from __future__ import annotations from dataclasses import dataclass from typing import Any from .manager import PlanModeManagerV2 from .models import PLAN_MODE_PLAN from .prompt_injector import build_plan_mode_prefix, inject_plan_context from .trace import emit_plan_agent_v2_trace @dataclass(frozen=True) class PlanAgentV2Decision: action: str # enter_plan | stay_plan | run_agent reply_text: str plan_state: dict[str, Any] system_prompt_override: str = "" def _is_confirm_text(text: str) -> bool: t = str(text or "").strip().lower() return t in {"确认", "确认计划", "同意", "通过", "approve", "approved", "confirm", "yes"} def _normalize_user_text(text: str) -> str: return " ".join(str(text or "").strip().lower().split()) def _is_low_signal_continue(text_norm: str) -> bool: t = str(text_norm or "").strip().lower() return t in { "继续", "继续啊", "继续吧", "可以", "好的", "好", "ok", "okay", "go on", "continue", } def _confirm_strategy(store: Any) -> str: try: raw = str(store.get_setting("AIA_EXPERT_PLAN_CONFIRM_STRATEGY") or "").strip().lower() except Exception: raw = "" if raw in {"auto", "strict", "off"}: return raw return "strict" def _last_user_text_norm_from_history(*, store: Any, session_id: str) -> str: """Most recent persisted user message (current turn is usually not persisted yet).""" try: msgs = store.get_messages(session_id=session_id, limit=120) except Exception: return "" for m in reversed(msgs): if str(getattr(m, "role", "") or "").strip().lower() == "user": return _normalize_user_text(str(getattr(m, "content", "") or "")) return "" def _agent_conversation_stall_suffix(*, lang: str) -> str: is_en = str(lang or "").startswith("en") if is_en: return ( "\n\n[Conversation stall guard — agent mode]\n" "The user's latest message matches their previous user message in this session.\n" "- Do not repeat your last assistant reply or restate \"I will now…\" boilerplate.\n" "- Make substantive progress: execute the next concrete tool step, produce new actionable output, " "or ask exactly one specific blocking question.\n" ) return ( "\n\n【对话停滞防护 · agent 模式】\n" "检测到用户本条输入与上一轮用户输入相同（会话已持久化部分）。\n" "- 禁止复述上一轮助手回复或重复「接下来我将…」式独白。\n" "- 必须给出实质进展：执行具体工具步骤、写出新的可执行结果，或只提一个关键追问。\n" ) def evaluate_for_expert_mode( *, store: Any, session_id: str, lang: str, requested_specialist: str, user_text: str, execution_mode: str = "agent", base_system_prompt: str, trace_id: str | None = None, parent_span_id: str | None = None, ) -> PlanAgentV2Decision: mgr = PlanModeManagerV2(store=store) st = mgr.load_state(session_id=session_id) txt = str(user_text or "").strip() txt_norm = _normalize_user_text(txt) exec_mode = str(execution_mode or "").strip().lower() if exec_mode not in {"agent", "plan"}: exec_mode = "plan" confirm_strategy = _confirm_strategy(store) if exec_mode == "agent" and st.mode != PLAN_MODE_PLAN: emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="plan_mode_bypassed", payload={"requested_mode": "agent", "plan_mode_state": str(st.mode or "")}, ) last_user_norm = _last_user_text_norm_from_history(store=store, session_id=session_id) stall = bool(txt_norm and last_user_norm and txt_norm == last_user_norm) override = "" if stall: emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="agent_mode_conversation_stall", payload={"reason": "repeated_user_message"}, ) base = str(base_system_prompt or "").strip() suffix = _agent_conversation_stall_suffix(lang=lang).strip() override = f"{base}\n\n{suffix}".strip() return PlanAgentV2Decision( action="run_agent", reply_text="", plan_state=st.to_dict(), system_prompt_override=override, ) if st.mode != PLAN_MODE_PLAN: entered = mgr.enter(session_id=session_id, owner_specialist=requested_specialist, force_new_plan=False) emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="plan_mode_entered", payload={"owner_specialist": entered.owner_specialist, "plan_id": entered.plan_id}, ) prefix = build_plan_mode_prefix(state=entered, lang=lang) return PlanAgentV2Decision( action="run_agent", reply_text="", plan_state=entered.to_dict(), system_prompt_override=f"{prefix}\n\n{str(base_system_prompt or '').strip()}".strip(), ) st = mgr.refresh_plan_content(session_id=session_id) st = mgr.update_loop_guard(session_id=session_id, user_text_norm=txt_norm) if _is_confirm_text(txt): if exec_mode != "agent" and confirm_strategy == "strict": emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="plan_mode_confirm_blocked", payload={ "reason": "execution_mode_not_agent", "requested_mode": exec_mode, "confirm_strategy": confirm_strategy, }, ) blocked_reply = ( "Plan is ready. Please switch to agent mode, then confirm to execute." if str(lang or "").startswith("en") else "计划已就绪。请先切换到 agent 模式，再回复“确认”开始执行。" ) return PlanAgentV2Decision( action="stay_plan", reply_text=blocked_reply, plan_state=st.to_dict(), system_prompt_override="", ) if exec_mode != "agent" and confirm_strategy == "auto": emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="plan_mode_confirm_auto_switched", payload={"from_mode": exec_mode, "to_mode": "agent", "confirm_strategy": confirm_strategy}, ) confirmed = mgr.confirm(session_id=session_id) emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="plan_mode_confirmed", payload={ "plan_id": confirmed.plan_id, "plan_confirmed": bool(confirmed.plan_confirmed), "confirm_strategy": confirm_strategy, }, ) next_system = inject_plan_context(base_system=base_system_prompt, state=confirmed, lang=lang) reply = mgr.build_approved_execution_message(state=confirmed, lang=lang) return PlanAgentV2Decision( action="run_agent", reply_text=reply, plan_state=confirmed.to_dict(), system_prompt_override=next_system, ) if _is_low_signal_continue(txt_norm): emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="plan_mode_active", payload={"plan_id": st.plan_id, "plan_path": st.plan_path, "loop_guard": "low_signal_continue"}, ) low_signal_reply = ( "Plan mode detected a low-information continuation. " "Please provide concrete plan adjustments, or switch to agent mode and reply 'confirm' to execute." if str(lang or "").startswith("en") else "检测到低信息续写（如“继续/可以”）。请给出具体计划修改点，或切换到 agent 模式后回复“确认”直接执行。" ) return PlanAgentV2Decision( action="stay_plan", reply_text=low_signal_reply, plan_state=st.to_dict(), system_prompt_override="", ) if int(st.plan_loop_count or 0) >= 2: emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="plan_mode_active", payload={"plan_id": st.plan_id, "plan_path": st.plan_path, "loop_guard": "hard_block"}, ) anti_loop_reply = ( "I am in plan mode. I will only output a concise executable plan. " "If you want me to execute, switch to agent mode and reply 'confirm'." if str(lang or "").startswith("en") else "当前为 plan 模式，我只输出可执行计划。若要开始执行，请切换到 agent 模式并回复“确认”。" ) return PlanAgentV2Decision( action="stay_plan", reply_text=anti_loop_reply, plan_state=st.to_dict(), system_prompt_override="", ) prefix = build_plan_mode_prefix(state=st, lang=lang) anti_loop_suffix = ( "\n\n[Anti-loop guard]\n" "- Do not repeat the previous response.\n" "- If user asks similarly, refine with more concrete steps, checks, and fallback.\n" "- Keep output as plan only; do not pretend execution is complete." ) emit_plan_agent_v2_trace( store=store, session_id=session_id, trace_id=trace_id, parent_span_id=parent_span_id, event_type="plan_mode_active", payload={"plan_id": st.plan_id, "plan_path": st.plan_path, "loop_count": int(st.plan_loop_count or 0)}, ) return PlanAgentV2Decision( action="run_agent", reply_text="", plan_state=st.to_dict(), system_prompt_override=f"{prefix}{anti_loop_suffix}\n\n{str(base_system_prompt or '').strip()}".strip(), ) __all__ = ["PlanAgentV2Decision", "evaluate_for_expert_mode"]