878def _compile_react(cfg: Dict, overrides: Dict) -> Term:
879    """
880    type: react -> Loop(react_step, condition, max_steps)
881
882    v2 optimizations:
883      - Context passed through via closure over shared ctx
884      - Sliding window state compression
885      - Early termination on implicit signals
886      - Tool call caching
887    """
888    # Inject tool parameter docs into system prompt
889    cfg = dict(cfg)  # shallow copy to avoid mutating original
890    tool_docs = _generate_tool_schema_docs(cfg)
891    if tool_docs:
892        cfg["systemPrompt"] = cfg.get("systemPrompt", "") + tool_docs
893
894    # 可选:磁盘进度纪律(react.progressDiscipline)。把 PROGRESS.md 待办纪律注入 systemPrompt,
895    # 防多步任务"上下文截断→失忆→工具调用乱跳/重做"。默认 off,只对显式开启的 react agent 注入;
896    # simple 子智能体走 _compile_lam 不走这里,天然不受影响。见 docs/PROGRESS_DISCIPLINE_DESIGN.md。
897    if (cfg.get("react", {}) or {}).get("progressDiscipline"):
898        cfg["systemPrompt"] = cfg.get("systemPrompt", "") + _PROGRESS_DISCIPLINE_FRAGMENT
899
900    think = _compile_lam(cfg, "think", overrides=overrides)
901    tools = _compile_tools(cfg, overrides)
902
903    # S05: Enforce mcp.policy.mode at runtime
904    mcp_cfg = cfg.get("mcp", {})
905    mcp_policy_mode = mcp_cfg.get("policy", {}).get("mode", "auto")
906    if mcp_policy_mode == "disable":
907        tools = {"terminate": tools.get("terminate", Tool("terminate", fn=lambda x: x))}
908    elif mcp_policy_mode == "force":
909        pass  # All tools available, forced execution
910    # "auto" and "intelligence" are default behavior (LLM decides)
911
912    react_cfg = cfg.get("react", {})
913    max_steps = react_cfg.get("maxSteps", 10)
914    tool_timeout = react_cfg.get("toolTimeout", 30)
915    observation_enabled = react_cfg.get("observationEnabled", True)
916    verbose = react_cfg.get("verbose", False)
917    agent_name = cfg.get("name", cfg.get("agentId", "agent"))
918
919    # ── 原生 function-calling 车道(react.nativeToolCalls,阶段 2)──
920    # FC-capable provider(OpenAI 兼容:qwen/openai/deepseek…)+ 开关 → 用结构化
921    # tool_calls 直接执行,根治"文本求 JSON、正则抠"那一类 0 执行/格式飘 bug。
922    # 旧文本路径(下方)保留作 fallback:provider 不支持 FC 或开关 off 时走它。
923    # 见 docs/NATIVE_FUNCTION_CALLING_DESIGN.md。
924    if react_cfg.get("nativeToolCalls"):
925        _prov = getattr(think, "provider", None)
926        if (_prov is not None and hasattr(_prov, "chat_with_tools")
927                and getattr(_prov, "supports_function_calling", lambda: False)()):
928            try:
929                from lambdagent.agentruntime import cancel as _fc_cancel
930            except Exception:
931                _fc_cancel = None
932            # enforceLoop 计数模式(单体 agent 用):terminate 前 tool 须调够 minCount 次。
933            # 序列模式(orchestrator)FC 暂不支持——那些 agent 不设 nativeToolCalls。

60class ConversationLam(Term):
61    """
62    Lambda abstraction with conversation persistence.
63
64    Lambda semantics preserved:
65        ConversationLam(provider, prompt) = lambda x. provider(history + x)
66        apply() = beta-reduction with memory
67
68    The key difference from stateless Lam:
69        Lam:             each apply() is independent
70        ConversationLam: each apply() builds on all previous calls
71
72    L03: Now supports both legacy dict-based and ChatMessage-based flows.
73    """
74
75    def __init__(
76        self,
77        name: str,
78        provider: LLMProvider,
79        system_prompt: str,
80        max_history_tokens: int = 80000,
81        keep_recent_turns: int = 20,
82        output_parser: Callable[[str], Any] | None = None,
83        # L03: New unified interface params
84        model: str = "",
85        temperature: float = 0.0,
86        max_tokens: int = 4096,
87        # Hard character ceiling for the *rendered* request (system + history +
88        # new input). Providers like DashScope/qwen reject inputs over a fixed
89        # CHARACTER count (qwen: 258048 → "Range of input length should be
90        # [1, 258048]"). The token-based budget above is not enough on its own:
91        # for CJK text 1 token ≈ 1 char, so an 80k-"token" history estimated as
92        # 320k chars sails past the char limit. We keep a conservative margin
93        # below 258048 for the system prompt + the next user turn.
94        max_input_chars: int = 200000,
95    ):
96        super().__init__(name)
97        self.provider = provider
98        self.system_prompt = system_prompt
99        self.max_history_tokens = max_history_tokens
100        self.keep_recent_turns = keep_recent_turns
101        self.max_input_chars = max_input_chars
102        self.output_parser = output_parser or (lambda x: x)
103        # L03: Store model/temperature/max_tokens for chat_typed calls
104        self._model = model
105        self._temperature = temperature
106        self._max_tokens = max_tokens
107
108        # Conversation history (system message always first)
109        self.messages: List[dict] = [
110            {"role": "system", "content": system_prompt}
111        ]
112
113        # L03: ChatMessage-based history (parallel to dict-based for typed path)
114        self._typed_messages: List[ChatMessage] = []
115
116    # Expose model name for react_step logging
117    @property
118    def model(self) -> str:
119        return self._model or self.provider.model_name
120
121    # Expose _session_id for react_step session detection
122    @property
123    def _session_id(self):
124        return getattr(self.provider, '_session_id', None)
125
126    def apply(self, input: Any, ctx: Context | None = None) -> Any: