docs: Clarify Gemini CLI marker system and analysis_source assignment logic

- Add explicit analysis_source assignment rules in plan.md (manual/gemini/auto-detected)
- Enhance GEMINI_CLI_REQUIRED marker documentation in code-developer.md and execute.md
- Specify fixed analysis_source="gemini" setting for plan-deep.md
- Add mapping rules from analysis_source values to CLI markers

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

.claude/agents/code-developer.md

@@ -37,12 +37,18 @@ You are a code execution specialist focused on implementing high-quality, produc
 ```
 IF context sufficient for implementation:
     → Proceed with execution
-ELIF context insufficient:
-    → Use Gemini CLI for codebase analysis
+ELIF context insufficient OR task has [GEMINI_CLI_REQUIRED] marker:
+    → Execute Gemini CLI for codebase analysis (MANDATORY)
     → Extract patterns and conventions
     → Proceed with execution
 ```
 
+**Gemini CLI Marker System**:
+- **[GEMINI_CLI_REQUIRED]**: Mandatory analysis flag
+  - **Trigger**: Auto-added when task.analysis_source = "gemini" or scope > 3 files
+  - **Action**: MUST run Gemini CLI first to gather context
+  - **Purpose**: Ensures code aligns with existing patterns
+
 **Gemini CLI Usage Standards**:
 Follow unified Gemini CLI guidelines: @~/.claude/workflows/gemini-unified.md
 

.claude/commands/workflow/execute.md

@@ -13,9 +13,6 @@ examples:
 Coordinates multiple agents for executing existing workflow tasks through automatic discovery and intelligent task orchestration. Analyzes workflow folders, checks task statuses, and coordinates agent execution based on discovered plans.
 
 ## Core Principles
-  
-**Session Management:** @~/.claude/workflows/workflow-architecture.md
-**Agent Orchestration:** @~/.claude/workflows/agent-orchestration-patterns.md
 
 ## Execution Philosophy
 
@@ -26,7 +23,15 @@ The intelligent execution approach focuses on:
 - **Dynamic task orchestration** - Coordinate based on discovered task relationships
 - **Progress tracking** - Update task status after agent completion
 
-**IMPORTANT**: Gemini context analysis is automatically applied based on discovered task scope and requirements.
+**GEMINI_CLI_REQUIRED Marker**:
+- **Purpose**: Forces agent to analyze existing codebase before implementation
+- **Auto-trigger**: When task.analysis_source = "gemini" OR scope > 3 files
+- **Agent Action**: MUST execute Gemini CLI as first step
+
+**analysis_source 到标记的映射**:
+- **"gemini"** → 添加 [GEMINI_CLI_REQUIRED]
+- **"auto-detected"** + scope > 3 files → 添加 [GEMINI_CLI_REQUIRED]
+- **"manual"** → 不添加标记
 
 ## Execution Flow
 
@@ -60,6 +65,8 @@ Workflow Discovery:
 - [ ] **TASK-002**: [Agent: code-developer] [GEMINI_CLI_REQUIRED] Implement auth logic (impl-1.2)  
 - [ ] **TASK-003**: [Agent: code-review-agent] Review implementations
 - [ ] **TASK-004**: Update task statuses and session state
+
+**Marker Legend**: [GEMINI_CLI_REQUIRED] = Agent must analyze codebase context first
 ```
 
 ### 3. Agent Context Assignment
@@ -130,7 +137,7 @@ For each executable task:
 - **Risk Awareness**: Alert agents to implementation.context_notes.risks before execution
 - **Workflow Integration**: Include session state and IMPL_PLAN.md context
 - **Scope Focus**: Direct agents to specific files from implementation.files[].path
-- **Gemini Flags**: Auto-add [GEMINI_CLI_REQUIRED] when analysis_source is "gemini"
+- **Gemini Marker**: Auto-add [GEMINI_CLI_REQUIRED] when analysis_source = "gemini"
 
 ### 4. Agent Execution & Progress Tracking
 

.claude/commands/workflow/plan-deep.md

@@ -85,6 +85,7 @@ Task(action-planning-agent):
     - Skip PRD processing (no PRD provided)
     - Skip session inheritance (standalone planning)
     - Force GEMINI_CLI_REQUIRED flag = true
+    - Set analysis_source = "gemini" (深度分析固定值)
     - Generate hierarchical task decomposition
     - Create detailed IMPL_PLAN.md with subtasks
     - Generate TODO_LIST.md for tracking
@@ -120,6 +121,7 @@ def process_plan_deep_command(input):
     
     MANDATORY FLAGS:
     - GEMINI_CLI_REQUIRED = true
+    - analysis_source = "gemini" (固定设置)
     - FORCE_PARALLEL_ANALYSIS = true
     - SKIP_PRD = true
     - SKIP_SESSION_INHERITANCE = true

.claude/commands/workflow/plan.md

@@ -83,6 +83,16 @@ The command automatically detects input type:
 ### Implementation Field Requirements
 ⚠️ **CRITICAL**: All generated tasks must include detailed implementation guidance
 
+**analysis_source 赋值规则**:
+- **"manual"**: 用户提供完整实现细节（包含具体文件、代码片段）
+- **"gemini"**: 信息不足，需要 Gemini 分析（缺少文件路径或代码上下文）  
+- **"auto-detected"**: 系统自动推断实现细节（基于模式识别）
+
+**判断流程**:
+1. **IF** 用户提供文件路径 + 代码片段 → "manual"
+2. **ELIF** 系统能推断实现位置 → "auto-detected"
+3. **ELSE** → "gemini" (需要深度分析)
+
 **Auto-fill Strategy**:
 1. **Sufficient Information**: Auto-fill implementation field based on user input and project context
 2. **Insufficient Information**: Mark analysis_source as "gemini" and prompt: