mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 01:21:59 +00:00
助手改图不再把没说完当成「画布无需修改」：输出被截断、或一个操作都没给时报错，不再对着原图自查、交付；改图额度从 8192 提到 32768，服务方不收时退回 8192；提示词要求不在思考里草拟 JSON
This commit is contained in:
ThearchyHelios
@@ -756,6 +756,8 @@ async def generate(
|
||||
_STREAM_PROTOCOL = """\
|
||||
输出格式（严格遵守）：
|
||||
- 每行一个完整的 JSON 对象，除此之外不输出任何东西——不要 markdown 围栏、不要解释文字、不要空行注释
|
||||
- 思考和输出共用同一份长度额度：思考里只定方案（加哪些节点、怎么连、每个节点干什么），
|
||||
不要在思考里把 JSON、提示词、表达式先写一遍——那些直接写进输出。额度用完时还没输出的改动就全丢了
|
||||
- 可用操作：
|
||||
{"op":"plan","summary":"一句话说明打算怎么搭"} ← 第一行
|
||||
{"op":"add_node","node":{"id":"...","type":"...","label":"中文标签","config":{...}}}
|
||||
@@ -913,6 +915,73 @@ def _apply_op(nodes: dict[str, dict[str, Any]], edges: list[dict[str, Any]], op:
|
||||
return False
|
||||
|
||||
|
||||
class CopilotIncomplete(Exception):
|
||||
"""模型这一轮没说完：输出被截断，或者一个操作都没给。画布不能按「无需修改」交付。"""
|
||||
|
||||
def __init__(self, reason: str, *, hint: str = "", detail: str = "") -> None:
|
||||
super().__init__(reason)
|
||||
self.hint = hint or "画布未修改。可以重试，或把需求拆成几步分别提出"
|
||||
self.detail = detail
|
||||
|
||||
|
||||
def _cut_off(chunk: Any) -> str:
|
||||
"""这一块是不是在说「额度用完了」：OpenAI 系是 finish_reason=length，Anthropic 是 stop_reason=max_tokens。"""
|
||||
meta = getattr(chunk, "response_metadata", None) or {}
|
||||
reason = meta.get("finish_reason") or meta.get("stop_reason")
|
||||
return str(reason) if reason in ("length", "max_tokens") else ""
|
||||
|
||||
|
||||
#: 流式改图的输出额度。会思考的模型（DeepSeek、Claude 的 adaptive thinking）思考也算在里面：
|
||||
#: 8192 时复杂的改图常常想完就没额度了，正文一个字没写（2026-09 部署实例上同类请求 5 轮坏了 4 轮）
|
||||
COPILOT_MAX_TOKENS = 32768
|
||||
#: 服务方不收大额度时退回这个（原来的默认值）
|
||||
_FALLBACK_MAX_TOKENS = 8192
|
||||
#: 服务方嫌额度太大时报错里会出现的字样：超过模型的输出上限（DeepSeek chat 系只到 8K），
|
||||
#: 或者输入加额度超过上下文长度（本地 vLLM）
|
||||
_BUDGET_WORDS = ("max_tokens", "max_completion_tokens", "max_output_tokens", "maximum context length")
|
||||
|
||||
|
||||
def _budget_rejected(e: BaseException) -> bool:
|
||||
status = getattr(e, "status_code", None) or getattr(getattr(e, "response", None), "status_code", None)
|
||||
return status in (400, 422) and any(w in str(e).lower() for w in _BUDGET_WORDS)
|
||||
|
||||
|
||||
class _BudgetFallback:
|
||||
"""先按大额度调；服务方一开口就嫌额度太大时按原来的额度重来，之后的调用（自查修正轮）都用小的。
|
||||
|
||||
只在还没吐出任何东西时重来：这类拒绝都发生在请求刚发出去的时候，重来不会重复输出。
|
||||
"""
|
||||
|
||||
def __init__(self, big: Any, small: Any) -> None:
|
||||
self.model, self._small = big, small
|
||||
|
||||
@property
|
||||
def max_tokens(self) -> Any:
|
||||
return getattr(self.model, "max_tokens", None)
|
||||
|
||||
async def astream(self, messages: list[Any]):
|
||||
started = False
|
||||
try:
|
||||
async for chunk in self.model.astream(messages):
|
||||
started = True
|
||||
yield chunk
|
||||
return
|
||||
except Exception as e:
|
||||
if started or self.model is self._small or not _budget_rejected(e):
|
||||
raise
|
||||
self.model = self._small
|
||||
async for chunk in self.model.astream(messages):
|
||||
yield chunk
|
||||
|
||||
|
||||
async def _op_model(session: AsyncSession, payload: GenerateIn) -> tuple[Any, str]:
|
||||
"""按操作流改图用的模型：额度见 COPILOT_MAX_TOKENS。"""
|
||||
spec = await copilot_model_spec(session, payload, max_tokens=COPILOT_MAX_TOKENS)
|
||||
big, model_id = await get_chat_model(session, spec)
|
||||
small, _ = await get_chat_model(session, spec.model_copy(update={"max_tokens": _FALLBACK_MAX_TOKENS}))
|
||||
return _BudgetFallback(big, small), model_id
|
||||
|
||||
|
||||
async def _iter_ops(model: Any, messages: list[Any]):
|
||||
"""流式读模型输出，按行切出操作。done 之后即停——后面就算有闲话也不要了。
|
||||
|
||||
@@ -923,11 +992,15 @@ async def _iter_ops(model: Any, messages: list[Any]):
|
||||
思考走 thinking_text()（Claude 4.6+ 的 thinking 块），不是"把非 JSON 行
|
||||
当成思考"：后者会把模型的 markdown 注释、协议跑偏时的乱输出一并当作思考
|
||||
展示出来，反而误导。不支持 thinking 的模型就只有心跳，不硬凑。
|
||||
|
||||
没等到 done 就因为额度用完停下的，抛 CopilotIncomplete：截断之前的操作照常转出去，
|
||||
由调用方决定整轮作废。
|
||||
"""
|
||||
from app.engine.state import message_text as _text
|
||||
from app.engine.state import thinking_text
|
||||
|
||||
buffer = ""
|
||||
cut = ""
|
||||
async for chunk in model.astream(messages):
|
||||
reasoning = thinking_text(chunk)
|
||||
if reasoning:
|
||||
@@ -941,9 +1014,19 @@ async def _iter_ops(model: Any, messages: list[Any]):
|
||||
yield op
|
||||
if op.get("op") == "done":
|
||||
return
|
||||
cut = _cut_off(chunk) or cut
|
||||
op = _parse_op_line(buffer)
|
||||
if op:
|
||||
yield op
|
||||
if op.get("op") == "done":
|
||||
return
|
||||
if cut:
|
||||
# 截断的流和说完了的流读起来一模一样：不在这里拦下，上面就会拿原图（或半张图）
|
||||
# 去自查、交付，界面上写「无需修改」
|
||||
limit = getattr(model, "max_tokens", None)
|
||||
raise CopilotIncomplete(
|
||||
"模型的输出超出了长度上限，修改没有写完",
|
||||
detail=f"finish_reason={cut}" + (f"，本轮输出上限 {limit} token" if limit else ""))
|
||||
|
||||
|
||||
# 多久没动静就补一拍心跳。3 秒是"用户开始怀疑是不是卡了"的量级。
|
||||
@@ -1553,8 +1636,7 @@ async def generate_stream(payload: GenerateIn, session: AsyncSession = Depends(g
|
||||
user = _user_message(payload, patch=False)
|
||||
|
||||
try:
|
||||
spec_ = await copilot_model_spec(session, payload)
|
||||
model, model_id = await get_chat_model(session, spec_)
|
||||
model, model_id = await _op_model(session, payload)
|
||||
except ProviderNotConfigured as e:
|
||||
raise HTTPException(400, _unconfigured(e)) from e
|
||||
|
||||
@@ -1598,6 +1680,14 @@ async def generate_stream(payload: GenerateIn, session: AsyncSession = Depends(g
|
||||
skipped_types.append(str(bad_type))
|
||||
if changed or kind in ("plan", "done"):
|
||||
yield f"data: {json.dumps(op, ensure_ascii=False)}\n\n"
|
||||
# 流正常结束，却没有 done、也没有一个操作能认（格式跑偏，每一行都被跳过）：
|
||||
# 往下走就是拿原图去自查、交付，界面上写「无需修改」——而模型什么都没说
|
||||
if all(o.get("op") == "plan" for o in seen_ops):
|
||||
raise CopilotIncomplete("模型没有给出任何修改，也没有说明原因",
|
||||
hint="画布未修改。可以重试，或把需求描述得更具体")
|
||||
except CopilotIncomplete as e:
|
||||
yield _sse({"op": "error", "message": f"助手本轮未完成：{e}", "hint": e.hint, "detail": e.detail})
|
||||
return
|
||||
except Exception as e: # noqa: BLE001
|
||||
reason, hint = explain_error(e)
|
||||
yield _sse({"op": "error", "message": f"助手本轮未完成：{reason}",
|
||||
@@ -1784,9 +1874,8 @@ async def assist_publish_fix(
|
||||
out["summary"] = "没有阻止发布的错误，无需交给助手"
|
||||
return out
|
||||
try:
|
||||
spec_ = await copilot_model_spec(
|
||||
chat, _ = await _op_model(
|
||||
session, GenerateIn(instruction="发布前自动修复", provider=provider, model=model))
|
||||
chat, _ = await get_chat_model(session, spec_)
|
||||
except ProviderNotConfigured as e:
|
||||
out["summary"] = _unconfigured(e)
|
||||
return out
|
||||
@@ -1959,9 +2048,8 @@ async def assist_upgrade(
|
||||
known = {_sig(e) for e in base_errors}
|
||||
out["warnings"] = _kept_code(spec)
|
||||
try:
|
||||
spec_ = await copilot_model_spec(
|
||||
chat, _ = await _op_model(
|
||||
session, GenerateIn(instruction="升级为可追溯结构", provider=provider, model=model))
|
||||
chat, _ = await get_chat_model(session, spec_)
|
||||
except ProviderNotConfigured as e:
|
||||
out.update(ok=False, summary=_unconfigured(e))
|
||||
return out
|
||||
|
||||
@@ -278,3 +278,25 @@ async def test_renaming_a_node_while_fixing_another_error_is_adopted(client, mon
|
||||
assert "复核团队（待拆成固定步骤）" in out["remaining"][0]["message"] or any(
|
||||
"复核团队（待拆成固定步骤）" in i["message"] for i in out["remaining"])
|
||||
assert {(c["node_id"], c["field"]) for c in out["changes"]} == {("team", "label"), ("done", "contract")}
|
||||
|
||||
|
||||
async def test_a_fix_cut_off_at_the_output_limit_is_not_adopted(client, monkeypatch):
|
||||
"""修正写到一半额度用完（finish_reason=length）：哪怕已经写出来的那几条能让 error 变少，也不采纳半截修改。"""
|
||||
from langchain_core.messages import AIMessageChunk
|
||||
|
||||
model = script(monkeypatch, [{"op": "plan", "summary": "给出具节点补契约"},
|
||||
{"op": "update_node", "id": "done", "config": {"contract": CONTRACT}}])
|
||||
lines = model.lines
|
||||
|
||||
async def cut_off(messages):
|
||||
model.calls.append(messages)
|
||||
for line in lines:
|
||||
yield AIMessageChunk(content=line + "\n")
|
||||
yield AIMessageChunk(content="", response_metadata={"finish_reason": "length"})
|
||||
|
||||
model.astream = cut_off
|
||||
wf = await create(client, two_exits())
|
||||
out = await assist(client, wf)
|
||||
assert out["assist"]["ok"] is False and "assist" not in out["applied"]
|
||||
assert out["assist"]["summary"].startswith("助手本轮未完成：") and "长度上限" in out["assist"]["summary"]
|
||||
assert out["graph"] == two_exits()
|
||||
|
||||
@@ -0,0 +1,295 @@
|
||||
"""Copilot 没说完就不能按「无需修改」交付。
|
||||
|
||||
2026-09 部署实例上，复杂的改图需求（并行算同比、环比，再加一个单独的 agent）连续两次得到
|
||||
「已检查，画布无需修改 · 自查通过」。按同一模型、同一张底图、同一句指令重放：输出额度 8192
|
||||
全部用在思考上（finish_reason=length），正文一个字都没有。读流时不看 finish_reason，也不要求
|
||||
收到 done，于是自查对着原图报「通过」，final 交回原图，前端比出来没变化就说「无需修改」。
|
||||
|
||||
这里守几件事：
|
||||
- 输出被截断（OpenAI 系的 finish_reason=length、Anthropic 的 stop_reason=max_tokens）要报错，
|
||||
不跑自查、不交 final；
|
||||
- 流正常结束却一个操作都没给、也没有 done（格式跑偏，全被跳过了）同样报错；
|
||||
- 模型明确说完了（有 done）但没改任何东西，照旧交付——那才是真的无需修改；
|
||||
- 改图的额度要够思考之后还写得下操作；服务方一开口就嫌额度太大时退回 8192 重来。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
|
||||
import pytest
|
||||
from httpx import ASGITransport, AsyncClient
|
||||
from langchain_core.messages import AIMessageChunk
|
||||
|
||||
from app.api import copilot as cp
|
||||
from app.api.copilot import _iter_ops
|
||||
from app.main import app
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
async def client():
|
||||
async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
|
||||
yield c
|
||||
|
||||
|
||||
def _node(nid: str, ntype: str, config: dict) -> dict:
|
||||
return {"id": nid, "type": ntype, "position": {"x": 0, "y": 0}, "data": {"label": nid, "config": config}}
|
||||
|
||||
|
||||
#: 一张本来就能过自查的图：截断时自查对着它报「通过」，正是那次的样子
|
||||
BASE = {
|
||||
"nodes": [
|
||||
_node("start", "input", {"fields": [{"name": "question"}]}),
|
||||
_node("answer", "llm", {"prompt": "{{ input.question }}", "assign_to": "answer"}),
|
||||
_node("out", "output", {"fields": [{"name": "答案", "value": "{{ vars.answer }}"}]}),
|
||||
],
|
||||
"edges": [{"source": "start", "target": "answer"}, {"source": "answer", "target": "out"}],
|
||||
}
|
||||
|
||||
|
||||
def _line(op: dict) -> AIMessageChunk:
|
||||
return AIMessageChunk(content=json.dumps(op, ensure_ascii=False) + "\n")
|
||||
|
||||
|
||||
def _thinking(text: str) -> AIMessageChunk:
|
||||
# OpenAI 兼容系（DeepSeek 经 LiteLLM）的思考走 reasoning_content
|
||||
return AIMessageChunk(content="", additional_kwargs={"reasoning_content": text})
|
||||
|
||||
|
||||
def _openai_cut() -> AIMessageChunk:
|
||||
return AIMessageChunk(content="", response_metadata={"finish_reason": "length"})
|
||||
|
||||
|
||||
def _anthropic_cut() -> AIMessageChunk:
|
||||
return AIMessageChunk(content=[], response_metadata={"stop_reason": "max_tokens"})
|
||||
|
||||
|
||||
class _Scripted:
|
||||
"""按剧本吐 chunk 的模型。each 是每次调用吐的那一串；调用次数多于剧本时重复最后一串。"""
|
||||
|
||||
def __init__(self, *each: list[AIMessageChunk], max_tokens: int | None = None) -> None:
|
||||
self.each = each
|
||||
self.calls = 0
|
||||
self.max_tokens = max_tokens
|
||||
|
||||
async def astream(self, messages):
|
||||
chunks = self.each[min(self.calls, len(self.each) - 1)]
|
||||
self.calls += 1
|
||||
for chunk in chunks:
|
||||
yield chunk
|
||||
|
||||
|
||||
async def _events(client, monkeypatch, model, base: dict | None = BASE) -> list[dict]:
|
||||
async def _model(*_a, **_k):
|
||||
return model, "mock-fast"
|
||||
|
||||
monkeypatch.setattr(cp, "get_chat_model", _model)
|
||||
out: list[dict] = []
|
||||
async with client.stream("POST", "/api/copilot/generate-stream",
|
||||
json={"instruction": "并行算同比和环比", "base_graph": base}) as r:
|
||||
async for line in r.aiter_lines():
|
||||
if line.startswith("data: "):
|
||||
out.append(json.loads(line[6:]))
|
||||
return out
|
||||
|
||||
|
||||
def _kinds(events: list[dict]) -> list[str]:
|
||||
return [e["op"] for e in events if e["op"] not in ("thinking", "heartbeat", "model")]
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 读流这一层
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
@pytest.mark.parametrize("cut", [_openai_cut, _anthropic_cut], ids=["finish_reason", "stop_reason"])
|
||||
async def test_a_cut_off_stream_raises(cut):
|
||||
model = _Scripted([_thinking("先想想怎么拆成并行分支……"), _line({"op": "plan", "summary": "拆成三支"}),
|
||||
AIMessageChunk(content='{"op":"add_node","node":{"id":"yo'), cut()])
|
||||
got: list[dict] = []
|
||||
with pytest.raises(cp.CopilotIncomplete) as info:
|
||||
async for op in _iter_ops(model, []):
|
||||
got.append(op)
|
||||
assert [o["op"] for o in got] == ["thinking", "plan"], "截断之前的操作照常转出去"
|
||||
assert "长度上限" in str(info.value)
|
||||
assert info.value.hint
|
||||
|
||||
|
||||
async def test_done_that_lands_exactly_at_the_limit_is_complete():
|
||||
"""最后一行 done 刚好写完（没来得及换行）就到了上限：模型已经说完了，不算截断。"""
|
||||
model = _Scripted([_line({"op": "plan", "summary": "x"}),
|
||||
AIMessageChunk(content='{"op":"done","explanation":"好了"}'), _openai_cut()])
|
||||
ops = [o async for o in _iter_ops(model, [])]
|
||||
assert [o["op"] for o in ops] == ["plan", "done"]
|
||||
|
||||
|
||||
async def test_a_normal_finish_is_not_a_cut():
|
||||
model = _Scripted([_line({"op": "plan", "summary": "x"}), _line({"op": "done", "explanation": "y"}),
|
||||
AIMessageChunk(content="", response_metadata={"finish_reason": "stop"})])
|
||||
ops = [o async for o in _iter_ops(model, [])]
|
||||
assert [o["op"] for o in ops] == ["plan", "done"]
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 画布改图：截断和空手而归都不能冒充「无需修改」
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
async def test_thinking_that_eats_the_whole_budget_is_an_error_not_unchanged(client, monkeypatch):
|
||||
"""就是截图里那一轮：只有思考、没有正文，额度用完。"""
|
||||
model = _Scripted([_thinking("Let me design the modification. "), _thinking("report_ans has 4 incoming edges."),
|
||||
_openai_cut()], max_tokens=8192)
|
||||
events = await _events(client, monkeypatch, model)
|
||||
|
||||
kinds = _kinds(events)
|
||||
assert "final" not in kinds, "没说完的一轮不能把原图当成结果交回去"
|
||||
assert "check" not in kinds, "自查对着原图报「通过」，是在替一个没发生的修改作证"
|
||||
[err] = [e for e in events if e["op"] == "error"]
|
||||
assert err["message"].startswith("助手本轮未完成：") and "长度上限" in err["message"]
|
||||
assert "画布未修改" in err["hint"]
|
||||
assert "length" in err["detail"] and "8192" in err["detail"]
|
||||
|
||||
|
||||
async def test_a_cut_in_the_middle_of_the_ops_is_not_delivered_as_done(client, monkeypatch):
|
||||
"""写到一半被截断：半张图不能当成品交付，更不能自动开跑。"""
|
||||
model = _Scripted([
|
||||
_line({"op": "plan", "summary": "加一支"}),
|
||||
_line({"op": "add_node", "node": {"id": "yoy", "type": "llm", "label": "同比",
|
||||
"config": {"prompt": "算同比", "assign_to": "yoy"}}}),
|
||||
_anthropic_cut(),
|
||||
])
|
||||
events = await _events(client, monkeypatch, model)
|
||||
kinds = _kinds(events)
|
||||
assert "final" not in kinds
|
||||
assert kinds[-1] == "error"
|
||||
|
||||
|
||||
async def test_output_that_parses_to_nothing_is_an_error(client, monkeypatch):
|
||||
"""格式跑偏（整段放进 markdown 围栏、JSON 分了行），每一行都被跳过：一个操作都没有，也没有 done。"""
|
||||
pretty = "```json\n" + json.dumps({"op": "add_node", "node": {"id": "yoy", "type": "llm"}}, indent=2) + "\n```\n"
|
||||
model = _Scripted([AIMessageChunk(content=pretty),
|
||||
AIMessageChunk(content="", response_metadata={"finish_reason": "stop"})])
|
||||
events = await _events(client, monkeypatch, model)
|
||||
kinds = _kinds(events)
|
||||
assert "final" not in kinds
|
||||
[err] = [e for e in events if e["op"] == "error"]
|
||||
assert err["message"].startswith("助手本轮未完成：") and "画布未修改" in err["hint"]
|
||||
|
||||
|
||||
async def test_a_model_that_says_done_without_changes_is_still_unchanged(client, monkeypatch):
|
||||
"""模型看过、明确收尾、一处不改：这才是真的「无需修改」，照旧交付。"""
|
||||
model = _Scripted([_line({"op": "plan", "summary": "现有的图已经满足"}),
|
||||
_line({"op": "done", "explanation": "不需要改"})])
|
||||
events = await _events(client, monkeypatch, model)
|
||||
final = events[-1]
|
||||
assert final["op"] == "final" and final["explanation"] == "不需要改"
|
||||
assert {n["id"] for n in final["graph"]["nodes"]} == {"start", "answer", "out"}
|
||||
assert "error" not in _kinds(events)
|
||||
|
||||
|
||||
async def test_a_cut_in_the_repair_round_is_reported(client, monkeypatch):
|
||||
"""自查交回去修的那一轮被截断：要说修正没完成，不能说「自查通过」。"""
|
||||
broken = [
|
||||
_line({"op": "plan", "summary": "加个循环"}),
|
||||
_line({"op": "add_node", "node": {"id": "lp", "type": "loop", "label": "循环",
|
||||
"config": {"mode": "while", "condition": "foo(vars.x)",
|
||||
"max_iterations": 2}}}),
|
||||
_line({"op": "add_edge", "edge": {"source": "answer", "target": "lp"}}),
|
||||
_line({"op": "add_edge", "edge": {"source": "lp", "target": "out", "sourceHandle": "done"}}),
|
||||
_line({"op": "done", "explanation": "加好了"}),
|
||||
]
|
||||
model = _Scripted(broken, [_thinking("修一下条件"), _openai_cut()])
|
||||
events = await _events(client, monkeypatch, model)
|
||||
checks = [e for e in events if e["op"] == "check"]
|
||||
assert checks[0]["status"] == "repairing"
|
||||
assert any(c["status"] == "error" and "长度上限" in c["message"] for c in checks), checks
|
||||
assert not any(c["status"] == "passed" for c in checks)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 额度
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
async def test_the_stream_asks_for_a_budget_that_leaves_room_after_thinking(client, monkeypatch):
|
||||
asked: list[int | None] = []
|
||||
|
||||
async def _model(_session, spec):
|
||||
asked.append(spec.max_tokens)
|
||||
return _Scripted([_line({"op": "plan", "summary": "x"}), _line({"op": "done", "explanation": "y"})]), "m"
|
||||
|
||||
monkeypatch.setattr(cp, "get_chat_model", _model)
|
||||
async with client.stream("POST", "/api/copilot/generate-stream",
|
||||
json={"instruction": "改一下", "base_graph": BASE}) as r:
|
||||
[line async for line in r.aiter_lines()]
|
||||
assert asked and asked[0] >= 32768, asked
|
||||
|
||||
|
||||
class _BudgetTooLarge(Exception):
|
||||
"""仿 openai / anthropic SDK 的 400：一发请求就被拒，还没吐出任何东西。"""
|
||||
|
||||
status_code = 400
|
||||
|
||||
|
||||
class _Rejects:
|
||||
def __init__(self, message: str) -> None:
|
||||
self.message = message
|
||||
self.calls = 0
|
||||
|
||||
async def astream(self, messages):
|
||||
self.calls += 1
|
||||
raise _BudgetTooLarge(self.message)
|
||||
yield # pragma: no cover - 让它成为异步生成器
|
||||
|
||||
|
||||
@pytest.mark.parametrize("message", [
|
||||
"Invalid max_tokens value, the valid range of max_tokens is [1, 8192]",
|
||||
"max_tokens: 32768 > 8192, which is the maximum allowed number of output tokens",
|
||||
"This model's maximum context length is 32768 tokens. However, you requested 40369 tokens "
|
||||
"(7601 in the messages, 32768 in the completion).",
|
||||
])
|
||||
async def test_a_rejected_budget_falls_back_to_the_old_one(client, monkeypatch, message):
|
||||
big = _Rejects(message)
|
||||
small = _Scripted([_line({"op": "plan", "summary": "x"}),
|
||||
_line({"op": "add_node", "node": {"id": "yoy", "type": "llm", "label": "同比",
|
||||
"config": {"prompt": "算同比", "assign_to": "yoy"}}}),
|
||||
_line({"op": "add_edge", "edge": {"source": "answer", "target": "yoy"}}),
|
||||
_line({"op": "add_edge", "edge": {"source": "yoy", "target": "out"}}),
|
||||
_line({"op": "done", "explanation": "加好了"})])
|
||||
asked: list[int | None] = []
|
||||
|
||||
async def _model(_session, spec):
|
||||
asked.append(spec.max_tokens)
|
||||
return (big if spec.max_tokens and spec.max_tokens > 8192 else small), "m"
|
||||
|
||||
monkeypatch.setattr(cp, "get_chat_model", _model)
|
||||
events: list[dict] = []
|
||||
async with client.stream("POST", "/api/copilot/generate-stream",
|
||||
json={"instruction": "加一支同比", "base_graph": BASE}) as r:
|
||||
async for line in r.aiter_lines():
|
||||
if line.startswith("data: "):
|
||||
events.append(json.loads(line[6:]))
|
||||
assert 8192 in asked
|
||||
assert big.calls == 1 and small.calls == 1
|
||||
final = events[-1]
|
||||
assert final["op"] == "final" and "yoy" in {n["id"] for n in final["graph"]["nodes"]}
|
||||
|
||||
|
||||
async def test_other_rejections_are_not_retried(client, monkeypatch):
|
||||
"""别的 400（模型名写错之类）照实报错，不拿小额度再撞一次。"""
|
||||
big = _Rejects("model 'no-such-model' does not exist")
|
||||
small = _Scripted([_line({"op": "done", "explanation": "不该到这里"})])
|
||||
|
||||
async def _model(_session, spec):
|
||||
return (big if spec.max_tokens and spec.max_tokens > 8192 else small), "m"
|
||||
|
||||
monkeypatch.setattr(cp, "get_chat_model", _model)
|
||||
events: list[dict] = []
|
||||
async with client.stream("POST", "/api/copilot/generate-stream",
|
||||
json={"instruction": "改一下", "base_graph": BASE}) as r:
|
||||
async for line in r.aiter_lines():
|
||||
if line.startswith("data: "):
|
||||
events.append(json.loads(line[6:]))
|
||||
assert small.calls == 0
|
||||
assert events[-1]["op"] == "error"
|
||||
|
||||
|
||||
def test_the_protocol_asks_not_to_draft_the_json_while_thinking():
|
||||
"""思考里把整段 JSON 先写一遍，额度就去了一半：那一轮重放时思考的结尾正在逐字写口径卡的表达式。"""
|
||||
assert "思考" in cp._STREAM_PROTOCOL and "不要在思考里" in cp._STREAM_PROTOCOL
|
||||
Reference in New Issue
Block a user