多智能体会商上传文件

This commit is contained in:
fengxiangdebaozi committed 2026-06-10 19:38:58 +08:00
1 parent cf5550ab0a
commit f01789a17e
9 files changed
+411 -49

No files matched your search

@@ -328,9 +328,15 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
# config.yaml 的 models[0],与 /api/ai-writing 等其它路由一致,内网部署只配
# 自己的本地模型即可工作。
model_name = str(payload.get("model") or "")
# 用户在 Step 1 上传的参考文件(已经过 /api/threads/{thread_id}/uploads)。
# 前端传 [{filename, size, path, status}, ...],这里原样透传给 _run_payload →
# additional_kwargs.files → UploadsMiddleware 自动注入清单 + 解锁 read_file 等工具。
raw_files = payload.get("files")
files = raw_files if isinstance(raw_files, list) else []
if not thread_id:
raise HTTPException(status_code=400, detail="thread_id is required")
if not message:
# 允许「仅文件、无文本」的回合:带了文件就不强制要求 message。
if not message and not files:
raise HTTPException(status_code=400, detail="message is required")
# 调用前自检:防御性二次检查——若管理员在 init 之后手动删了目录,这里也能自愈。
@@ -381,8 +387,17 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
_MAX_CLARIFICATION_ROUNDS,
)
# 带了上传文件时,解锁「读」类工具(read_file / ls / grep / view_image),
# 否则 UploadsMiddleware 注入了文件清单、模型却没工具读取。仍保留对
# write_file / str_replace / present_files / bash / agent_orchestration 的封锁,
# 防止意图阶段越界做方案/派活。
if files:
read_tools = {"read_file", "ls", "grep", "view_image"}
base_excluded_tools = [t for t in base_excluded_tools if t not in read_tools]
# 2. message 前缀——给模型一个明确的"本轮必须收敛"信号。
effective_message = message
# 仅文件、无文本时给一句兜底引导,让模型先读文件再澄清意图。
effective_message = message or "我上传了文件,请先阅读文件内容,再据此理解并澄清我的任务意图。"
if past_clarifications >= _MAX_CLARIFICATION_ROUNDS:
effective_message = (
f"[系统强制] 已澄清 {past_clarifications} 轮,触发硬上限。"
@@ -409,6 +424,7 @@ async def stream_intent(request: Request, payload: dict = Body(...)) -> Streamin
# not reliable on its own, hard exclusion is.
excluded_tools=base_excluded_tools,
skill_stop_names=[],
files=files,
)
# captured:上游 SSE 全部 data 行,用于流结束后从 values 快照取 messages。
@@ -638,6 +638,7 @@ def _run_payload(
reasoning_effort: str = "medium",
multitask_strategy: str = "reject",
extra_context: dict[str, Any] | None = None,
files: list[dict[str, Any]] | None = None,
) -> dict[str, Any]:
"""构造 ``POST /api/threads/{thread_id}/runs/stream`` 的请求体。
@@ -650,13 +651,19 @@ def _run_payload(
应传 ``thinking_enabled=False`` + ``reasoning_effort="low"``,省掉模型先吐
reasoning tokens 才开口的时间,显著降低前端"调度中"首字节延迟。
"""
# 用户上传的文件元数据透传给 UploadsMiddleware(与普通聊天同链路):
# before_agent 钩子读 additional_kwargs.files,生成 <uploaded_files> 清单注入消息,
# 之后席位/总控用内置 read_file / grep / view_image 工具按需读取。空时落空 dict,行为不变。
additional_kwargs: dict[str, Any] = {}
if files:
additional_kwargs["files"] = files
payload: dict[str, Any] = {
"input": {
"messages": [
{
"type": "human",
"content": [{"type": "text", "text": new_message}],
"additional_kwargs": {},
"additional_kwargs": additional_kwargs,
}
]
},
@@ -1258,6 +1265,7 @@ async def _leader_run(
seat_meta: dict[str, dict[str, Any]] | None = None,
synthesis_mode: bool = False,
suppress_pre_broadcast: bool = False,
files: list[dict[str, Any]] | None = None,
) -> AsyncIterator[str | dict[str, Any]]:
"""总控(leader)单轮:透传上游 SSE + 末尾 yield 一个 dict 状态帧(由路由层 ``json.dumps``)。
@@ -1346,18 +1354,26 @@ async def _leader_run(
# 调用方可再叠加自己的 skill_stop。后台进程内网关用同一份 policy,保证传参一致。
policy = roundtable_run_policy("leader")
effective_stop_names = merge_skill_stop_names(policy.skill_stop_names, skill_stop_names)
# 带了上传文件时,给总控解锁「读」类工具(leader policy 默认禁 read_file/ls/view_image),
# 否则文件清单注入了却无工具读取。写/派活类仍按 policy 封锁,不放开。
leader_excluded = policy.excluded_tools
if files:
_read_tools = {"read_file", "ls", "grep", "view_image"}
leader_excluded = [t for t in leader_excluded if t not in _read_tools]
body = _run_payload(
agent_name=agent_name,
model_name=model_name,
thread_id=thread_id,
# 注入了席位清单的输入(leader_input);seat_roster 为空时等同 new_message。
new_message=leader_input,
excluded_tools=policy.excluded_tools,
excluded_tools=leader_excluded,
skill_stop_names=effective_stop_names,
thinking_enabled=policy.thinking_enabled,
reasoning_effort=policy.reasoning_effort,
# 人工干预/恢复会重发 leader run 到同一总控 thread;抢占掉没结束的旧 run,避免 409。
multitask_strategy="rollback",
# 用户随干预上传的参考文件 → 总控可用 read_file 等工具读取后再派活。
files=files,
)
t_payload_built = time.perf_counter()
logger.info(
@@ -1532,6 +1548,7 @@ async def _special_run(
display_name: str = "",
no_file: bool = False,
peer_deliveries: list[dict[str, Any]] | None = None,
files: list[dict[str, Any]] | None = None,
) -> AsyncIterator[str | dict[str, Any]]:
"""子智能体(special)单轮:执行总控下发的 task 文本,交付 content。
@@ -1560,12 +1577,18 @@ async def _special_run(
else:
role = "seat"
policy = roundtable_run_policy(role)
# 带了上传文件时,给席位解锁「读」类工具(view_image 默认被禁;read_file 本就可用),
# 让席位能读取/识别用户上传的文件。写/派活类仍按 policy 封锁。
seat_excluded = policy.excluded_tools
if files:
_read_tools = {"read_file", "ls", "grep", "view_image"}
seat_excluded = [t for t in seat_excluded if t not in _read_tools]
body = _run_payload(
agent_name=agent_name,
model_name=model_name,
thread_id=thread_id,
new_message=new_message,
excluded_tools=policy.excluded_tools,
excluded_tools=seat_excluded,
skill_stop_names=merge_skill_stop_names(policy.skill_stop_names, skill_stop_names),
thinking_enabled=policy.thinking_enabled,
reasoning_effort=policy.reasoning_effort,
@@ -1575,6 +1598,8 @@ async def _special_run(
extra_context=(
{"roundtable_peer_deliveries": peer_deliveries} if peer_deliveries else None
),
# 用户随干预上传的参考文件 → 席位可用 read_file 等工具读取。
files=files,
)
t_payload_built = time.perf_counter()
logger.info(
@@ -1677,6 +1702,11 @@ async def run_multi_agent_stream(request: Request, payload: dict = Body(...)) ->
display_name: str = str(payload.get("display_name") or "").strip()
new_message: str = str(payload.get("new_message") or "")
skill_stop_names: list[str] = payload.get("skill_stop_names") or []
# 用户随干预消息上传的参考文件(已经过 /api/threads/{thread_id}/uploads)。
# 前端传 [{filename, size, path, status}, ...],透传给 _run_payload →
# additional_kwargs.files → UploadsMiddleware 注入清单 + 解锁 read_file 等工具。
_files_raw = payload.get("files")
files: list[dict[str, Any]] = _files_raw if isinstance(_files_raw, list) else []
# 综合轮标记(仅 leader 用):前端在「派活后让总控判断继续/综合」的轮次置真,网关据此
# 额外把各已交付席位的完整交付正文注入总控输入,供其写最终方案。缺省 False(派活决策轮
# 只给摘要,保持上下文精简)。旧前端不传 → False,行为安全退化。
@@ -1741,6 +1771,7 @@ async def run_multi_agent_stream(request: Request, payload: dict = Body(...)) ->
seat_meta=seat_meta,
synthesis_mode=synthesis_mode,
suppress_pre_broadcast=suppress_pre_broadcast,
files=files,
)
else:
gen = _special_run(
@@ -1748,6 +1779,7 @@ async def run_multi_agent_stream(request: Request, payload: dict = Body(...)) ->
display_name=display_name,
no_file=no_file,
peer_deliveries=peer_deliveries,
files=files,
)
async for chunk in gen:
if isinstance(chunk, dict):