feat(skills): 自我进化后端 — curator 配置可编辑 + 运行日志可读

后端骨架:
- curator.py / usage.py / skill_tools.py:基于使用计数的生命周期管理
  + LLM 合并/归档建议(同事已有原型,集中提交)
- skill_evolution_config.py:CuratorConfig 嵌套模型
- app.py:启动时按间隔触发 curator 后台线程
- tools.py:开启自我进化时一并注册 skill_list / skill_view 工具
- skill_manage_tool.py:agent 创建技能时标记 source=agent

本次新增:
- skill_evolution_runtime.py:skill_evolution.json 读写,原子写入
- app_config.py:把 overrides 深合并到 AppConfig.skill_evolution,
  并把 overrides 的 mtime 纳入缓存失效条件,前端 PUT 后热生效
- curator_log.py:进程内 ring buffer (maxlen=200),所有 record/get
  操作内部兜异常,保证 logging 永远不会让 curator 跑挂
- curator.py 关键节点插桩:run_started / 自动转换计数 / 候选数 /
  LLM agent invoke / 解析结果 / archive 动作 / run complete
- routers/skills.py:GET/PUT /api/curator/config(写到
  skill_evolution.json + reset_app_config)+ GET /api/curator/logs

测试:
- tests/test_skill_evolution_runtime.py:overrides 往返、损坏 JSON
  容错、AppConfig 合并、ring buffer 线程安全、record 异常吞咽

杂项:
- .gitignore:忽略 .deer-flow/* 与 skills/* 等运行时目录
- uv.lock:跟随依赖锁更新
- 清理已废弃的 agent runtime 数据 SOUL.md

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
xuhdandClaude Opus 4.7 committed 2026-05-14 12:00:21 +08:00
1 parent 230eec1baf
commit 9b6b94d919
14 files changed
+3981 -2698

No files matched your search

@@ -20,6 +20,8 @@ from deerflow.config.run_events_config import RunEventsConfig
from deerflow.config.runtime_paths import existing_project_file
from deerflow.config.sandbox_config import SandboxConfig
from deerflow.config.skill_evolution_config import SkillEvolutionConfig
from deerflow.config.skill_evolution_runtime import get_overrides_mtime as _get_skill_evolution_overrides_mtime
from deerflow.config.skill_evolution_runtime import load_overrides as _load_skill_evolution_overrides
from deerflow.config.skills_config import SkillsConfig
from deerflow.config.stream_bridge_config import StreamBridgeConfig, load_stream_bridge_config_from_dict
from deerflow.config.subagents_config import SubagentsAppConfig, load_subagents_config_from_dict
@@ -54,6 +56,22 @@ def _legacy_config_candidates() -> tuple[Path, ...]:
return (backend_dir / "config.yaml", repo_root / "config.yaml")
def _deep_merge_dict(base: dict[str, Any], overrides: dict[str, Any]) -> dict[str, Any]:
"""Deep-merge ``overrides`` onto a copy of ``base`` and return the result.
Nested dicts are merged recursively; any other value type in ``overrides``
replaces the corresponding value in ``base`` outright.
"""
merged: dict[str, Any] = dict(base)
for key, value in overrides.items():
existing = merged.get(key)
if isinstance(existing, dict) and isinstance(value, dict):
merged[key] = _deep_merge_dict(existing, value)
else:
merged[key] = value
return merged
def logging_level_from_config(name: str | None) -> int:
"""Map ``config.yaml`` ``log_level`` string to a :mod:`logging` level constant."""
mapping = logging.getLevelNamesMapping()
@@ -204,6 +222,15 @@ class AppConfig(BaseModel):
extensions_config = ExtensionsConfig.from_file()
config_data["extensions"] = extensions_config.model_dump()
# Apply skill_evolution runtime overrides from skill_evolution.json
# (file is optional; load_overrides returns None if missing/invalid)
overrides = _load_skill_evolution_overrides()
if overrides:
yaml_section = config_data.get("skill_evolution") or {}
if not isinstance(yaml_section, dict):
yaml_section = {}
config_data["skill_evolution"] = _deep_merge_dict(yaml_section, overrides)
result = cls.model_validate(config_data)
return result
@@ -329,6 +356,7 @@ class AppConfig(BaseModel):
_app_config: AppConfig | None = None
_app_config_path: Path | None = None
_app_config_mtime: float | None = None
_app_config_skill_evolution_overrides_mtime: float | None = None
_app_config_is_custom = False
_current_app_config: ContextVar[AppConfig | None] = ContextVar("deerflow_current_app_config", default=None)
_current_app_config_stack: ContextVar[tuple[AppConfig | None, ...]] = ContextVar("deerflow_current_app_config_stack", default=())
@@ -344,12 +372,13 @@ def _get_config_mtime(config_path: Path) -> float | None:
def _load_and_cache_app_config(config_path: str | None = None) -> AppConfig:
"""Load config from disk and refresh cache metadata."""
global _app_config, _app_config_path, _app_config_mtime, _app_config_is_custom
global _app_config, _app_config_path, _app_config_mtime, _app_config_skill_evolution_overrides_mtime, _app_config_is_custom
resolved_path = AppConfig.resolve_config_path(config_path)
_app_config = AppConfig.from_file(str(resolved_path))
_app_config_path = resolved_path
_app_config_mtime = _get_config_mtime(resolved_path)
_app_config_skill_evolution_overrides_mtime = _get_skill_evolution_overrides_mtime()
_app_config_is_custom = False
return _app_config
@@ -373,8 +402,14 @@ def get_app_config() -> AppConfig:
resolved_path = AppConfig.resolve_config_path()
current_mtime = _get_config_mtime(resolved_path)
current_overrides_mtime = _get_skill_evolution_overrides_mtime()
should_reload = _app_config is None or _app_config_path != resolved_path or _app_config_mtime != current_mtime
should_reload = (
_app_config is None
or _app_config_path != resolved_path
or _app_config_mtime != current_mtime
or _app_config_skill_evolution_overrides_mtime != current_overrides_mtime
)
if should_reload:
if _app_config_path == resolved_path and _app_config_mtime is not None and current_mtime is not None and _app_config_mtime != current_mtime:
logger.info(
@@ -382,6 +417,12 @@ def get_app_config() -> AppConfig:
_app_config_mtime,
current_mtime,
)
elif _app_config_skill_evolution_overrides_mtime != current_overrides_mtime:
logger.info(
"skill_evolution.json overrides changed (mtime: %s -> %s), reloading AppConfig",
_app_config_skill_evolution_overrides_mtime,
current_overrides_mtime,
)
_load_and_cache_app_config(str(resolved_path))
return _app_config
@@ -409,10 +450,11 @@ def reset_app_config() -> None:
`get_app_config()` to reload from file. Useful for testing
or when switching between different configurations.
"""
global _app_config, _app_config_path, _app_config_mtime, _app_config_is_custom
global _app_config, _app_config_path, _app_config_mtime, _app_config_skill_evolution_overrides_mtime, _app_config_is_custom
_app_config = None
_app_config_path = None
_app_config_mtime = None
_app_config_skill_evolution_overrides_mtime = None
_app_config_is_custom = False
@@ -424,10 +466,11 @@ def set_app_config(config: AppConfig) -> None:
Args:
config: The AppConfig instance to use.
"""
global _app_config, _app_config_path, _app_config_mtime, _app_config_is_custom
global _app_config, _app_config_path, _app_config_mtime, _app_config_skill_evolution_overrides_mtime, _app_config_is_custom
_app_config = config
_app_config_path = None
_app_config_mtime = None
_app_config_skill_evolution_overrides_mtime = None
_app_config_is_custom = True
@@ -1,6 +1,15 @@
from pydantic import BaseModel, Field
class CuratorConfig(BaseModel):
"""Configuration for the background skill curator."""
enabled: bool = Field(default=True, description="Whether the curator runs automatically.")
interval_hours: int = Field(default=168, description="Hours between curator runs (default: 7 days).")
stale_after_days: int = Field(default=30, description="Days of inactivity before a skill is marked stale.")
archive_after_days: int = Field(default=90, description="Days of inactivity before a stale skill is archived.")
class SkillEvolutionConfig(BaseModel):
"""Configuration for agent-managed skill evolution."""
@@ -12,3 +21,7 @@ class SkillEvolutionConfig(BaseModel):
default=None,
description="Optional model name for skill security moderation. Defaults to the primary chat model.",
)
curator: CuratorConfig = Field(
default_factory=CuratorConfig,
description="Configuration for the background skill curator.",
)
@@ -0,0 +1,115 @@
"""Runtime overrides for `skill_evolution` config, stored in skill_evolution.json.
The file format mirrors the ``skill_evolution`` YAML section — its content is
deep-merged onto the defaults loaded from ``config.yaml`` when ``AppConfig`` is
instantiated. The file is optional; absence (or a malformed payload) is treated
as "no overrides" so the system keeps the YAML-only behaviour.
"""
from __future__ import annotations
import json
import logging
import os
import tempfile
from pathlib import Path
from typing import Any
from deerflow.config.runtime_paths import existing_project_file
logger = logging.getLogger(__name__)
_FILENAME = "skill_evolution.json"
def resolve_path() -> Path:
"""Resolve the on-disk location of ``skill_evolution.json``.
Priority:
1. ``DEER_FLOW_SKILL_EVOLUTION_CONFIG_PATH`` environment variable
2. ``skill_evolution.json`` in the project root (current and parents)
3. Legacy ``backend/`` or repo-root sibling locations
4. Falls back to ``backend/skill_evolution.json`` (does not need to exist)
"""
override = os.getenv("DEER_FLOW_SKILL_EVOLUTION_CONFIG_PATH")
if override:
return Path(override)
project_file = existing_project_file((_FILENAME,))
if project_file is not None:
return project_file
backend_dir = Path(__file__).resolve().parents[4]
repo_root = backend_dir.parent
for candidate in (backend_dir / _FILENAME, repo_root / _FILENAME):
if candidate.exists():
return candidate
return backend_dir / _FILENAME
def load_overrides() -> dict[str, Any] | None:
"""Read overrides from disk.
Returns ``None`` when the file does not exist, is empty, contains
non-dict JSON, or fails to parse. Errors are logged at WARNING level
so production stays silent in the common "no overrides" case.
"""
path = resolve_path()
if not path.exists():
return None
try:
raw = path.read_text(encoding="utf-8").strip()
if not raw:
return None
data = json.loads(raw)
except OSError as exc:
logger.warning("Failed to read %s: %s", path, exc)
return None
except json.JSONDecodeError as exc:
logger.warning("Skill evolution overrides at %s are not valid JSON: %s", path, exc)
return None
if not isinstance(data, dict):
logger.warning("Skill evolution overrides at %s must be a JSON object, got %s", path, type(data).__name__)
return None
return data
def save_overrides(payload: dict[str, Any]) -> Path:
"""Atomically persist ``payload`` to ``skill_evolution.json``.
The caller is responsible for shape validation (e.g. via the
``SkillEvolutionConfig`` Pydantic model). Returns the path that was
written so callers can log it.
"""
if not isinstance(payload, dict):
raise TypeError("skill_evolution overrides payload must be a dict")
path = resolve_path()
path.parent.mkdir(parents=True, exist_ok=True)
fd, tmp_path_str = tempfile.mkstemp(prefix=".skill-evolution-", suffix=".tmp", dir=str(path.parent))
tmp_path = Path(tmp_path_str)
try:
with os.fdopen(fd, "w", encoding="utf-8") as f:
json.dump(payload, f, indent=2, ensure_ascii=False)
os.replace(tmp_path, path)
except Exception:
tmp_path.unlink(missing_ok=True)
raise
logger.info("Wrote skill_evolution overrides to %s", path)
return path
def get_overrides_mtime() -> float | None:
"""Return the mtime of ``skill_evolution.json`` if it exists, else None.
Used by the AppConfig cache to decide whether to reload — when the
file appears, disappears, or is modified, the mtime tuple changes
and ``get_app_config()`` rebuilds.
"""
path = resolve_path()
try:
return path.stat().st_mtime
except OSError:
return None
@@ -0,0 +1,440 @@
"""Skill curator — automatic lifecycle transitions and LLM consolidation pass."""
from __future__ import annotations
import json
import logging
import os
import re
import tempfile
import threading
from datetime import UTC, datetime, timedelta
from pathlib import Path
logger = logging.getLogger(__name__)
_CURATOR_STATE_FILENAME = ".curator_state"
def _get_skills_root() -> Path:
from deerflow.skills.storage import get_or_new_skill_storage
return Path(get_or_new_skill_storage().get_skills_root_path())
def _curator_state_path() -> Path:
return _get_skills_root() / _CURATOR_STATE_FILENAME
def _read_curator_state() -> dict:
path = _curator_state_path()
if not path.exists():
return {"run_count": 0, "paused": False}
try:
return json.loads(path.read_text(encoding="utf-8"))
except Exception:
return {"run_count": 0, "paused": False}
def _write_curator_state(state: dict) -> None:
path = _curator_state_path()
try:
path.parent.mkdir(parents=True, exist_ok=True)
tmp_fd, tmp_path = tempfile.mkstemp(dir=path.parent, suffix=".tmp")
try:
with os.fdopen(tmp_fd, "w", encoding="utf-8") as f:
json.dump(state, f, indent=2, default=str)
os.replace(tmp_path, path)
except Exception:
try:
os.unlink(tmp_path)
except Exception:
pass
raise
except Exception:
logger.debug("Failed to write curator state", exc_info=True)
def _get_curator_config():
try:
from deerflow.config import get_app_config
config = get_app_config()
skill_evo = getattr(config, "skill_evolution", None)
return getattr(skill_evo, "curator", None)
except Exception:
return None
def should_run_now() -> bool:
"""Return True if the curator interval has elapsed and it is not paused."""
curator_cfg = _get_curator_config()
if curator_cfg is None or not getattr(curator_cfg, "enabled", True):
return False
state = _read_curator_state()
if state.get("paused"):
return False
last_run = state.get("last_run_at")
if not last_run:
return True
interval_hours = getattr(curator_cfg, "interval_hours", 168)
try:
last_dt = datetime.fromisoformat(last_run)
return datetime.now(UTC) >= last_dt + timedelta(hours=interval_hours)
except Exception:
return True
def apply_automatic_transitions(now: datetime | None = None) -> dict:
"""Pure age-based state transitions — no LLM involved.
Only processes skills with source=='agent'. Skips pinned skills.
Returns counters: {marked_stale, archived, checked}.
"""
from deerflow.skills.usage import STATE_ACTIVE, STATE_ARCHIVED, STATE_STALE, agent_created_entries, latest_activity_at, set_state
if now is None:
now = datetime.now(UTC)
curator_cfg = _get_curator_config()
stale_days = getattr(curator_cfg, "stale_after_days", 30) if curator_cfg else 30
archive_days = getattr(curator_cfg, "archive_after_days", 90) if curator_cfg else 90
entries = agent_created_entries()
counters = {"marked_stale": 0, "archived": 0, "checked": len(entries)}
for name, entry in entries.items():
if entry.get("pinned"):
continue
last_activity_str = latest_activity_at(name)
if not last_activity_str:
continue
try:
last_dt = datetime.fromisoformat(last_activity_str)
if last_dt.tzinfo is None:
last_dt = last_dt.replace(tzinfo=UTC)
except Exception:
continue
days_idle = (now - last_dt).days
current_state = entry.get("state", STATE_ACTIVE)
if days_idle >= archive_days and current_state != STATE_ARCHIVED:
set_state(name, STATE_ARCHIVED)
counters["archived"] += 1
elif days_idle >= stale_days and current_state == STATE_ACTIVE:
set_state(name, STATE_STALE)
counters["marked_stale"] += 1
return counters
def _render_candidate_list() -> str:
from deerflow.skills.usage import STATE_ARCHIVED, all_entries
entries = {k: v for k, v in all_entries().items() if v.get("source") == "agent" and v.get("state") != STATE_ARCHIVED}
if not entries:
return "No agent-created active skills found."
lines = []
for name, entry in sorted(entries.items()):
state = entry.get("state", "active")
use = entry.get("use_count", 0)
view = entry.get("view_count", 0)
patch = entry.get("patch_count", 0)
last = entry.get("last_used_at") or entry.get("created_at") or "unknown"
lines.append(f"- {name} [state={state}, use={use}, view={view}, patches={patch}, last_active={last}]")
return "\n".join(lines)
_CURATOR_REVIEW_PROMPT = (
"You are running as DeerFlow's background skill CURATOR. This is an "
"UMBRELLA-BUILDING consolidation pass, not a passive audit and not a "
"duplicate-finder.\n\n"
"The goal of the skill collection is a LIBRARY OF CLASS-LEVEL "
"INSTRUCTIONS AND EXPERIENTIAL KNOWLEDGE. A collection of hundreds of "
"narrow skills where each one captures one session's specific bug is "
"a FAILURE of the library — not a feature. An agent searching skills "
"matches on descriptions, not on exact names; one broad umbrella "
"skill with labeled subsections beats five narrow siblings for "
"discoverability, not the other way around.\n\n"
"The right target shape is CLASS-LEVEL skills with rich SKILL.md "
"bodies + `references/`, `templates/`, and `scripts/` subfiles for "
"session-specific detail — not one-session-one-skill micro-entries.\n\n"
"Hard rules — do not violate:\n"
"1. DO NOT touch public (built-in) skills. The candidate list below "
"is already filtered to agent-created custom skills only.\n"
"2. DO NOT delete any skill. Archiving (moving the skill's directory "
"into the .archive/ folder under the skills root) is the maximum "
"destructive action. Archives are recoverable; deletion is not.\n"
"3. DO NOT touch skills shown as pinned=yes. Skip them entirely.\n"
"4. DO NOT use usage counters as a reason to skip consolidation. The "
"counters are new and often mostly zero. Judge overlap on CONTENT, "
"not on use_count. 'use=0' is not evidence a skill is valuable; it's "
"absence of evidence either way.\n"
"5. DO NOT reject consolidation on the grounds that 'each skill has "
"a distinct trigger'. Pairwise distinctness is the wrong bar. The "
"right bar is: 'would a human maintainer write this as N separate "
"skills, or as one skill with N labeled subsections?' When the "
"answer is the latter, merge.\n\n"
"How to work — not optional:\n"
"1. Scan the full candidate list. Identify PREFIX CLUSTERS (skills "
"sharing a first word or domain keyword). Examples you are likely "
"to find: search-*, code-*, data-*, api-*, file-*, analysis-*, "
"workflow-*, report-*, monitor-*, debug-*, python-*, security-*, etc. "
"Expect 10-25 clusters.\n"
"2. For each cluster with 2+ members, do NOT ask 'are these pairs "
"overlapping?' — ask 'what is the UMBRELLA CLASS these skills all "
"serve? Would a maintainer name that class and write one skill for "
"it?' If yes, pick (or create) the umbrella and absorb the siblings "
"into it.\n"
"3. Three ways to consolidate — use the right one per cluster:\n"
" a. MERGE INTO EXISTING UMBRELLA — one skill in the cluster is "
"already broad enough to be the umbrella. Patch it to add a labeled "
"section for each sibling's unique insight, then archive the "
"siblings.\n"
" b. CREATE A NEW UMBRELLA SKILL.md — no existing member is broad "
"enough. Create a new class-level skill whose SKILL.md covers the "
"shared workflow and has short labeled subsections. Archive the "
"now-absorbed narrow siblings.\n"
" c. DEMOTE TO REFERENCES/TEMPLATES/SCRIPTS — a sibling has "
"narrow-but-valuable session-specific content. Move it into the "
"umbrella's appropriate support directory:\n"
" • `references/<topic>.md` for session-specific detail OR "
"condensed knowledge banks (quoted research, API docs excerpts, "
"domain notes, provider quirks, reproduction recipes)\n"
" • `templates/<name>.<ext>` for starter files meant to be "
"copied and modified\n"
" • `scripts/<name>.<ext>` for statically re-runnable actions "
"(verification scripts, fixture generators, probes)\n"
" Then archive the old sibling.\n"
"4. Also flag skills whose NAME is too narrow (contains a specific "
"issue number, a feature codename, a specific error string, an "
"'audit' / 'diagnosis' / 'salvage' session artifact). These almost "
"always belong as a subsection or support file under a class-level "
"umbrella.\n"
"5. Iterate. After one consolidation round, scan the remaining set "
"and look for the NEXT umbrella opportunity. Don't stop after 3 "
"merges.\n\n"
"'keep' is a legitimate decision ONLY when the skill is already a "
"class-level umbrella and none of the proposed merges would improve "
"discoverability. 'This is narrow but distinct from its siblings' "
"is NOT a reason to keep — it's a reason to move it under an "
"umbrella as a subsection or support file.\n\n"
"Expected output: real umbrella-ification. Process every obvious "
"cluster. If you end the pass with fewer than 10 archives, you "
"stopped too early — go back and look at the clusters you left "
"alone.\n\n"
"When done, write a human summary AND a structured machine-readable "
"block so downstream tooling can distinguish consolidation from "
"pruning. Format EXACTLY:\n\n"
"## Structured summary (required)\n"
"```yaml\n"
"consolidations:\n"
" - from: <old-skill-name>\n"
" into: <umbrella-skill-name>\n"
" reason: <one short sentence — why merged, not just 'similar'>\n"
"prunings:\n"
" - name: <skill-name>\n"
" reason: <one short sentence — why archived with no merge target>\n"
"```\n\n"
"Every skill you moved to .archive/ MUST appear in exactly one of the "
"two lists. If you consolidated X into umbrella Y (patched Y, wrote "
"a references file to Y, or created Y with X's content absorbed), X "
"goes under `consolidations` with `into: Y`. If you archived X with "
"no absorption — truly stale, irrelevant, or obsolete — X goes under "
"`prunings`. Leave a list empty (`consolidations: []`) if none. Do "
"not omit the block. The block comes AFTER your human-readable "
"summary of clusters processed, patches made, and decisions left alone."
)
def _make_skill_archive_tool():
"""Curator-only tool: move a skill directory to .archive/ (recoverable)."""
from langchain.tools import tool as lc_tool
from deerflow.skills import curator_log
from deerflow.skills.storage import get_or_new_skill_storage
@lc_tool("skill_archive")
def _skill_archive(name: str) -> str:
"""Move an agent-created skill to the .archive/ directory (recoverable). Use this instead of delete.
Args:
name: The skill name to archive.
"""
from deerflow.skills.usage import STATE_ARCHIVED, get_entry, set_state
storage = get_or_new_skill_storage()
if not storage.custom_skill_exists(name):
curator_log.record("warning", f"skill_archive: '{name}' not found or already archived")
return f"Skill '{name}' not found or already archived."
if get_entry(name).get("pinned"):
curator_log.record("info", f"skill_archive: '{name}' is pinned — skipping")
return f"Skill '{name}' is pinned — skipping."
skill_dir = storage.get_custom_skill_dir(name)
archive_root = Path(storage.get_skills_root_path()) / "custom" / ".archive"
archive_root.mkdir(parents=True, exist_ok=True)
stamp = datetime.now(UTC).strftime("%Y%m%d-%H%M%S")
dest = archive_root / f"{name}-{stamp}"
try:
skill_dir.rename(dest)
except Exception as e:
curator_log.record("error", f"skill_archive: archive failed for '{name}': {e}")
return f"Archive failed: {e}"
set_state(name, STATE_ARCHIVED)
curator_log.record("info", f"Archived '{name}' → custom/.archive/{dest.name}")
return f"Archived '{name}' → custom/.archive/{dest.name}."
return _skill_archive
def _run_llm_review(candidate_list: str) -> str:
"""Run the curator consolidation pass as a React agent with skill tools bound."""
from deerflow.skills import curator_log
try:
from langchain_core.messages import HumanMessage
from langgraph.prebuilt import create_react_agent
from deerflow.models import create_chat_model
from deerflow.tools.builtins.skill_tools import skill_list_tool, skill_view_tool
from deerflow.tools.skill_manage_tool import skill_manage_tool
model = create_chat_model()
tools = [skill_list_tool, skill_view_tool, skill_manage_tool, _make_skill_archive_tool()]
agent = create_react_agent(model, tools)
curator_log.record("info", "LLM review: invoking React agent")
prompt = _CURATOR_REVIEW_PROMPT + f"\n\nSkill candidates for this pass:\n{candidate_list}"
result = agent.invoke({"messages": [HumanMessage(content=prompt)]})
messages = result.get("messages", [])
for msg in reversed(messages):
if not getattr(msg, "tool_calls", None):
content = getattr(msg, "content", "")
if isinstance(content, str) and content.strip():
curator_log.record("info", "LLM review: agent returned final answer")
return content
if isinstance(content, list):
text = " ".join(
block.get("text", "") if isinstance(block, dict) else str(block)
for block in content
).strip()
curator_log.record("info", "LLM review: agent returned final answer")
return text
curator_log.record("warning", "LLM review: no final assistant message produced")
return ""
except Exception as e:
logger.warning("Curator LLM review failed: %s", e)
curator_log.record("warning", f"LLM review failed: {e}")
return ""
def _parse_structured_summary(text: str) -> dict:
pattern = re.compile(r"```yaml\s*(.*?)```", re.DOTALL)
match = pattern.search(text)
if not match:
return {"consolidations": [], "prunings": []}
try:
import yaml
data = yaml.safe_load(match.group(1)) or {}
return {"consolidations": data.get("consolidations") or [], "prunings": data.get("prunings") or []}
except Exception:
return {"consolidations": [], "prunings": []}
def run_curator_review() -> dict:
"""Main curator run: automatic transitions → LLM consolidation → write state."""
from deerflow.skills import curator_log
curator_log.start_new_run()
now = datetime.now(UTC)
logger.info("Curator: starting run at %s", now.isoformat())
curator_log.record("info", f"Curator starting run at {now.isoformat()}")
transition_counters = apply_automatic_transitions(now)
logger.info("Curator: transitions=%s", transition_counters)
curator_log.record(
"info",
(
f"Age-based transitions: checked={transition_counters['checked']}, "
f"marked_stale={transition_counters['marked_stale']}, "
f"archived={transition_counters['archived']}"
),
)
candidate_list = _render_candidate_list()
candidate_line_count = 0 if not candidate_list.strip() else len([line for line in candidate_list.splitlines() if line.startswith("- ")])
curator_log.record("info", f"Candidates for LLM review: {candidate_line_count} skill(s)")
llm_text = _run_llm_review(candidate_list)
suggestions = _parse_structured_summary(llm_text)
curator_log.record(
"info",
(
f"LLM suggestions parsed: consolidations={len(suggestions['consolidations'])}, "
f"prunings={len(suggestions['prunings'])}"
),
)
state = _read_curator_state()
state["last_run_at"] = now.isoformat()
state["run_count"] = state.get("run_count", 0) + 1
state["last_run_summary"] = (
f"auto: {transition_counters['marked_stale']} marked stale, "
f"{transition_counters['archived']} archived; "
f"llm: {len(suggestions['consolidations'])} consolidations suggested"
)
state["last_consolidation_suggestions"] = suggestions
_write_curator_state(state)
logger.info("Curator: run complete — %s", state["last_run_summary"])
curator_log.record("info", f"Curator run complete — {state['last_run_summary']}")
return {"transitions": transition_counters, "suggestions": suggestions}
def _curator_thread_target() -> None:
from deerflow.skills import curator_log
try:
run_curator_review()
except Exception as exc:
logger.exception("Curator background run failed")
curator_log.record("error", f"Curator background run failed: {exc}")
async def maybe_run_curator() -> None:
"""Async entry called at startup — spawns a daemon thread if the interval has elapsed."""
if not should_run_now():
return
logger.info("Curator: interval elapsed, spawning background thread")
thread = threading.Thread(target=_curator_thread_target, daemon=True, name="skill-curator")
thread.start()
def get_curator_status() -> dict:
"""Return curator state for the API status endpoint."""
from deerflow.skills.usage import STATE_ACTIVE, STATE_ARCHIVED, STATE_STALE, agent_created_entries
state = _read_curator_state()
entries = agent_created_entries()
counts: dict[str, int] = {STATE_ACTIVE: 0, STATE_STALE: 0, STATE_ARCHIVED: 0}
for entry in entries.values():
s = entry.get("state", STATE_ACTIVE)
counts[s] = counts.get(s, 0) + 1
return {
"enabled": _get_curator_config() is not None and getattr(_get_curator_config(), "enabled", True),
"run_count": state.get("run_count", 0),
"last_run_at": state.get("last_run_at"),
"last_run_summary": state.get("last_run_summary"),
"paused": state.get("paused", False),
"agent_created_skills": {
"total": len(entries),
"active": counts[STATE_ACTIVE],
"stale": counts[STATE_STALE],
"archived": counts[STATE_ARCHIVED],
},
}
@@ -0,0 +1,82 @@
"""In-memory ring buffer holding the most recent curator run's execution log.
The buffer is intentionally process-local and non-persistent — restarting the
server clears it. The Gateway exposes the buffer over ``GET /api/curator/logs``
so the frontend can show what the latest curator pass actually did.
All public functions swallow internal exceptions so a logging failure can never
break the curator run itself.
"""
from __future__ import annotations
import logging
import threading
from collections import deque
from datetime import UTC, datetime
from typing import Any
logger = logging.getLogger(__name__)
_MAX_RECORDS = 200
_lock = threading.Lock()
_buffer: deque[dict[str, Any]] = deque(maxlen=_MAX_RECORDS)
def _now_iso() -> str:
return datetime.now(UTC).isoformat()
def start_new_run() -> None:
"""Clear the buffer and emit a ``run_started`` record.
Called once at the top of each curator run so the frontend sees a fresh
log stream rather than the previous run's tail.
"""
try:
with _lock:
_buffer.clear()
_buffer.append({"ts": _now_iso(), "level": "info", "message": "run_started"})
except Exception:
logger.debug("curator_log.start_new_run failed", exc_info=True)
def record(level: str, message: str, **fields: Any) -> None:
"""Append a single log record. Never raises.
``level`` is one of ``info``/``warning``/``error``; unknown values are
accepted and rendered verbatim by the frontend. Extra keyword arguments
are merged into the record so callers can attach structured context.
"""
try:
entry: dict[str, Any] = {
"ts": _now_iso(),
"level": str(level or "info"),
"message": str(message),
}
if fields:
entry.update({k: v for k, v in fields.items() if k not in entry})
with _lock:
_buffer.append(entry)
except Exception:
logger.debug("curator_log.record failed", exc_info=True)
def get_recent() -> list[dict[str, Any]]:
"""Return a snapshot copy of the buffer contents. Never raises."""
try:
with _lock:
return list(_buffer)
except Exception:
logger.debug("curator_log.get_recent failed", exc_info=True)
return []
def clear() -> None:
"""Drop all buffered records. Used by tests."""
try:
with _lock:
_buffer.clear()
except Exception:
logger.debug("curator_log.clear failed", exc_info=True)
@@ -0,0 +1,159 @@
"""Skill usage sidecar — per-skill telemetry stored in .usage.json."""
from __future__ import annotations
import json
import logging
import os
import tempfile
from datetime import UTC, datetime
from pathlib import Path
logger = logging.getLogger(__name__)
_USAGE_FILENAME = ".usage.json"
STATE_ACTIVE = "active"
STATE_STALE = "stale"
STATE_ARCHIVED = "archived"
SOURCE_AGENT = "agent"
SOURCE_UPLOAD = "upload"
_DEFAULT_ENTRY: dict = {
"use_count": 0,
"view_count": 0,
"patch_count": 0,
"last_used_at": None,
"last_viewed_at": None,
"last_patched_at": None,
"created_at": None,
"state": STATE_ACTIVE,
"pinned": False,
"archived_at": None,
"source": SOURCE_UPLOAD,
}
def _usage_path() -> Path:
from deerflow.skills.storage import get_or_new_skill_storage
return Path(get_or_new_skill_storage().get_skills_root_path()) / _USAGE_FILENAME
def _read_all() -> dict:
path = _usage_path()
if not path.exists():
return {}
try:
return json.loads(path.read_text(encoding="utf-8"))
except Exception:
logger.debug("Failed to read %s", path, exc_info=True)
return {}
def _write_all(data: dict) -> None:
path = _usage_path()
try:
path.parent.mkdir(parents=True, exist_ok=True)
tmp_fd, tmp_path = tempfile.mkstemp(dir=path.parent, suffix=".tmp")
try:
with os.fdopen(tmp_fd, "w", encoding="utf-8") as f:
json.dump(data, f, indent=2, default=str)
os.replace(tmp_path, path)
except Exception:
try:
os.unlink(tmp_path)
except Exception:
pass
raise
except Exception:
logger.debug("Failed to write %s", path, exc_info=True)
def _now_iso() -> str:
return datetime.now(UTC).isoformat()
def _update_entry(name: str, updates: dict) -> None:
try:
data = _read_all()
entry = {**_DEFAULT_ENTRY, **data.get(name, {})}
entry.update(updates)
data[name] = entry
_write_all(data)
except Exception:
logger.debug("Failed to update usage entry for %r", name, exc_info=True)
def get_entry(name: str) -> dict:
"""Return the usage entry for *name*, filling missing fields with defaults."""
data = _read_all()
return {**_DEFAULT_ENTRY, **data.get(name, {})}
def all_entries() -> dict:
"""Return all usage entries keyed by skill name."""
return _read_all()
def agent_created_entries() -> dict:
"""Return only entries where source == 'agent' (eligible for curation)."""
return {k: v for k, v in _read_all().items() if v.get("source") == SOURCE_AGENT}
def bump_view(name: str) -> None:
entry = get_entry(name)
_update_entry(name, {"view_count": entry.get("view_count", 0) + 1, "last_viewed_at": _now_iso()})
def bump_use(name: str) -> None:
entry = get_entry(name)
_update_entry(name, {"use_count": entry.get("use_count", 0) + 1, "last_used_at": _now_iso()})
def bump_patch(name: str) -> None:
entry = get_entry(name)
_update_entry(name, {"patch_count": entry.get("patch_count", 0) + 1, "last_patched_at": _now_iso()})
def set_state(name: str, state: str) -> None:
updates: dict = {"state": state}
if state == STATE_ARCHIVED:
updates["archived_at"] = _now_iso()
elif state == STATE_ACTIVE:
updates["archived_at"] = None
_update_entry(name, updates)
def set_pinned(name: str, pinned: bool) -> None:
_update_entry(name, {"pinned": pinned})
def forget(name: str) -> None:
try:
data = _read_all()
data.pop(name, None)
_write_all(data)
except Exception:
logger.debug("Failed to forget usage entry for %r", name, exc_info=True)
def mark_agent_created(name: str) -> None:
"""Record a skill as agent-created; sets source='agent' and initialises created_at."""
_update_entry(name, {"source": SOURCE_AGENT, "created_at": _now_iso(), "state": STATE_ACTIVE})
def mark_uploaded(name: str) -> None:
"""Record a skill as user-uploaded; sets source='upload'."""
_update_entry(name, {"source": SOURCE_UPLOAD, "created_at": _now_iso()})
def latest_activity_at(name: str) -> str | None:
"""Return the most recent activity timestamp, falling back to created_at."""
entry = get_entry(name)
candidates = [entry.get("last_used_at"), entry.get("last_viewed_at"), entry.get("last_patched_at")]
valid = [ts for ts in candidates if ts]
if valid:
return max(valid)
return entry.get("created_at")
@@ -0,0 +1,83 @@
"""skill_view and skill_list — read-only tools for agent skill introspection."""
from __future__ import annotations
import json
import logging
from langchain.tools import tool
logger = logging.getLogger(__name__)
@tool("skill_view")
def skill_view_tool(name: str) -> str:
"""Read the full SKILL.md content of a skill by name.
Args:
name: The skill name (hyphen-case, e.g. 'my-skill').
"""
from deerflow.skills.storage import get_or_new_skill_storage
from deerflow.skills.usage import bump_view
storage = get_or_new_skill_storage()
try:
if storage.custom_skill_exists(name):
content = storage.read_custom_skill(name)
elif storage.public_skill_exists(name):
skill_file = storage.get_skills_root_path()
from pathlib import Path
from deerflow.skills.types import SKILL_MD_FILE
path = Path(skill_file) / "public" / name / SKILL_MD_FILE
if not path.exists():
return f"Skill '{name}' not found."
content = path.read_text(encoding="utf-8")
else:
return f"Skill '{name}' not found."
except Exception as e:
logger.debug("skill_view failed for %r: %s", name, e)
return f"Error reading skill '{name}': {e}"
try:
bump_view(name)
except Exception:
pass
return content
@tool("skill_list")
def skill_list_tool() -> str:
"""List all available skills with metadata and usage stats. Does NOT read SKILL.md content.
Returns a JSON array with fields: name, description, category, enabled, state,
use_count, view_count, patch_count, last_activity_at, pinned, source.
"""
from deerflow.skills.storage import get_or_new_skill_storage
from deerflow.skills.usage import all_entries, latest_activity_at
try:
storage = get_or_new_skill_storage()
skills = storage.load_skills(enabled_only=False)
usage = all_entries()
result = []
for skill in skills:
entry = usage.get(skill.name, {})
result.append({
"name": skill.name,
"description": skill.description,
"category": skill.category.value if hasattr(skill.category, "value") else str(skill.category),
"enabled": skill.enabled,
"state": entry.get("state", "active"),
"use_count": entry.get("use_count", 0),
"view_count": entry.get("view_count", 0),
"patch_count": entry.get("patch_count", 0),
"last_activity_at": latest_activity_at(skill.name),
"pinned": entry.get("pinned", False),
"source": entry.get("source", "upload"),
})
return json.dumps(result, ensure_ascii=False, indent=2)
except Exception as e:
logger.debug("skill_list failed: %s", e)
return f"Error listing skills: {e}"
@@ -104,6 +104,11 @@ async def _skill_manage_impl(
name,
_history_record(action="create", file_path=SKILL_MD_FILE, prev_content=None, new_content=content, thread_id=thread_id, scanner=scan),
)
try:
from deerflow.skills.usage import mark_agent_created
await _to_thread(mark_agent_created, name)
except Exception:
pass
await refresh_skills_system_prompt_cache_async()
return f"Created custom skill '{name}'."
@@ -86,9 +86,10 @@ def get_available_tools(
builtin_tools.append(scheduled_task_tool)
skill_evolution_config = getattr(config, "skill_evolution", None)
if getattr(skill_evolution_config, "enabled", False):
from deerflow.tools.builtins.skill_tools import skill_list_tool, skill_view_tool
from deerflow.tools.skill_manage_tool import skill_manage_tool
builtin_tools.append(skill_manage_tool)
builtin_tools.extend([skill_list_tool, skill_view_tool, skill_manage_tool])
# Add subagent tools only if enabled via runtime parameter
if subagent_enabled: