Feat/3.0.0 beta1

src/backend/bisheng/common/services/llm_error_classifier.py

@@ -116,6 +116,33 @@ class ErrorType(str, enum.Enum):
     "限流",
 )
 
+# Transient TRANSPORT failures a provider reports from ITS OWN upstream fetch,
+# wrapped in an otherwise non-transient status. When an endpoint fetches something
+# on our behalf — staging a data-URI image into object storage, a sandbox pulling a
+# file, a RAG fetcher — its I/O error surfaces as a 4xx even though the REQUEST was
+# fine; only their fetch of it failed. Retrying is exactly right, and a plain 400
+# otherwise DEGRADEs, which on the main graph means the whole run dies.
+#
+# Measured on 180 (2026-08-09, session 649ba617): a deck screenshot handed to the
+# model came back as
+#   400 {"code":"DOWNLOAD_FAILED", ... "read tcp 10.0.2.60:46976->10.86.10.104:6000:
+#        read: connection reset by peer"}
+# — both endpoints inside the vendor's own network — and killed a 21-minute run
+# with zero retries.
+#
+# These are Go/POSIX transport phrases, not vendor error codes, so this stays
+# vendor-agnostic like the other flat signature sets. DELIBERATELY transport-only:
+# a fetch that failed with 404 / bad URL / unsupported format is a genuine client
+# error and MUST keep degrading rather than burn three backoffs on a sure failure.
+_TRANSIENT_TRANSPORT_SIGNATURES: tuple[str, ...] = (
+    "connection reset",
+    "connection refused",
+    "broken pipe",
+    "i/o timeout",
+    "unexpected eof",
+    "no route to host",
+)
+
 # Content-moderation / safety-guardrail codes across mainstream vendors:
 # Aliyun DashScope (data_inspection_failed / inappropriate_content),
 # Baidu Qianfan (336003/336005), Zhipu (1301), MiniMax (2013),
@@ -231,6 +258,16 @@ def _is_rate_limit(exc: BaseException) -> bool:
     return any(sig in text for sig in _RATE_LIMIT_SIGNATURES)
 
 
+def _is_transient_transport(exc: BaseException) -> bool:
+    """A transport-level failure the PROVIDER hit on its own upstream fetch.
+
+    Checked only after the FAIL_FAST branches, so a quota/auth error carrying such
+    wording in a nested message can never be turned into a retry.
+    """
+    text = _exc_text(exc)
+    return any(sig in text for sig in _TRANSIENT_TRANSPORT_SIGNATURES)
+
+
 def _is_task_aborted(exc: BaseException) -> bool:
     """The task was aborted by a framework/loop guard, not a provider error.
 
@@ -300,6 +337,8 @@ def classify_behavior(exc: BaseException) -> Behavior:
         return Behavior.RETRYABLE
     if isinstance(exc, (TimeoutError, ConnectionError)):
         return Behavior.RETRYABLE
+    if _is_transient_transport(exc):  # provider's own upstream fetch blew up mid-flight
+        return Behavior.RETRYABLE
 
     # 3) DEGRADABLE — default bucket: content filter, plain 400, anything else.
     return Behavior.DEGRADABLE
@@ -351,6 +390,10 @@ def label_error(exc: BaseException) -> ErrorType:
     status = _status_code(exc)
     if isinstance(exc, openai.InternalServerError) or (status is not None and 500 <= status < 600):
         return ErrorType.SERVICE_UNAVAILABLE
+    # Retries exhausted on the provider's own upstream fetch: "service busy" is the
+    # honest story for the user, not the generic unknown card.
+    if _is_transient_transport(exc):
+        return ErrorType.SERVICE_UNAVAILABLE
     return ErrorType.UNKNOWN
 
 

src/backend/bisheng/initdb_config.yaml

@@ -104,14 +104,18 @@ linsight:
   tool_buffer: 100000
   # 单个任务最大执行步骤数（LangGraph recursion_limit），防止死循环。
   # ⚠️ 这里的“步骤”是 LangGraph super-step，不是模型轮次：一轮模型调用约消耗 4 个 super-step
-  #   （model → 工具循环熔断器 after_model → TodoList after_model → tools），
-  #   所以 500 ≈ 115 轮。它是跑飞时的保险丝，真正的业务上限请调 max_model_turns。
-  #   若本值低于 max_model_turns 所需，后端会自动抬高并打 warning 日志。
-  max_steps: 500
+  #   （model → 工具循环熔断器 after_model → TodoList after_model → tools）。
+  #   另外「只产出 write_todos」的空转轮会退还预算（每个预算桶最多退 10 轮），
+  #   所以实际模型调用可能超过 max_model_turns，保险丝要留出这部分余量：
+  #   (115 + 10) × 4 + 20 = 520，取 600 留头。
+  #   它是跑飞时的保险丝，真正的业务上限请调 max_model_turns。
+  #   若本值低于所需，后端会自动抬高并打 warning 日志。
+  max_steps: 600
   # 主图轮次预算：一次任务运行最多允许多少次模型调用，到顶前会提示模型收尾（软着陆）。
   # ask_user 打断恢复后重新计数（中间件随 agent 重建）。
   max_model_turns: 115
-  # 子代理（researcher）自己的轮次预算，独立计数（子图跑自己的 Pregel 循环）
+  # 子代理（researcher）轮次预算。⚠️ 按「每次 task 调用」独立计数，不是整次任务共享：
+  # 主图并行委派 2 个子代理时，每个各得完整一份，互不抢额度。
   max_model_turns_subagent: 30
   # 距离预算耗尽还剩多少轮时开始提示模型收尾；最后 2 轮只保留写文件/导出工具，
   # 归零时不再提供任何工具，模型只能输出文本，图正常结束

src/backend/bisheng/linsight/domain/services/agent_factory.py

@@ -157,12 +157,12 @@
 
 - ask_user(reason, questions)：第 0 步澄清；整个会话最多调用一次。
 - write_todos(todos)：维护有编号的待办清单；只翻转 status，不改写已有文案。
-__KB_TOOL_LINE__- write_file / read_file / edit_file / ls：工作区文件工具；交付物写 output/，中间产物写 scratch/。
+__KB_TOOL_LINE__- write_file / read_file / edit_file / ls：工作区文件工具；交付物写 output/，中间产物写 scratch/（工具回显的 /output/x 与 output/x 是同一个文件）。
 - export_docx(source_path, dest_path)：把 output/ 下的 markdown 转 Word(.docx)，必须在对应 .md 写好之后。
 - export_pdf(source_path, dest_path)：把 output/ 下的 markdown 转 PDF，必须在对应 .md 写好之后。
 - task(description, subagent_type="general-purpose")：把独立、可隔离、较重的调研子任务委派给子代理。description 必须自包含——子代理看不到你的对话历史与上下文，只能读到 description，因此完成该子任务所需的全部背景、目标、约束与必要标识都要写进去；不得委派最终交付物撰写，也不得委派“问用户/澄清”。
 
-# 风格
+__PATH_NAMESPACE_LINE__# 风格
 
 - 简洁，工具调用之间不加多余解释性文字。
 - 不要凭空编造事实；交付内容应基于检索到的资料/依据。
@@ -320,10 +320,7 @@ def _build_linsight_system_prompt(
                 "此时**禁止**调用 export_docx / export_pdf 从 markdown 派生同一份交付物——"
                 "那条路径丢掉技能规定的全部排版细节，等于没用技能；"
                 "只有代码执行连续失败、确实产不出成品时，才允许退回 export_docx / export_pdf 兜底，"
-                "并在收尾里说明退化。3a 的 markdown 规范源仍然要写（界面预览依赖它）。"
-                "注意：代码执行器直接写在工作区本地目录，它生成的文件**不会**出现在 ls / glob 的结果里"
-                "（那两个工具读的是对象存储视图）。只要执行返回 exitcode 0 且日志显示写成功，"
-                "就视为交付物已产出，继续下一步；不要反复 ls / glob 找它，也不要因为“找不到”而重新生成。\n"
+                "并在收尾里说明退化。3a 的 markdown 规范源仍然要写（界面预览依赖它）。\n"
             )
         else:
             # No executor bound: the skill's route cannot run, so banning the
@@ -337,12 +334,38 @@ def _build_linsight_system_prompt(
     else:
         skill_exec_line = ""
         skill_deliverable_line = ""
+
+    # The two path namespaces only need explaining when a code interpreter is bound —
+    # that is the ONLY tool with a cwd. Gated on ``has_code_interpreter`` ALONE:
+    # this used to ride along inside ``skill_deliverable_line``, so a run that bound
+    # the executor but selected no skill (exactly the 180 session) never saw it and
+    # burned three model round-trips rediscovering the mapping. Keeping the
+    # ``has_code_interpreter`` gate preserves the prompt↔tool lockstep this module
+    # enforces everywhere else (a bound-tool name must never appear otherwise).
+    path_namespace_line = ""
+    if has_code_interpreter:
+        path_namespace_line = (
+            "# 工作区路径（同一份文件的两种写法）\n\n"
+            "- 文件工具（ls / read_file / write_file / edit_file / export_*）用**带前导斜杠**的工作区路径："
+            "/output/x.md、/scratch/x.png、/uploads/x.xlsx、/skills/<name>/SKILL.md；不带斜杠的相对写法等价。\n"
+            "- bisheng_code_interpreter 的**当前工作目录就是工作区根**，代码里一律用**去掉前导斜杠**的相对路径："
+            'open("skills/<name>/assets/t.html")、plt.savefig("output/chart.png")。'
+            '写成 open("/skills/…") / open("/output/…") 会落到容器根目录——读必然 FileNotFoundError，'
+            "写会被丢弃、进不了交付。\n"
+            "- 反向同理：**不要**把代码里看到的宿主机绝对路径（形如 /root/.cache/…/<8位任务号>/output/a.png）"
+            "传给 read_file / edit_file，去掉前缀只传 output/a.png。\n"
+            "- 执行器直接写本地工作目录，它生成的文件**不会**出现在 ls / glob 的结果里"
+            "（那两个工具读的是对象存储视图）。只要执行返回 exitcode 0 且日志显示写成功，"
+            "就视为交付物已产出，继续下一步；不要反复 ls / glob 找它，也不要因为“找不到”而重新生成。\n\n"
+        )
+
     return (
         _LINSIGHT_SYSTEM_PROMPT_TEMPLATE_ZH.replace("__KB_EXEC_LINE__", exec_line)
         .replace("__KB_TOOL_LINE__", tool_line)
         .replace("__KB_DELEGATE_LINE__", delegate_line)
         .replace("__SKILL_EXEC_LINE__", skill_exec_line)
         .replace("__SKILL_DELIVERABLE_LINE__", skill_deliverable_line)
+        .replace("__PATH_NAMESPACE_LINE__", path_namespace_line)
     )