `{base_dir}/agents/{name}/`."""
"""Legacy per-agent directory (no user isolation): `{base_dir}/agents/{name}/`."""
return self.agents_dir / name.lower()
def agent_memory_file(self, name: str) -> Path:
"""Per-agent memory file: `{base_dir}/agents/{name}/memory.json`."""
"""Legacy per-agent memory file: `{base_dir}/agents/{name}/memory.json`."""
return self.agent_dir(name) / "memory.json"
def user_dir(self, user_id: str) -> Path:
@@ -151,9 +156,17 @@ class Paths:
"""Per-user memory file: `{base_dir}/users/{user_id}/memory.json`."""
return self.user_dir(user_id) / "memory.json"
def user_agents_dir(self, user_id: str) -> Path:
"""Per-user root for that user's custom agents: `{base_dir}/users/{user_id}/agents/`."""
return self.user_dir(user_id) / "agents"
def user_agent_dir(self, user_id: str, agent_name: str) -> Path:
"""Per-user per-agent directory: `{base_dir}/users/{user_id}/agents/{name}/`."""
return self.user_agents_dir(user_id) / agent_name.lower()
def user_agent_memory_file(self, user_id: str, agent_name: str) -> Path:
"""Per-user per-agent memory: `{base_dir}/users/{user_id}/agents/{name}/memory.json`."""
return self.user_dir(user_id) / "agents" / agent_name.lower() / "memory.json"
return self.user_agent_dir(user_id, agent_name) / "memory.json"
def thread_dir(self, thread_id: str, *, user_id: str | None = None) -> Path:
"""
@@ -2,10 +2,12 @@ from .clarification_tool import ask_clarification_tool
from .present_file_tool import present_file_tool
from .setup_agent_tool import setup_agent
from .task_tool import task_tool
from .update_agent_tool import update_agent
from .view_image_tool import view_image_tool
__all__ = [
"setup_agent",
"update_agent",
"present_file_tool",
"ask_clarification_tool",
"view_image_tool",
@@ -8,6 +8,7 @@ from langgraph.types import Command
from deerflow.config.agents_config import validate_agent_name
from deerflow.config.paths import get_paths
from deerflow.runtime.user_context import get_effective_user_id
logger = logging.getLogger(__name__)
@@ -34,7 +35,14 @@ def setup_agent(
try:
agent_name = validate_agent_name(agent_name)
paths = get_paths()
agent_dir = paths.agent_dir(agent_name) if agent_name else paths.base_dir
if agent_name:
# Custom agents are persisted under the current user's bucket so
# different users do not see each other's agents.
user_id = get_effective_user_id()
agent_dir = paths.user_agent_dir(user_id, agent_name)
else:
# Default agent (no agent_name): SOUL.md lives at the global base dir.
agent_dir = paths.base_dir
is_new_dir = not agent_dir.exists()
agent_dir.mkdir(parents=True, exist_ok=True)
@@ -0,0 +1,241 @@
"""update_agent tool — let a custom agent persist updates to its own SOUL.md / config.
Bound to the lead agent only when ``runtime.context['agent_name']`` is set
(i.e. inside an existing custom agent's chat). The default agent does not see
this tool, and the bootstrap flow continues to use ``setup_agent`` for the
initial creation handshake.
The tool writes back to ``{base_dir}/users/{user_id}/agents/{agent_name}/{config.yaml,SOUL.md}``
so an agent created by one user is never visible to (or mutable by) another.
Writes are staged into temp files first; both files are renamed into place only
after both temp files are successfully written, so a partial failure cannot leave
config.yaml updated while SOUL.md still holds stale content.
"""
from __future__ import annotations
import logging
import tempfile
from pathlib import Path
from typing import Any
import yaml
from langchain_core.messages import ToolMessage
from langchain_core.tools import tool
from langgraph.prebuilt import ToolRuntime
from langgraph.types import Command
from deerflow.config.agents_config import load_agent_config, validate_agent_name
from deerflow.config.app_config import get_app_config
from deerflow.config.paths import get_paths
from deerflow.runtime.user_context import get_effective_user_id
logger = logging.getLogger(__name__)
def _stage_temp(path: Path, text: str) -> Path:
"""Write ``text`` into a sibling temp file and return its path.
The caller is responsible for ``Path.replace``-ing the temp into the target
once every staged file is ready, or for unlinking it on failure.
"""
path.parent.mkdir(parents=True, exist_ok=True)
fd = tempfile.NamedTemporaryFile(
mode="w",
dir=path.parent,
suffix=".tmp",
delete=False,
encoding="utf-8",
)
try:
fd.write(text)
fd.flush()
fd.close()
return Path(fd.name)
except BaseException:
fd.close()
Path(fd.name).unlink(missing_ok=True)
raise
def _cleanup_temps(temps: list[Path]) -> None:
"""Best-effort removal of staged temp files."""
for tmp in temps:
try:
tmp.unlink(missing_ok=True)
except OSError:
logger.debug("Failed to clean up temp file %s", tmp, exc_info=True)
@tool
def update_agent(
runtime: ToolRuntime,
soul: str | None = None,
description: str | None = None,
skills: list[str] | None = None,
tool_groups: list[str] | None = None,
model: str | None = None,
) -> Command:
"""Persist updates to the current custom agent's SOUL.md and config.yaml.
Use this when the user asks to refine the agent's identity, description,
skill whitelist, tool-group whitelist, or default model. Only the fields
you explicitly pass are updated; omitted fields keep their existing values.
Pass ``soul`` as the FULL replacement SOUL.md content — there is no patch
semantics, so always start from the current SOUL and apply your edits.
Pass ``skills=[]`` to disable all skills for this agent. Omit ``skills``
entirely to keep the existing whitelist.
Args:
soul: Optional full replacement SOUL.md content.
description: Optional new one-line description.
skills: Optional skill whitelist. ``[]`` = no skills, omit = unchanged.
tool_groups: Optional tool-group whitelist. ``[]`` = empty, omit = unchanged.
model: Optional model override (must match a configured model name).
Returns:
Command with a ToolMessage describing the result. Changes take effect
on the next user turn (when the lead agent is rebuilt with the fresh
SOUL.md and config.yaml).
"""
tool_call_id = runtime.tool_call_id
agent_name_raw: str | None = runtime.context.get("agent_name") if runtime.context else None
def _err(message: str) -> Command:
return Command(update={"messages": [ToolMessage(content=f"Error: {message}", tool_call_id=tool_call_id)]})
if soul is None and description is None and skills is None and tool_groups is None and model is None:
return _err("No fields provided. Pass at least one of: soul, description, skills, tool_groups, model.")
try:
agent_name = validate_agent_name(agent_name_raw)
except ValueError as e:
return _err(str(e))
if not agent_name:
return _err("update_agent is only available inside a custom agent's chat. There is no agent_name in the current runtime context, so there is nothing to update. If you are inside the bootstrap flow, use setup_agent instead.")
# Resolve the active user so that updates only affect this user's agent.
# ``get_effective_user_id`` returns DEFAULT_USER_ID when no auth context
# is set (matching how memory and thread storage behave).
user_id = get_effective_user_id()
# Reject an unknown ``model`` *before* touching the filesystem. Otherwise
# ``_resolve_model_name`` silently falls back to the default at runtime
# and the user sees confusing repeated warnings on every later turn.
if model is not None and get_app_config().get_model_config(model) is None:
return _err(f"Unknown model '{model}'. Pass a model name that exists in config.yaml's models section.")
paths = get_paths()
agent_dir = paths.user_agent_dir(user_id, agent_name)
if not agent_dir.exists() and paths.agent_dir(agent_name).exists():
return _err(f"Agent '{agent_name}' only exists in the legacy shared layout and is not scoped to a user. Run scripts/migrate_user_isolation.py to move legacy agents into the per-user layout before updating.")
try:
existing_cfg = load_agent_config(agent_name, user_id=user_id)
except FileNotFoundError:
return _err(f"Agent '{agent_name}' does not exist for the current user. Use setup_agent to create a new agent first.")
except ValueError as e:
return _err(f"Agent '{agent_name}' has an unreadable config: {e}")
if existing_cfg is None:
return _err(f"Agent '{agent_name}' could not be loaded.")
updated_fields: list[str] = []
# Force the on-disk ``name`` to match the directory we are writing into,
# even if ``existing_cfg.name`` had drifted (e.g. from manual yaml edits).
config_data: dict[str, Any] = {"name": agent_name}
new_description = description if description is not None else existing_cfg.description
config_data["description"] = new_description
if description is not None and description != existing_cfg.description:
updated_fields.append("description")
new_model = model if model is not None else existing_cfg.model
if new_model is not None:
config_data["model"] = new_model
if model is not None and model != existing_cfg.model:
updated_fields.append("model")
new_tool_groups = tool_groups if tool_groups is not None else existing_cfg.tool_groups
if new_tool_groups is not None:
config_data["tool_groups"] = new_tool_groups
if tool_groups is not None and tool_groups != existing_cfg.tool_groups:
updated_fields.append("tool_groups")
new_skills = skills if skills is not None else existing_cfg.skills
if new_skills is not None:
config_data["skills"] = new_skills
if skills is not None and skills != existing_cfg.skills:
updated_fields.append("skills")
config_changed = bool({"description", "model", "tool_groups", "skills"} & set(updated_fields))
# Stage every file we intend to rewrite into a temp sibling. Only after
# *all* temp files exist do we rename them into place — so a failure on
# SOUL.md cannot leave config.yaml already replaced.
pending: list[tuple[Path, Path]] = []
staged_temps: list[Path] = []
try:
agent_dir.mkdir(parents=True, exist_ok=True)
if config_changed:
yaml_text = yaml.dump(config_data, default_flow_style=False, allow_unicode=True, sort_keys=False)
config_target = agent_dir / "config.yaml"
config_tmp = _stage_temp(config_target, yaml_text)
staged_temps.append(config_tmp)
pending.append((config_tmp, config_target))
if soul is not None:
soul_target = agent_dir / "SOUL.md"
soul_tmp = _stage_temp(soul_target, soul)
staged_temps.append(soul_tmp)
pending.append((soul_tmp, soul_target))
updated_fields.append("soul")
# Commit phase. ``Path.replace`` is atomic per file on POSIX/NTFS and
# the staging step above means any earlier failure has already been
# reported. The remaining failure mode is a crash *between* two
# ``replace`` calls, which is reported via the partial-write error
# branch below so the caller knows which files are now on disk.
committed: list[Path] = []
try:
for tmp, target in pending:
tmp.replace(target)
committed.append(target)
except Exception as e:
_cleanup_temps([t for t, _ in pending if t not in committed])
if committed:
logger.error(
"[update_agent] Partial write for agent '%s' (user=%s): committed=%s, failed during rename: %s",
agent_name,
user_id,
[p.name for p in committed],
e,
exc_info=True,
)
return _err(f"Partial update for agent '{agent_name}': {[p.name for p in committed]} were updated, but the rest failed ({e}). Re-run update_agent to retry the remaining fields.")
raise
except Exception as e:
_cleanup_temps(staged_temps)
logger.error("[update_agent] Failed to update agent '%s' (user=%s): %s", agent_name, user_id, e, exc_info=True)
return _err(f"Failed to update agent '{agent_name}': {e}")
if not updated_fields:
return Command(update={"messages": [ToolMessage(content=f"No changes applied to agent '{agent_name}'. The provided values matched the existing config.", tool_call_id=tool_call_id)]})
logger.info("[update_agent] Updated agent '%s' (user=%s) fields: %s", agent_name, user_id, updated_fields)
return Command(
update={
"messages": [
ToolMessage(
content=(f"Agent '{agent_name}' updated successfully. Changed: {', '.join(updated_fields)}. The new configuration takes effect on the next user turn."),
tool_call_id=tool_call_id,
)
]
}
)