fix(ai): 分析类调用放开 max_tokens 限制 — 修复推理模型正文 0 字失败

根因(实测钉死): deepseek-v4-pro 等推理模型把思考 token 计入 max_tokens 预算。个股分析真实调用(62KB prompt, max_tokens=4500): usage completion=4500 全是 reasoning, finish=length, 正文 0 字, 流"正常"结束 → 前端兜底报「分析未返回内容」→ 表现为"经常失败" (推理长度随机, 逼近 4500 即失败)。与 180s 超时无关(实际 65s 结束)。改动: - ai_provider: generate/stream 的 max_tokens 支持 None=不传该参数, 输出上限交给服务端默认; _openai_kwargs None 时省略; codex 提示词 None 时跳过长度约束行 - 四个分析器(个股/复盘/财务/概念轮动)改传 max_tokens=None; 新增空正文守卫: 流结束但 0 个 delta → 明确报错+日志(原先静默, backend.log 无痕迹无法排查)。策略标题生成(max_tokens=8)等小任务保持原限制对照实测(真实 62KB prompt): 不传 → finish=stop, 推理 3311 token +正文 2407 字, 59s; max_tokens=16000 → 同样成功但推理撒欢 6239 token/93s — 不限制即最优。验证: 新增 kwargs/codex 单测, ai_provider 21 通过, 相关 174 通过; 线上接口复测: 修复前 0 delta 直接 done, 修复后 1682 delta/2467 字正常流式; ruff 相对 HEAD 无新增。
2026-09-12 15:34:16 +08:00 · 2026-08-17 17:47:46 +08:00
parent 6f0786abac
commit 0ee3aa8f1f
6 changed files with 77 additions and 15 deletions
@@ -216,10 +216,15 @@ async def generate_ai_text(
    messages: Sequence[Message],
    *,
    temperature: float | None = 0.3,
-    max_tokens: int = 3000,
+    max_tokens: int | None = 3000,
    timeout: float = 180.0,
 ) -> str:
-    """Return a complete AI response from the currently configured provider."""
+    """Return a complete AI response from the currently configured provider.
+
+    max_tokens=None 表示不传该参数(输出上限交给服务端默认) — 推理型模型
+    (如 deepseek reasoner 系)的思考 token 计入 max_tokens 预算, 显式限制
+    会挤占正文甚至全部吃光(正文 0 字 + finish=length), 长分析类调用应放开。
+    """
    if is_codex_cli_provider():
        return await _run_codex_cli(messages, max_tokens=max_tokens, timeout=max(timeout, 600.0))
    return await _run_openai_once(
@@ -234,13 +239,15 @@ async def stream_ai_text(
    messages: Sequence[Message],
    *,
    temperature: float | None = 0.5,
-    max_tokens: int = 4000,
+    max_tokens: int | None = 4000,
    timeout: float = 180.0,
 ) -> AsyncIterator[str]:
    """Yield text deltas from the configured provider.

    Codex CLI only exposes the final assistant message for this use case, so it
    yields one complete chunk after the command exits.
+
+    max_tokens=None 表示不限制输出(同 generate_ai_text 的说明)。
    """
    if is_codex_cli_provider():
        yield await _run_codex_cli(messages, max_tokens=max_tokens, timeout=max(timeout, 600.0))
@@ -259,7 +266,7 @@ async def _run_openai_once(
    messages: Sequence[Message],
    *,
    temperature: float | None,
-    max_tokens: int,
+    max_tokens: int | None,
    timeout: float,
 ) -> str:
    ai_key = secrets_store.get_ai_key()
@@ -295,7 +302,7 @@ async def _stream_openai(
    messages: Sequence[Message],
    *,
    temperature: float | None,
-    max_tokens: int,
+    max_tokens: int | None,
    timeout: float,
 ) -> AsyncIterator[str]:
    ai_key = secrets_store.get_ai_key()
@@ -402,9 +409,15 @@ def _openai_retry_kwargs(exc: Exception, kwargs: dict) -> dict | None:
    return None


-def _openai_kwargs(*, temperature: float | None, max_tokens: int) -> dict:
-    """Build OpenAI create() kwargs; optional parameters are omitted when empty."""
-    kwargs: dict = {"max_tokens": max_tokens}
+def _openai_kwargs(*, temperature: float | None, max_tokens: int | None) -> dict:
+    """Build OpenAI create() kwargs; optional parameters are omitted when empty.
+
+    max_tokens=None 时不传 — 由服务端默认上限管理(推理模型的思考 token 也
+    计入该参数预算, 限制会挤占正文, 见 stream_ai_text 文档)。
+    """
+    kwargs: dict = {}
+    if max_tokens is not None:
+        kwargs["max_tokens"] = max_tokens
    if temperature is not None:
        kwargs["temperature"] = temperature
    if current_ai_provider() == OPENAI_PROVIDER:
@@ -507,7 +520,7 @@ def _compact_error_text(text: str) -> str:
 async def _run_codex_cli(
    messages: Sequence[Message],
    *,
-    max_tokens: int,
+    max_tokens: int | None,
    timeout: float,
 ) -> str:
    prompt = _codex_prompt(messages, max_tokens=max_tokens)
@@ -640,14 +653,14 @@ def _make_writable_and_retry(
        raise exc_info[1] from None


-def _codex_prompt(messages: Sequence[Message], *, max_tokens: int) -> str:
+def _codex_prompt(messages: Sequence[Message], *, max_tokens: int | None) -> str:
    parts = [
        "You are Tick Stock Panel's local AI provider.",
        "This is a text-generation task. The working directory is intentionally empty.",
        "Use only the user-provided prompt content below; do not inspect or modify local files.",
        "Return only the final requested content; do not include execution logs.",
    ]
-    if max_tokens > 0:
+    if max_tokens:
        parts.append(f"Keep the final answer within about {max_tokens} output tokens.")
    for message in messages:
        role = message.get("role", "user")
@@ -365,18 +365,26 @@ async def analyze_rotation_stream(
            return

        user_prompt = _build_user_prompt(signals, overview, days, dates, focus, kind)
+        got_content = False
        async for delta in stream_ai_text(
            [
                {"role": "system", "content": _build_system_prompt(kind)},
                {"role": "user", "content": user_prompt},
            ],
            temperature=0.5,
-            max_tokens=4000,
+            # 不限制输出(推理模型思考 token 计入预算, 见 ai_provider.stream_ai_text)
+            max_tokens=None,
        ):
+            got_content = True
            yield json.dumps({"type": "delta", "content": delta}, ensure_ascii=False)

    except Exception as e:  # noqa: BLE001
        logger.exception("AI %s rotation analyze failed: %s", kind, e)
        yield json.dumps({"type": "error", "message": f"AI 轮动分析失败: {e}"}, ensure_ascii=False)
+        return

+    if not got_content:
+        logger.warning("AI %s rotation analyze ended with empty content", kind)
+        yield json.dumps({"type": "error", "message": "AI 未返回正文(输出被截断), 请重试"}, ensure_ascii=False)
+        return
    yield json.dumps({"type": "done"}, ensure_ascii=False)
@@ -178,14 +178,17 @@ async def analyze_financials_stream(
        from app.services.ai_provider import stream_ai_text

        user_prompt = _build_user_prompt(fins, symbol, focus)
+        got_content = False
        async for delta in stream_ai_text(
            [
                {"role": "system", "content": _SYSTEM_PROMPT},
                {"role": "user", "content": user_prompt},
            ],
            temperature=0.4,
-            max_tokens=4000,
+            # 不限制输出(推理模型思考 token 计入预算, 见 ai_provider.stream_ai_text)
+            max_tokens=None,
        ):
+            got_content = True
            yield json.dumps({"type": "delta", "content": delta}, ensure_ascii=False)

    except Exception as e:  # noqa: BLE001
@@ -193,4 +196,8 @@ async def analyze_financials_stream(
        yield json.dumps({"type": "error", "message": f"AI 分析失败: {e}"}, ensure_ascii=False)
        return

+    if not got_content:
+        logger.warning("AI financial analysis ended with empty content for %s", symbol)
+        yield json.dumps({"type": "error", "message": "AI 未返回正文(输出被截断), 请重试"}, ensure_ascii=False)
+        return
    yield json.dumps({"type": "done"}, ensure_ascii=False)
@@ -301,14 +301,17 @@ async def recap_market_stream(
        from app.services.ai_provider import stream_ai_text

        user_prompt = _build_user_prompt(overview, news or [], focus)
+        got_content = False
        async for delta in stream_ai_text(
            [
                {"role": "system", "content": _SYSTEM_PROMPT},
                {"role": "user", "content": user_prompt},
            ],
            temperature=0.5,
-            max_tokens=4500,
+            # 不限制输出(推理模型思考 token 计入预算, 见 ai_provider.stream_ai_text)
+            max_tokens=None,
        ):
+            got_content = True
            yield json.dumps({"type": "delta", "content": delta}, ensure_ascii=False)

    except Exception as e:  # noqa: BLE001
@@ -316,6 +319,10 @@ async def recap_market_stream(
        yield json.dumps({"type": "error", "message": f"AI 复盘失败: {e}"}, ensure_ascii=False)
        return

+    if not got_content:
+        logger.warning("AI market recap ended with empty content for %s", as_of_str)
+        yield json.dumps({"type": "error", "message": "AI 未返回正文(输出被截断), 请重试"}, ensure_ascii=False)
+        return
    yield json.dumps({"type": "done"}, ensure_ascii=False)


@@ -315,14 +315,18 @@ async def analyze_stock_stream(
        kline_tail = _clean_rows(df, _KLINE_KEEP_COLS)
        user_prompt = _build_user_prompt(kline_tail, fins, levels, close, symbol, focus,
                                         asset_type=repo.resolve_asset_type(symbol))
+        got_content = False
        async for delta in stream_ai_text(
            [
                {"role": "system", "content": _SYSTEM_PROMPT},
                {"role": "user", "content": user_prompt},
            ],
            temperature=0.5,
-            max_tokens=4500,
+            # 不限制输出: 推理模型(deepseek reasoner 系)思考 token 计入 max_tokens
+            # 预算, 固定上限会把正文挤光(实测 4500 全被推理吃掉 → 正文 0 字)。
+            max_tokens=None,
        ):
+            got_content = True
            yield json.dumps({"type": "delta", "content": delta}, ensure_ascii=False)

    except Exception as e:  # noqa: BLE001
@@ -330,4 +334,9 @@ async def analyze_stock_stream(
        yield json.dumps({"type": "error", "message": f"AI 分析失败: {e}"}, ensure_ascii=False)
        return

+    if not got_content:
+        # 流正常结束但一个正文块都没有(典型: 输出上限被思考吃光后静默截断)
+        logger.warning("AI stock analysis ended with empty content for %s", symbol)
+        yield json.dumps({"type": "error", "message": "AI 未返回正文(输出被截断), 请重试"}, ensure_ascii=False)
+        return
    yield json.dumps({"type": "done"}, ensure_ascii=False)
@@ -197,6 +197,24 @@ def test_openai_kwargs_include_configured_reasoning_effort(monkeypatch):
    assert "reasoning_effort" not in ai_provider._openai_kwargs(temperature=None, max_tokens=1000)


+def test_openai_kwargs_none_max_tokens_omits_limit():
+    """max_tokens=None → 不传上限(推理模型思考 token 计入预算, 分析类调用放开)。"""
+    kwargs = ai_provider._openai_kwargs(temperature=0.5, max_tokens=None)
+    assert "max_tokens" not in kwargs
+    assert kwargs.get("temperature") == 0.5
+
+    # 显式数值仍正常下发(策略标题生成等小任务依赖)
+    assert ai_provider._openai_kwargs(temperature=None, max_tokens=8) == {"max_tokens": 8}
+
+
+def test_codex_prompt_none_max_tokens_skips_length_hint():
+    prompt = ai_provider._codex_prompt([{"role": "user", "content": "hi"}], max_tokens=None)
+    assert "Keep the final answer" not in prompt
+
+    bounded = ai_provider._codex_prompt([{"role": "user", "content": "hi"}], max_tokens=300)
+    assert "Keep the final answer" in bounded
+
+
 def test_ai_settings_keep_provider_models_separate(monkeypatch):
    stored = {
        "ai_provider": "openai_compat",