diff --git a/.claude/agents/conceptual-planning-agent.md b/.claude/agents/conceptual-planning-agent.md
index a89ae158..b4afc473 100644
--- a/.claude/agents/conceptual-planning-agent.md
+++ b/.claude/agents/conceptual-planning-agent.md
@@ -308,3 +308,37 @@ When analysis is complete, ensure:
 - **Relevance**: Directly addresses user's specified requirements
 - **Actionability**: Provides concrete next steps and recommendations
 
+## Output Size Limits (Context Overflow Prevention)
+
+**⚠️ CRITICAL**: To prevent API context overflow in brainstorm workflows, enforce these limits:
+
+### Document Size Limits
+- `analysis.md` (main): **< 3000 words** (~12KB)
+- `analysis-*.md` (sub-docs): **< 2000 words each** (~8KB)
+- Maximum 5 sub-documents per role
+- **Total per-role**: **< 15000 words** (~60KB)
+
+### Strategies for Staying Within Limits
+1. **Be concise**: Focus on key insights, not exhaustive detail
+2. **Use bullet points**: More efficient than prose paragraphs
+3. **Reference don't repeat**: Use `@filename.md` references instead of duplicating content
+4. **Prioritize**: Include top 3-5 recommendations per section, not all possible ones
+5. **Defer details**: Mark "further analysis needed" for complex areas instead of expanding
+
+### Self-Check Before Output
+```
+Before writing files, verify:
+□ Main analysis.md < 3000 words
+□ Each sub-document < 2000 words
+□ Total sub-documents ≤ 5
+□ No duplicated content across files
+□ References used where appropriate
+```
+
+### If Limit Exceeded
+If analysis naturally exceeds limits:
+1. Split into essential vs nice-to-have content
+2. Move nice-to-have to a separate `analysis-appendix.md` (counts toward 5 sub-doc limit)
+3. Use executive summary style for main document
+4. Note "detailed analysis available on request" for deferred topics
+
diff --git a/.claude/commands/workflow/brainstorm/auto-parallel.md b/.claude/commands/workflow/brainstorm/auto-parallel.md
index 62a77be1..14183836 100644
--- a/.claude/commands/workflow/brainstorm/auto-parallel.md
+++ b/.claude/commands/workflow/brainstorm/auto-parallel.md
@@ -424,6 +424,48 @@ CONTEXT_VARS:
 - **Agent execution failure**: Agent-specific retry with minimal dependencies
 - **Template loading issues**: Agent handles graceful degradation
 - **Synthesis conflicts**: Synthesis highlights disagreements without resolution
+- **Context overflow protection**: See below for automatic context management
+
+## Context Overflow Protection
+
+**⚠️ IMPORTANT**: Brainstorm workflows can generate substantial content across multiple roles. To prevent API context limit issues:
+
+### Per-Role Output Limits
+- Each `analysis.md` should be **< 3000 words**
+- Sub-documents (`analysis-*.md`) should be **< 2000 words each**
+- Maximum 5 sub-documents per role
+- Total per-role content: **< 15000 words**
+
+### Synthesis Phase Protection
+Before Phase 3 (synthesis), check total analysis size:
+```javascript
+// Check combined analysis size before synthesis
+const analysisFiles = Glob(".workflow/active/WFS-{session}/.brainstorming/*/analysis*.md");
+let totalSize = 0;
+for (const file of analysisFiles) {
+  totalSize += getFileSize(file);
+}
+
+// If total > 100KB, use summary mode
+if (totalSize > 100 * 1024) {
+  console.log("⚠️ Large context detected. Using summary mode for synthesis.");
+  // Synthesis will read only analysis.md (index) files, not sub-documents
+  useSummaryMode = true;
+}
+```
+
+### Recovery from Context Overflow
+If API returns empty response or timeout:
+1. **Identify overflow**: Check last request size in logs
+2. **Reduce scope**: Re-run synthesis with fewer roles (use `--count 2`)
+3. **Use summary mode**: Pass `--summary-only` to synthesis command
+4. **Manual synthesis**: Combine key insights from each role's `analysis.md` manually
+
+### Prevention Best Practices
+- Start with `--count 3` (default) before trying more roles
+- Use structured topic format: `"GOAL: [x] SCOPE: [y] CONTEXT: [z]"`
+- Review each role's output size before proceeding to synthesis
+- For complex topics, consider running brainstorm in phases (3 roles at a time)
 
 ## Reference Information
 
diff --git a/DASHBOARD_GUIDE.md b/DASHBOARD_GUIDE.md
index 6d9246f4..524bfd1e 100644
--- a/DASHBOARD_GUIDE.md
+++ b/DASHBOARD_GUIDE.md
@@ -148,6 +148,36 @@ CCW Dashboard 是一个单页应用（SPA），界面由四个核心部分组成
 - **模型配置**: 配置每个工具的主要和次要模型
 - **安装/卸载**: 通过向导安装或卸载工具
 
+#### API Endpoint 配置（无需安装 CLI）
+
+如果您没有安装 Gemini/Qwen CLI，但有 API 访问权限（如反向代理服务），可以在 `~/.claude/cli-tools.json` 中配置 `api-endpoint` 类型的工具：
+
+```json
+{
+  "version": "3.2.0",
+  "tools": {
+    "gemini-api": {
+      "enabled": true,
+      "type": "api-endpoint",
+      "id": "your-api-id",
+      "primaryModel": "gemini-2.5-pro",
+      "secondaryModel": "gemini-2.5-flash",
+      "tags": ["analysis"]
+    }
+  }
+}
+```
+
+**配置说明**：
+- `type: "api-endpoint"`: 表示使用 API 调用而非 CLI
+- `id`: API 端点标识符，用于路由请求
+- API Endpoint 工具仅支持**分析模式**（只读），不支持文件写入操作
+
+**使用示例**：
+```bash
+ccw cli -p "分析代码结构" --tool gemini-api --mode analysis
+```
+
 #### CodexLens 管理
 - **索引路径**: 查看和修改索引存储位置
 - **索引操作**: