mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-21 12:51:59 +00:00
工具调用失败能自救;agent 改成一步一个工具
起因是一条没人看得懂的报错:
工具 db_query__shop 执行失败:
_make_query_tool.<locals>._run() got an unexpected keyword argument 'query'
参数名写成了 query(它叫 sql),而 call_tool 对动态工具是裸 **args 展开的,
跳过了 args_schema —— pydantic 那句"缺少 sql 字段"漏成了只剩内部闭包名的
TypeError,既指不到是哪个工具,也说不出该填什么。同一个函数里内置工具那条
分支本来是校验的,两条路径不一致。
registry 里补上 prepare_args / describe_args,两条路径共用:
- 候选唯一时(多一个不认识的、少一个必填的)改名重试,并发一条 warn 说破
—— 配置里那个错名还在,替它跑通一次却不吭声,下次照样踩
- 有歧义就不猜,报错改成"参数不对:query 不是它的参数;缺少必填参数 sql。
它接受:sql(必填,字符串)— 要执行的 SQL…",模型照着这句就能改对。
这才是失败之后真正能自愈的重试
- 顺带堵了一类静默失败:pydantic 默认把认不出的键悄悄丢掉,全可选 schema 上
{"tabel": "orders"} 会"成功"执行一次什么都没查的调用
agent 循环默认改成串行:一轮只调一个工具,看到结果再想下一步。并行更快,但
一批调用中间没有任何新的思考——第二个是照着"还没看到第一个结果"时的判断发出
去的。节点上留了「并行执行工具」开关。两个必须守住的点:没执行的调用绝不发
tool.start(前端靠 call_id 配对,发了没结尾的 start 会永远转圈),但必须各有
一条配对的 tool_result(少一条下次调用直接 400)。
交错思考跟着补上。4.6 起 adaptive thinking 自带交错,不需要也不该挂 beta
header(Opus 4.6 上手动模式压根没有交错);4/4.5 那一代要手动开 thinking +
interleaved-thinking beta,光挂 header 是一条什么都不做的空设置。开了 thinking
就不能再传 temperature,这一代本来是收的,漏了就是 400。
还有两处是串行带出来的,一并治了:
- 步数用完时收尾取的是 messages[-1],而那往往是给模型看的 ToolMessage。
真漏到用户面前当过答案(run dd9927e6):一句"本轮只执行了第一个工具"冒充
结论。改成只取模型自己说过的话,并把"步数用完了、去调大最大步数"附在后面
- db_schema 一次能问多张表。一次调用就是一步,逐张问 6 张表就是 6 步——
实测 8 步的运行里 7 步花在这上面,最后只剩一步真查数据
This commit is contained in:
@@ -17,8 +17,21 @@ from app.db.models import Skill
|
||||
from app.engine.context import NodeContext, NodeError
|
||||
from app.engine.state import GraphState, message_text, template_context, thinking_text
|
||||
from app.providers import catalog
|
||||
from app.providers.factory import ModelSpec, ProviderNotConfigured, get_chat_model
|
||||
from app.tools.registry import ToolContext, build_tools, get_spec
|
||||
from app.providers.factory import (
|
||||
ModelSpec,
|
||||
ProviderNotConfigured,
|
||||
bind_tools_safely,
|
||||
get_chat_model,
|
||||
)
|
||||
from app.tools.registry import (
|
||||
ToolArgsError,
|
||||
ToolContext,
|
||||
args_model_of,
|
||||
build_tools,
|
||||
describe_args,
|
||||
get_spec,
|
||||
prepare_args,
|
||||
)
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 公共部分
|
||||
@@ -270,6 +283,33 @@ async def _resolve_tools(ctx: NodeContext, state: GraphState) -> list[BaseTool]:
|
||||
return await build_tools(names, tool_ctx, session=session)
|
||||
|
||||
|
||||
SKIPPED_NOTE = (
|
||||
"本轮只执行了第一个工具。看到它的结果之后,再决定这一个还要不要调、要用什么参数调。"
|
||||
)
|
||||
|
||||
|
||||
def split_tool_calls(
|
||||
tool_calls: list[dict[str, Any]], *, parallel: bool
|
||||
) -> tuple[list[dict[str, Any]], list[tuple[dict[str, Any], ToolMessage]]]:
|
||||
"""把一轮里的工具调用分成"这次跑"和"这次不跑"。
|
||||
|
||||
串行时只跑第一个。剩下的不能就这么丢掉:两家 API 都要求每个 tool_use 在
|
||||
下一轮有一条配对的 tool_result,少一条下次调用直接 400。所以这里给每个
|
||||
未执行的调用配一条说明性的 ToolMessage 一起交出去。
|
||||
|
||||
它们也**绝不能**发 tool.start —— 前端靠 call_id 配对 start/end
|
||||
(frontend/src/run/decode.ts),发了一条没有结尾的 start,那个工具卡片就会
|
||||
在界面上永远转圈。所以这个函数不碰事件,只产消息。
|
||||
"""
|
||||
if parallel or len(tool_calls) <= 1:
|
||||
return list(tool_calls), []
|
||||
deferred = [
|
||||
(call, ToolMessage(content=SKIPPED_NOTE, tool_call_id=call.get("id") or f"skipped-{i}"))
|
||||
for i, call in enumerate(tool_calls[1:])
|
||||
]
|
||||
return tool_calls[:1], deferred
|
||||
|
||||
|
||||
def _needs_approval(ctx: NodeContext, tool_name: str) -> bool:
|
||||
mode = ctx.cfg("approval", "dangerous") # never | dangerous | always
|
||||
if mode == "always":
|
||||
@@ -295,10 +335,13 @@ async def run_agent(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
|
||||
tools = await _resolve_tools(ctx, state)
|
||||
tool_map = {t.name: t for t in tools}
|
||||
model = base_model.bind_tools(tools) if tools else base_model
|
||||
# 默认串行:一轮只调一个工具,拿到结果再想下一步。并行更快,但一批工具
|
||||
# 中间没有任何新的思考——第二个调用是照着"还没看到第一个结果"时的判断发出来的
|
||||
parallel_tools = bool(ctx.cfg("parallel_tools", False))
|
||||
model = bind_tools_safely(base_model, tools, parallel=parallel_tools)
|
||||
|
||||
messages = await _build_messages(state, ctx)
|
||||
max_steps = min(int(ctx.cfg("max_steps", 8) or 8), settings.max_agent_steps)
|
||||
max_steps = min(int(ctx.cfg("max_steps", 12) or 12), settings.max_agent_steps)
|
||||
total_usage: dict[str, Any] = {"input_tokens": 0, "output_tokens": 0, "cost_usd": 0.0, "calls": 0}
|
||||
transcript: list[dict[str, Any]] = []
|
||||
|
||||
@@ -328,7 +371,9 @@ async def run_agent(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
final_text = message_text(response)
|
||||
break
|
||||
|
||||
for call in tool_calls:
|
||||
run_calls, deferred = split_tool_calls(tool_calls, parallel=parallel_tools)
|
||||
|
||||
for call in run_calls:
|
||||
name = call.get("name", "")
|
||||
args = call.get("args", {}) or {}
|
||||
call_id = call.get("id") or f"{ctx.node.id}-{step}-{name}"
|
||||
@@ -339,6 +384,24 @@ async def run_agent(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
)
|
||||
continue
|
||||
|
||||
# 参数先过 schema 再决定要不要打扰人审批:参数就不对的调用没必要问人,
|
||||
# 而纠正过的参数必须是人在审批框里看到的那一份
|
||||
schema = args_model_of(tool_map[name])
|
||||
if schema is not None:
|
||||
try:
|
||||
args, fix_note = prepare_args(schema, args)
|
||||
except ToolArgsError as e:
|
||||
# 不执行,把"它接受什么"原样喂回去。模型下一步照着改就行,
|
||||
# 这才是失败之后真正能自愈的重试——以前喂回去的是一句
|
||||
# TypeError,里面既没有正确参数名也没有字段说明
|
||||
ctx.emit(EventType.TOOL_ERROR, tool=name, call_id=call_id, error=str(e))
|
||||
messages.append(ToolMessage(content=str(e), tool_call_id=call_id))
|
||||
transcript.append({"tool": name, "args": args, "ok": False, "result": str(e)})
|
||||
continue
|
||||
if fix_note:
|
||||
ctx.emit(EventType.LOG, level="warn",
|
||||
message=f"工具 {name}:{fix_note}")
|
||||
|
||||
if _needs_approval(ctx, name):
|
||||
ctx.emit(
|
||||
EventType.HUMAN_REQUESTED,
|
||||
@@ -377,6 +440,10 @@ async def run_agent(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
ok = True
|
||||
except Exception as e: # noqa: BLE001 - 工具失败要喂回模型,让它自己纠错
|
||||
content = f"工具执行失败:{type(e).__name__}: {e}"
|
||||
if isinstance(e, TypeError) and schema is not None:
|
||||
# 签名对不上还能走到这儿,说明 schema 和函数本身不一致(工具的 bug)。
|
||||
# 模型改不了这个,但把参数表摊开至少让它别在同一个地方反复试
|
||||
content += f"\n{describe_args(schema, args)}"
|
||||
ok = False
|
||||
elapsed = int((time.perf_counter() - started) * 1000)
|
||||
snapshot_id: str | None = None
|
||||
@@ -402,13 +469,32 @@ async def run_agent(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
{"tool": name, "args": args, "ok": ok, "duration_ms": elapsed, "result": content[:4000]}
|
||||
)
|
||||
messages.append(ToolMessage(content=content, tool_call_id=call_id))
|
||||
|
||||
# 没跑的那些排在执行过的后面,保持模型原本的调用顺序
|
||||
for call, message in deferred:
|
||||
messages.append(message)
|
||||
transcript.append(
|
||||
{"tool": call.get("name", ""), "args": call.get("args", {}) or {}, "skipped": True}
|
||||
)
|
||||
else:
|
||||
final_text = message_text(messages[-1]) if messages else ""
|
||||
ctx.emit(
|
||||
EventType.LOG,
|
||||
level="warn",
|
||||
message=f"agent 达到最大步数 {max_steps},提前收尾",
|
||||
# 步数用完了。成果只能取模型自己说过的话——messages[-1] 很可能是一条
|
||||
# ToolMessage:工具的原始返回,或者串行模式下那句"本轮只执行了第一个
|
||||
# 工具"。后者真的漏到用户面前当过答案(run dd9927e6),一句内部管道
|
||||
# 文案冒充结论,比明说"没跑完"糟糕得多。
|
||||
final_text = next(
|
||||
(text for m in reversed(messages)
|
||||
if isinstance(m, AIMessage) and (text := message_text(m).strip())),
|
||||
"",
|
||||
)
|
||||
hint = (
|
||||
f"agent 用满了 {max_steps} 步还没给出结论。"
|
||||
"把节点上的「最大步数」调大;如果它大部分步数花在逐张表查结构上,"
|
||||
"也可以在提示词里点明该查哪几张表。"
|
||||
)
|
||||
# 一步一工具时步数消耗得比并行快得多,这里不说清楚,用户只会看到
|
||||
# 一个没头没尾的答案,而不知道是被步数掐断的
|
||||
final_text = f"{final_text}\n\n({hint})".strip() if final_text else f"({hint})"
|
||||
ctx.emit(EventType.LOG, level="warn", message=hint)
|
||||
|
||||
total_usage["total_tokens"] = total_usage["input_tokens"] + total_usage["output_tokens"]
|
||||
result = {
|
||||
|
||||
@@ -14,8 +14,14 @@ from app.db.models import Workflow
|
||||
from app.engine.context import NodeContext, NodeError
|
||||
from app.engine.state import GraphState, message_text, template_context
|
||||
from app.providers import catalog
|
||||
from app.providers.factory import ModelSpec, get_chat_model
|
||||
from app.tools.registry import ToolContext, build_tools
|
||||
from app.providers.factory import ModelSpec, bind_tools_safely, get_chat_model
|
||||
from app.tools.registry import (
|
||||
ToolArgsError,
|
||||
ToolContext,
|
||||
args_model_of,
|
||||
build_tools,
|
||||
prepare_args,
|
||||
)
|
||||
|
||||
_MAX_DEPTH = 3
|
||||
|
||||
@@ -122,7 +128,7 @@ async def run_supervisor(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
)
|
||||
tools = await build_tools(cfg.get("tools", []) or [], tool_ctx, session=session)
|
||||
|
||||
bound = model.bind_tools(tools) if tools else model
|
||||
bound = bind_tools_safely(model, tools, parallel=False)
|
||||
tool_map = {t.name: t for t in tools}
|
||||
messages: list[Any] = [
|
||||
SystemMessage(content=cfg.get("system") or f"你是 {name}。"),
|
||||
@@ -140,15 +146,32 @@ async def run_supervisor(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
tname = call.get("name", "")
|
||||
targs = call.get("args", {}) or {}
|
||||
cid = call.get("id") or f"{name}-{tname}"
|
||||
ctx.emit(EventType.TOOL_START, tool=tname, args=targs, agent=name, call_id=cid)
|
||||
if tname not in tool_map:
|
||||
ctx.emit(EventType.TOOL_START, tool=tname, args=targs, agent=name, call_id=cid)
|
||||
content = f"错误:没有名为 {tname} 的工具"
|
||||
else:
|
||||
try:
|
||||
raw = await tool_map[tname].ainvoke(targs)
|
||||
content = raw if isinstance(raw, str) else json.dumps(raw, ensure_ascii=False, default=str)
|
||||
except Exception as e: # noqa: BLE001
|
||||
content = f"工具失败:{type(e).__name__}: {e}"
|
||||
schema = args_model_of(tool_map[tname])
|
||||
fix_note = None
|
||||
args_error = None
|
||||
if schema is not None:
|
||||
try:
|
||||
targs, fix_note = prepare_args(schema, targs)
|
||||
except ToolArgsError as e:
|
||||
args_error = str(e)
|
||||
# 纠正后的参数才是真正要执行的那份,tool.start 要发它
|
||||
ctx.emit(EventType.TOOL_START, tool=tname, args=targs, agent=name, call_id=cid)
|
||||
if fix_note:
|
||||
ctx.emit(EventType.LOG, level="warn",
|
||||
message=f"工具 {tname}:{fix_note}")
|
||||
if args_error:
|
||||
# 参数就不对,没必要真调一次。把"它接受什么"喂回去让它改
|
||||
content = args_error
|
||||
else:
|
||||
try:
|
||||
raw = await tool_map[tname].ainvoke(targs)
|
||||
content = raw if isinstance(raw, str) else json.dumps(raw, ensure_ascii=False, default=str)
|
||||
except Exception as e: # noqa: BLE001
|
||||
content = f"工具失败:{type(e).__name__}: {e}"
|
||||
ctx.emit(EventType.TOOL_END, tool=tname, agent=name, call_id=cid, preview=content[:1500])
|
||||
calls.append({"tool": tname, "args": targs, "result": content[:2000]})
|
||||
messages.append(ToolMessage(content=content, tool_call_id=cid))
|
||||
|
||||
@@ -12,7 +12,7 @@ from app.engine.context import NodeContext, NodeError
|
||||
from app.engine.state import GraphState
|
||||
from app.sandbox.base import SandboxLimits
|
||||
from app.sandbox.manager import sandbox_manager
|
||||
from app.tools.registry import ToolContext, call_tool, get_spec
|
||||
from app.tools.registry import ToolArgsError, ToolContext, call_tool, get_spec
|
||||
|
||||
|
||||
def _tool_ctx(ctx: NodeContext) -> ToolContext:
|
||||
@@ -55,11 +55,25 @@ async def run_tool(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
|
||||
ctx.emit(EventType.TOOL_START, tool=name, args=args)
|
||||
started = time.perf_counter()
|
||||
|
||||
def _fixed(note: str) -> None:
|
||||
# 替它跑通了这一次,但节点配置里那个错的参数名原封不动,下次还会踩。
|
||||
# 所以纠正必须留一条看得见的痕迹,而不是安静地把事办了
|
||||
ctx.emit(EventType.LOG, level="warn",
|
||||
message=f"工具 {name}:{note}。请到节点里改正。")
|
||||
|
||||
try:
|
||||
async with SessionLocal() as session:
|
||||
result = await call_tool(name, args, _tool_ctx(ctx), session=session)
|
||||
result = await call_tool(name, args, _tool_ctx(ctx), session=session, on_fix=_fixed)
|
||||
except KeyError as e:
|
||||
raise NodeError(ctx.node.id, str(e)) from e
|
||||
except ToolArgsError as e:
|
||||
# 参数对不上且没有唯一候选可纠。报错里已经写清楚该填什么,
|
||||
# 不要再套一层 "执行失败:" 把它推远
|
||||
ctx.emit(EventType.TOOL_ERROR, tool=name, error=str(e))
|
||||
if ctx.cfg("fail_fast", True):
|
||||
raise NodeError(ctx.node.id, f"工具 {name}:{e}") from e
|
||||
result = {"error": str(e)}
|
||||
except Exception as e: # noqa: BLE001
|
||||
ctx.emit(EventType.TOOL_ERROR, tool=name, error=f"{type(e).__name__}: {e}")
|
||||
if ctx.cfg("fail_fast", True):
|
||||
|
||||
Reference in New Issue
Block a user