关闭多模态时，analyze_screen/capture_screen 不再出现在工具列表里。

Source/WulaFallenEmpire/WulaAI_DevDocs/qwen_agent/QWEN_AGENT_3PHASE_PROTOCOL.md

@@ -0,0 +1,80 @@
+Qwen-Agent 3-Phase Tool Calling Protocol (Design + Implementation Plan)
+
+Goal
+- Preserve the existing three-phase workflow: Query -> Action -> Reply.
+- Provide a strict, parseable tool-call format compatible with Qwen-Agent style.
+- Allow explicit switching between native tool calling and template parsing.
+
+Scope
+- Query phase: only query tools (get_*/search_*/analyze_*/recall_memories)
+- Action phase: only action tools (spawn_resources, send_reinforcement, call_bombardment, modify_goodwill, call_prefab_airdrop, set_overwatch_mode, remember_fact)
+- Reply phase: natural language only, no tool calls
+
+Tool Calling Formats
+Native Mode (useNativeToolApi = true)
+- Use provider-native tool calls (OpenAI/DeepSeek tool_calls, Gemini functionCall)
+- No text template parsing, only validation + execution
+
+Template Mode (useNativeToolApi = false)
+- Use Qwen-Agent style tool-call tags in text:
+  <tool_call>
+  {"name":"tool_name","arguments":{...}}
+  </tool_call>
+- Tool results are injected as tool-role messages (not visible to the user)
+- No natural language is allowed in tool phases
+- If no tool is needed: output exactly {"tool_calls":[]}
+
+Three-Phase Flow (Template Mode)
+Phase 1: Query
+- Prompt includes tool list and Qwen tool-call template
+- Model returns one or more <tool_call> blocks
+- Parser extracts tool calls -> validator -> tool execution
+- Tool results stored and appended to history
+- One retry max if tool parsing/validation fails
+
+Phase 2: Action
+- Same as Query, but only action tools allowed
+- One retry max if tool parsing/validation fails
+
+Phase 3: Reply
+- Natural language only
+- If tool-call tags or JSON are detected, one retry is triggered
+
+Validation and Guardrails
+- ToolCallValidator ensures:
+  - Tool exists
+  - Required parameters present
+  - Tool is valid for current phase
+- Failures are returned to the model as tool results for self-correction
+- No parameter guessing or defName guessing
+
+Implementation Checklist (Files)
+1) WulaFallenEmpireSettings.cs
+   - Add useNativeToolApi flag (default true for backward compatibility)
+2) WulaFallenEmpireMod.cs
+   - Expose a UI checkbox to toggle useNativeToolApi
+3) AIIntelligenceCore.cs
+   - Select native loop vs template loop based on useNativeToolApi
+   - Template loop uses Qwen tool-call instructions in Query/Action
+   - Reply phase unchanged
+4) JsonToolCallParser.cs
+   - Add parsing for <tool_call>...</tool_call>
+   - Convert parsed calls to tool_calls list for execution
+5) ToolCallValidator.cs
+   - Enforce phase/tool rules and required args
+
+Rollout Plan
+1) Implement parser + settings toggle
+2) Switch flow selection by useNativeToolApi
+3) Add template tool-call instructions
+4) Validate in-game with:
+   - Item request (Query -> search_thing_def)
+   - Action call (spawn_resources)
+   - Reply text only
+
+Testing Checklist
+- Query phase outputs <tool_call> blocks
+- Action phase outputs <tool_call> blocks
+- Reply phase never outputs tool calls
+- Tool results are linked to tool_call_id (native) or injected safely (template)
+- One retry only per phase when parsing fails