feat: 标签管理功能 + 问答前缀对用户隐藏

标签管理:
- 新增 tags/tag_assignments 持久化层、迁移、/api/tags 路由
- 标签分智能体/技能/定时任务三类, 仅同类型可互打, 仅 admin 可增删改与打标签
- 三类资源列表支持按名称模糊检索与按标签筛选(走后端查询参数)
- 前端新增标签管理页、TagPickerDialog(点击即时绑定)、彩色标签 chip

问答前缀:
- 前缀不再拼入消息 content, 改存 additional_kwargs 由 PromptPrefixMiddleware
  在调模型前临时注入, 用户界面看不到前缀但模型仍收到
- 修正 thread_existed 闸门导致首轮前缀被跳过的问题

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
吉鑫升andClaude Opus 4.7 committed 2026-05-15 16:19:08 +08:00
1 parent aeba15eb16
commit 646cb8caeb
47 files changed
+2180 -109

No files matched your search

@@ -100,12 +100,15 @@ _DEFAULT_ASSISTANT_ID = "lead_agent"
def _prepend_prompt_prefix(graph_input: dict[str, Any], prefix: str) -> dict[str, Any]:
"""Return ``graph_input`` with the prefix prepended to its first HumanMessage.
"""Return ``graph_input`` with the prefix *tagged* onto its first HumanMessage.
The prefix is stored in ``additional_kwargs['prompt_prefix']`` rather than
concatenated into ``content``. This keeps the persisted/displayed message
clean — the user never sees the prefix — while ``PromptPrefixMiddleware``
injects it into the content of the request actually sent to the LLM.
Leaves the input unchanged when there are no messages or the first message
is not a human turn. Idempotent in the sense that it only touches the first
human message, so re-running this with the same prefix would double-prepend
— callers must gate it on first-turn detection.
is not a human turn.
"""
if not prefix:
return graph_input
@@ -115,14 +118,9 @@ def _prepend_prompt_prefix(graph_input: dict[str, Any], prefix: str) -> dict[str
new_messages = list(messages)
for idx, msg in enumerate(new_messages):
if isinstance(msg, HumanMessage):
existing = msg.content
if isinstance(existing, str):
new_messages[idx] = HumanMessage(content=f"{prefix}\n\n{existing}".strip("\n"))
elif isinstance(existing, list):
# Structured content (text blocks + media). Prepend a new text block.
new_messages[idx] = HumanMessage(content=[{"type": "text", "text": prefix}, *existing])
else:
new_messages[idx] = HumanMessage(content=f"{prefix}\n\n{existing!s}")
extra = dict(getattr(msg, "additional_kwargs", None) or {})
extra["prompt_prefix"] = prefix
new_messages[idx] = msg.model_copy(update={"additional_kwargs": extra})
break
return {**graph_input, "messages": new_messages}
@@ -380,7 +378,6 @@ async def start_run(
# Upsert thread metadata so the thread appears in /threads/search,
# even for threads that were never explicitly created via POST /threads
# (e.g. stateless runs).
thread_existed = False
try:
existing = await run_ctx.thread_store.get(thread_id)
if existing is None:
@@ -390,7 +387,6 @@ async def start_run(
metadata=body.metadata,
)
else:
thread_existed = True
await run_ctx.thread_store.update_status(thread_id, "running")
except Exception:
logger.warning("Failed to upsert thread_meta for %s (non-fatal)", sanitize_log_param(thread_id))
@@ -398,13 +394,15 @@ async def start_run(
agent_factory = resolve_agent_factory(body.assistant_id)
graph_input = normalize_input(body.input)
# Server-side safety net for the admin-configurable prompt prefix: when the
# caller opted in via context.prompt_prefix_enabled AND this is the very
# first turn of the thread, prepend the configured prefix to the human
# message. We deliberately skip already-running threads so the prefix is
# never inserted twice — the frontend toggle is a per-message hint, but
# this check is what makes the feature safe against client bugs / replays.
if not thread_existed and _prompt_prefix_opted_in(body):
# Admin-configurable prompt prefix: when the caller opted in via
# context.prompt_prefix_enabled, tag the human message with the configured
# prefix (stored in additional_kwargs, not concatenated into content, so the
# user never sees it — PromptPrefixMiddleware injects it for the LLM only).
# The frontend only sends the flag on the opening turn, so this naturally
# applies once. We intentionally do NOT gate on thread-row existence: the
# thread_meta row is frequently pre-created, which previously skipped the
# prefix for genuine first turns.
if _prompt_prefix_opted_in(body):
prefix_settings = load_system_settings().prompt_prefix
if prefix_settings.enabled and prefix_settings.prompt_prefix.strip():
graph_input = _prepend_prompt_prefix(graph_input, prefix_settings.prompt_prefix.strip())