feat: 技能收藏与常驻(always-on)支持
- 新增 skill_favorites / skill_always_on 迁移与持久化模型字段 - skills 路由、存储、前端设置页支持收藏与常驻 - 调整摘要中间件与 lead agent 提示词 - 同步岗位 position_sync 及相关配置清理 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
1 parent
5ebbb82d66
commit
5182ea5d11
19 files changed
+472
-107
No files matched your search
+52
-2
@@ -86,8 +86,58 @@ def _drop_archived_skills(skills: list[Skill]) -> list[Skill]:
|
||||
return skills
|
||||
|
||||
|
||||
def _run_coro_blocking(coro):
|
||||
"""Run an async coroutine to completion from sync code.
|
||||
|
||||
Mirrors ``skill_tools._run_coro_blocking``: when already inside a running
|
||||
event loop (e.g. prompt built during an async agent run), hand the coroutine
|
||||
to a fresh loop in a worker thread; otherwise run it directly.
|
||||
"""
|
||||
try:
|
||||
asyncio.get_running_loop()
|
||||
in_loop = True
|
||||
except RuntimeError:
|
||||
in_loop = False
|
||||
|
||||
if in_loop:
|
||||
import concurrent.futures
|
||||
|
||||
with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
|
||||
return pool.submit(asyncio.run, coro).result()
|
||||
return asyncio.run(coro)
|
||||
|
||||
|
||||
def _load_always_on_names() -> set[str]:
|
||||
"""Names marked 常驻(免压缩), read from the **DB** (skills table).
|
||||
|
||||
Returns an empty set on any failure / no-DB deployment, so compression then
|
||||
simply treats every skill as compressible (safe default).
|
||||
"""
|
||||
try:
|
||||
from deerflow.persistence.engine import get_session_factory
|
||||
from deerflow.persistence.skills import make_skill_store
|
||||
|
||||
store = make_skill_store(get_session_factory())
|
||||
return set(_run_coro_blocking(store.list_always_on_names()))
|
||||
except Exception:
|
||||
logger.debug("Failed to load always_on skill names from DB", exc_info=True)
|
||||
return set()
|
||||
|
||||
|
||||
def _apply_always_on(skills: list[Skill]) -> list[Skill]:
|
||||
"""Stamp ``Skill.always_on`` from the DB-backed 常驻 set (in place)."""
|
||||
try:
|
||||
always_on = _load_always_on_names()
|
||||
if always_on:
|
||||
for skill in skills:
|
||||
skill.always_on = skill.name in always_on
|
||||
except Exception:
|
||||
logger.debug("Failed to apply always_on flags to skills", exc_info=True)
|
||||
return skills
|
||||
|
||||
|
||||
def _load_enabled_skills_sync() -> list[Skill]:
|
||||
return _drop_archived_skills(list(get_or_new_skill_storage().load_skills(enabled_only=True)))
|
||||
return _apply_always_on(_drop_archived_skills(list(get_or_new_skill_storage().load_skills(enabled_only=True))))
|
||||
|
||||
|
||||
def _start_enabled_skills_refresh_thread() -> None:
|
||||
@@ -188,7 +238,7 @@ def _get_enabled_skills_for_config(app_config: AppConfig | None = None) -> list[
|
||||
"""
|
||||
if app_config is None:
|
||||
return _get_enabled_skills()
|
||||
return _drop_archived_skills(list(get_or_new_skill_storage(app_config=app_config).load_skills(enabled_only=True)))
|
||||
return _apply_always_on(_drop_archived_skills(list(get_or_new_skill_storage(app_config=app_config).load_skills(enabled_only=True))))
|
||||
|
||||
|
||||
def _skill_mutability_label(category: SkillCategory | str) -> str:
|
||||
|
||||
+41
-2
@@ -314,16 +314,36 @@ class DeerFlowSummarizationMiddleware(SummarizationMiddleware):
|
||||
logger.debug("summarization: thread_id unresolved; cross-turn summary cache disabled for this call")
|
||||
cached = self._get_cache_entry(thread_id) if thread_id else None
|
||||
present_ids = {m.id for m in messages}
|
||||
if cached is not None and cached.reusable and cached.summarized_ids <= present_ids:
|
||||
cache_usable = cached is not None and cached.reusable and cached.summarized_ids <= present_ids
|
||||
if cache_usable:
|
||||
preserved = [m for m in messages if m.id not in cached.summarized_ids]
|
||||
effective = [self._summary_message(cached.summary_text), *preserved]
|
||||
if not self._should_summarize(effective, self.token_counter(effective)):
|
||||
# Decide re-summarization on the *counted* size of the compressed
|
||||
# context only — NOT ``_should_summarize`` — because that folds in the
|
||||
# last AIMessage's reported ``usage_metadata`` (system prompt + tool
|
||||
# schemas, fixed overhead compression can't shrink). Using it here
|
||||
# would re-summarize on every single turn once that overhead alone
|
||||
# approaches the trigger.
|
||||
if not self._exceeds_trigger_by_count(effective):
|
||||
# Refresh LRU recency and reuse without an LLM call.
|
||||
self._touch_cache_entry(thread_id)
|
||||
return ("reuse", effective)
|
||||
# New messages have pushed the compressed context back over the
|
||||
# threshold — fall through and re-summarize a larger prefix.
|
||||
|
||||
# Genuine (blocking) LLM summarize path. Log *why* we are paying for it so
|
||||
# an "it compacts every turn" report is diagnosable straight from the
|
||||
# backend log: cache=miss → thread_id/key problem; unusable → message-id
|
||||
# instability; stale → preserved genuinely grew back over the trigger.
|
||||
cache_state = "stale" if cache_usable else ("miss" if cached is None else "unusable")
|
||||
logger.info(
|
||||
"summarization: compacting thread=%s cache=%s counted_tokens=%d msgs=%d",
|
||||
thread_id,
|
||||
cache_state,
|
||||
self.token_counter(messages),
|
||||
len(messages),
|
||||
)
|
||||
|
||||
cutoff_index = self._determine_cutoff_index(messages)
|
||||
if cutoff_index <= 0:
|
||||
return None
|
||||
@@ -386,6 +406,25 @@ class DeerFlowSummarizationMiddleware(SummarizationMiddleware):
|
||||
while len(self._summary_cache) > _SUMMARY_CACHE_MAX_THREADS:
|
||||
self._summary_cache.popitem(last=False)
|
||||
|
||||
def _exceeds_trigger_by_count(self, messages: list[AnyMessage]) -> bool:
|
||||
"""Trigger check using only the *counted* size of ``messages``.
|
||||
|
||||
Deliberately ignores the stock middleware's reported-``usage_metadata``
|
||||
heuristic: that number includes fixed system-prompt + tool-schema
|
||||
overhead which compression can never shrink, so relying on it to decide
|
||||
*re*-summarization makes a long, tool-heavy thread re-compact on every
|
||||
turn. The first-ever decision to compress still uses the full
|
||||
``_should_summarize`` (reported usage is a fair signal there); only the
|
||||
cache-reuse-vs-recompact decision uses this stricter, content-only check.
|
||||
"""
|
||||
total = self.token_counter(messages)
|
||||
for kind, value in getattr(self, "_trigger_conditions", []):
|
||||
if kind == "messages" and len(messages) >= value:
|
||||
return True
|
||||
if kind == "tokens" and total >= value:
|
||||
return True
|
||||
return False
|
||||
|
||||
def _apply_stickiness_headroom(self, messages: list[AnyMessage], cutoff_index: int) -> int:
|
||||
"""Advance ``cutoff_index`` (keep fewer recent messages) so the preserved
|
||||
suffix lands below every trigger threshold with headroom.
|
||||
|
||||
Reference in new issue
Block a user