fix: per-node max_tokens + validation 502 guard + correct_jrxml output validity

- backend/llm.py: per-node max_tokens via get_llm(max_tokens=N), LLM_MAX_TOKENS env var (default 8192)
- agent/nodes.py: 5 generation nodes use max_tokens=32768, generate_skeleton retries at 65536
- agent/nodes.py: fix ns:field regex (<field → <[\w:]*field) to handle namespace prefixes
- agent/nodes.py: fix correct_jrxml never writing back to state["current_jrxml"]
- agent/nodes.py: correct_jrxml rejects non-JRXML output (no <jasperReport tag)
- agent/nodes.py: _strip_continuation_wrapper strips markdown/prefixes from continuation rounds
- agent/nodes.py: _extract_jrxml iterates multiple markdown code blocks, skips fragments
- agent/graph.py: route_after_validate skips correction loop when service_unavailable
- agent/graph.py: route_after_save skips validation for empty JRXML
- backend/validation.py: returns service_unavailable: True for ConnectError and HTTP 5xx
- Docs: CLAUDE.md v14 changelog, README.md LLM_MAX_TOKENS, .env.example LLM_MAX_TOKENS

backend/llm.py

@@ -156,8 +156,14 @@ class _LLMLoggingWrapper(_BaseLLM):
             raise
 
 
-def _build_raw_llm(caller: str = "") -> tuple[_BaseLLM, str, str]:
-    """构造原始 LLM 实例，返回 (实例, model名, backend名)。"""
+DEFAULT_MAX_TOKENS = int(os.getenv("LLM_MAX_TOKENS", "8192"))
+
+
+def _build_raw_llm(caller: str = "", max_tokens: int | None = None) -> tuple[_BaseLLM, str, str]:
+    """构造原始 LLM 实例，返回 (实例, model名, backend名)。
+
+    max_tokens: 覆盖默认输出 token 数。None 使用 LLM_MAX_TOKENS 环境变量或 8192。
+    """
     backend = os.getenv("LLM_BACKEND", "cloud")
     if backend == "local":
         from langchain_ollama import ChatOllama
@@ -183,18 +189,19 @@ def _build_raw_llm(caller: str = "") -> tuple[_BaseLLM, str, str]:
         base_url = os.getenv("ANTHROPIC_BASE_URL") or os.getenv("OPENAI_BASE_URL", "https://api.minimaxi.com/anthropic")
         model = os.getenv("LLM_MODEL", "MiniMax-M2.7")
         temperature = 0.1
-        max_tokens = 8192
+        _default_max_tokens = max_tokens if max_tokens is not None else DEFAULT_MAX_TOKENS
 
         client = Anthropic(api_key=api_key, base_url=base_url, timeout=120)
 
         class MiniMaxLLM(_BaseLLM):
             def __init__(self):
                 self._last_stop_reason = None
+                self._max_tokens = _default_max_tokens
 
             def invoke(self, prompt: str) -> Any:
                 resp = client.messages.create(
                     model=model,
-                    max_tokens=max_tokens,
+                    max_tokens=self._max_tokens,
                     temperature=temperature,
                     messages=[{"role": "user", "content": [{"type": "text", "text": prompt}]}],
                 )
@@ -208,7 +215,7 @@ def _build_raw_llm(caller: str = "") -> tuple[_BaseLLM, str, str]:
                 self._last_stop_reason = None
                 with client.messages.stream(
                     model=model,
-                    max_tokens=max_tokens,
+                    max_tokens=self._max_tokens,
                     temperature=temperature,
                     messages=[{"role": "user", "content": [{"type": "text", "text": prompt}]}],
                 ) as s:
@@ -250,9 +257,12 @@ def _build_raw_llm(caller: str = "") -> tuple[_BaseLLM, str, str]:
         return OpenAIWrapper(), model, f"cloud/openai/{model}"
 
 
-def get_llm(caller: str = "") -> _BaseLLM:
-    """返回带日志的 LLM 实例。caller 用于标识调用来源（如 generate、classify_intent）。"""
-    inner, model, backend = _build_raw_llm(caller)
+def get_llm(caller: str = "", max_tokens: int | None = None) -> _BaseLLM:
+    """返回带日志的 LLM 实例。caller 用于标识调用来源（如 generate、classify_intent）。
+
+    max_tokens: 覆盖默认输出 token 数。用于骨架生成等需要大量输出的节点。
+    """
+    inner, model, backend = _build_raw_llm(caller, max_tokens=max_tokens)
     return _LLMLoggingWrapper(inner, model=model, backend=backend, caller=caller)