first commit

autonomous-ai-agents/hermes-agent/references/custom-openai-compatible-providers.md

@@ -0,0 +1,81 @@
+# Custom OpenAI-Compatible Providers in Hermes
+
+When a provider isn't built-in but offers an OpenAI-compatible `/v1` endpoint, add it manually to `config.yaml`.
+
+## Steps
+
+1. **Find the base URL and API key env var** — usually `/v1` at the provider's domain.
+
+2. **List available models:**
+   ```bash
+   curl -s "https://<provider>/v1/models" \
+     -H "Authorization: Bearer $API_KEY" \
+     --max-time 15 | python3 -m json.tool
+   ```
+   Note model `id`, `input_modalities`, `context_length`, `supported_features`.
+
+3. **Add provider to config.yaml via Python** (don't hand-edit YAML — indentation errors break everything):
+   ```python
+   import yaml, json, os
+
+   config_path = os.path.expanduser("~/.hermes/config.yaml")
+   with open(config_path) as f:
+       config = yaml.safe_load(f)
+
+   # Read API key from .env
+   api_key = ""
+   with open(os.path.expanduser("~/.hermes/.env")) as f:
+       for line in f:
+           if line.startswith("YOUR_KEY_PREFIX="):
+               api_key = line.strip().split("=", 1)[1]
+               break
+
+   config.setdefault("providers", {})["your-provider"] = {
+       "api_key": api_key,
+       "base_url": "https://provider.example.com/v1",
+       "available_models_json": json.dumps([
+           {"id": "model-id", "name": "Display Name"},
+       ]),
+       "model": "default-model-id",
+       "model_display_name": "Default Display Name"
+   }
+
+   with open(config_path, "w") as f:
+       yaml.dump(config, f, default_flow_style=False, allow_unicode=True, sort_keys=False)
+   ```
+
+4. **Verify with a test call:**
+   ```bash
+   curl -s "https://provider.example.com/v1/chat/completions" \
+     -H "Authorization: Bearer $API_KEY" \
+     -H "Content-Type: application/json" \
+     -d '{"model":"model-id","messages":[{"role":"user","content":"hi"}],"max_tokens":100}'
+   ```
+
+5. **Use the model:** `hermes -m your-provider/model-id` or via `hermes model` picker.
+
+## Provider Config Fields
+
+| Field | Required | Notes |
+|-------|----------|-------|
+| `api_key` | Yes | Actual key value, not env var reference |
+| `base_url` | Yes | Must end with `/v1` (or `/v1/`) |
+| `available_models_json` | Yes | JSON string of `[{id, name}]` array |
+| `model` | No | Default model ID |
+| `model_display_name` | No | Human-readable default model name |
+| `api_mode` | No | Only set if non-standard (e.g. `anthropic-messages` for MiniMax). Omit for OpenAI-compatible. |
+| `protocol` | No | Usually leave as `''` |
+
+## Pitfalls
+
+- **Don't hand-edit YAML** — use Python `yaml.safe_load` + `yaml.dump` to avoid indentation corruption.
+- **`api_key` must be the actual value**, not `$ENV_VAR` — Hermes doesn't resolve env vars inside provider config (only in `.env`).
+- **No `api_mode` needed for OpenAI-compatible** — only set this for providers with custom protocols (Anthropic Messages, etc.).
+- **`reasoning` field in responses** — some providers (SenseNova, DeepSeek) return a `reasoning` field in the message object. Hermes handles this natively for reasoning-capable models.
+- **Model discovery** — always call `/v1/models` first; don't guess model IDs from documentation (they change).
+
+## Known Custom Providers
+
+| Provider | Base URL | Key Env Var | Models |
+|----------|----------|-------------|--------|
+| SenseNova | `https://token.sensenova.cn/v1` | `SN_API_KEY` | sensenova-6.7-flash-lite, deepseek-v4-flash, sensenova-u1-fast |

autonomous-ai-agents/hermes-agent/references/dashboard-remote-access.md

@@ -0,0 +1,68 @@
+# Dashboard Remote Access
+
+## Problem
+Dashboard binds to 127.0.0.1:9119 by default. Accessing from a different machine (e.g., local laptop → cloud VPS) requires either SSH tunnel or insecure bind.
+
+## Recommended: SSH Port Forwarding
+```bash
+# On your local machine
+ssh -L 9119:127.0.0.1:9119 user@server-ip
+# Then open http://127.0.0.1:9119 in browser
+```
+**Pitfall (Windows):** `ssh: connect to host ... port 22: Connection timed out` — almost always a cloud security group issue. Check your cloud provider's security group / firewall rules to allow inbound TCP 22. SSH socket activation (`ssh.socket`) is enabled by default on Ubuntu; the service itself may show `inactive (dead)` — that's normal, socket activation triggers it on connection.
+
+## Password Protection (Reverse Proxy)
+
+Dashboard has **no built-in password auth**. Options:
+
+### Nginx + Basic Auth
+```bash
+sudo apt install nginx apache2-utils
+sudo htpasswd -c /etc/nginx/.htpasswd your-username
+```
+```nginx
+server {
+    listen 8080;
+    location / {
+        auth_basic "Hermes Dashboard";
+        auth_basic_user_file /etc/nginx/.htpasswd;
+        proxy_pass http://127.0.0.1:9119;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        # WebSocket support for Chat TUI
+        proxy_http_version 1.1;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection "upgrade";
+    }
+}
+```
+
+### Caddy (simpler config)
+```
+:8080 {
+    basicauth * {
+        username $hashed_password
+    }
+    reverse_proxy localhost:9119
+}
+```
+Generate hash: `caddy hash-password --plaintext 'your-password'`
+
+## Alternative: Insecure Bind (⚠️ exposes API keys)
+```bash
+hermes dashboard --insecure --port 9119
+# Access via http://server-ip:9119
+```
+Only use on trusted/private networks. The dashboard exposes `.env` contents including API keys. The `--insecure` flag exists because there's no built-in auth — the warning is intentional.
+
+## TUI Mode (Embedded Chat)
+```bash
+hermes dashboard --tui --no-open
+```
+Adds a Chat tab to the web UI — a browser-based `hermes --tui` via PTY/WebSocket. Useful when CLI access is inconvenient.
+
+## Common Issues
+- Multiple dashboard processes: `hermes dashboard --stop` kills all
+- Port conflict: change port with `--port 8080`
+- Gateway must be running for Kanban dispatch to work (`hermes gateway status`)
+- SSH connection timeout from Windows: check cloud security group allows inbound TCP 22

autonomous-ai-agents/hermes-agent/references/dashboard-reverse-proxy.md

@@ -0,0 +1,110 @@
+# Hermes Dashboard Reverse Proxy with Nginx
+
+## Quick Setup (Nginx + Basic Auth)
+
+### 1. Install Dependencies
+```bash
+sudo apt update && sudo apt install -y nginx apache2-utils
+```
+
+### 2. Create Password File
+```bash
+# Generate password (will prompt for password twice)
+sudo htpasswd -c /etc/nginx/.htpasswd <username>
+
+# Or non-interactive:
+echo -n '<username>:' | sudo tee /etc/nginx/.htpasswd
+openssl passwd -apr1 '<password>' | sudo tee -a /etc/nginx/.htpasswd
+```
+
+### 3. Nginx Config (`/etc/nginx/sites-available/hermes-dashboard`)
+```nginx
+server {
+    listen 80;
+    server_name <your-domain-or-ip>;  # e.g., 111.230.53.30 or hermes.example.com
+
+    location / {
+        auth_basic "Hermes Dashboard";
+        auth_basic_user_file /etc/nginx/.htpasswd;
+        
+        proxy_pass http://127.0.0.1:9119;
+        # IMPORTANT: Use "localhost" for Host header, NOT $host
+        # Dashboard validates Host header and rejects non-localhost values
+        # This causes "Invalid Host header" error if set to $host
+        proxy_set_header Host localhost;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        
+        # WebSocket support (required for Chat TUI)
+        proxy_http_version 1.1;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection "upgrade";
+        
+        proxy_connect_timeout 60s;
+        proxy_send_timeout 60s;
+        proxy_read_timeout 60s;
+    }
+}
+```
+
+### 4. Enable & Reload
+```bash
+sudo ln -sf /etc/nginx/sites-available/hermes-dashboard /etc/nginx/sites-enabled/
+sudo nginx -t && sudo systemctl reload nginx
+```
+
+### 5. Ensure Dashboard Running
+```bash
+hermes dashboard --no-open --port 9119
+```
+
+## Access
+- URL: `http://<domain-or-ip>`
+- Auth: Browser popup for username/password
+
+## Commands
+```bash
+sudo systemctl status nginx
+sudo systemctl restart nginx
+hermes dashboard --status
+hermes dashboard --stop
+```
+
+## Cleanup (Remove Reverse Proxy)
+```bash
+# Stop services
+hermes dashboard --stop
+sudo systemctl stop nginx
+sudo systemctl disable nginx
+
+# Remove config files
+sudo rm -f /etc/nginx/sites-available/hermes-dashboard
+sudo rm -f /etc/nginx/sites-enabled/hermes-dashboard
+sudo rm -f /etc/nginx/.htpasswd
+```
+
+## HTTPS (Optional)
+Use Certbot for Let's Encrypt:
+```bash
+sudo apt install certbot python3-certbot-nginx
+sudo certbot --nginx -d hermes.example.com
+```
+
+## Pitfalls
+
+### Invalid Host Header Error
+If you see `{"detail":"Invalid Host header. Dashboard requests must use the hostname the server was bound to."}`:
+- **Cause**: Nginx is passing `$host` (the public domain/IP) but Dashboard only accepts `localhost`
+- **Fix**: Change `proxy_set_header Host $host;` to `proxy_set_header Host localhost;`
+
+### Domain Requires ICP Filing (China)
+If accessing via domain in China triggers ICP filing requirement:
+- **Solution**: Use IP address directly instead of domain
+- Update `server_name` to the server's public IP
+
+### Security Notes
+- Dashboard has NO built-in password auth
+- Without reverse proxy, anyone with network access can see API keys
+- Always use reverse proxy + basic auth for remote access
+- Consider SSH port forwarding as a more secure alternative

autonomous-ai-agents/hermes-agent/references/hermes-update-autostash-triage.md

@@ -0,0 +1,76 @@
+# Hermes Update Autostash Triage
+
+`hermes update` auto-stashes local changes before pulling. These accumulate as `stash@{N}` with the naming pattern:
+
+```
+hermes-update-autostash-YYYYMMDD-HHMMSS
+```
+
+## Triage Workflow
+
+### Step 1: List all stashes
+
+```bash
+git stash list
+```
+
+### Step 2: Quick scan each stash — file types matter
+
+```bash
+git stash show stash@{N} --stat
+```
+
+**Lock-file-only stashes** (only `package.json`, `package-lock.json`, `ui-tui/package-lock.json`):
+- Usually npm dependency resolution artifacts (registry mirror switches, peer dependency reclassification)
+- Safe to drop: `git stash drop stash@{N}`
+
+**Source-code stashes** (`.py`, `.ts`, `.tsx` files changed):
+- Need detailed analysis — these may contain valuable local features
+
+### Step 3: For source-code stashes — compare against current code
+
+Don't just `git stash pop`. First check if the features were already merged upstream:
+
+```bash
+# Get the full diff
+git stash show -p stash@{N}
+
+# For each key feature, search current code:
+grep -n "feature_keyword" path/to/file.py
+```
+
+**Classification:**
+- ✅ Already in current code → safe to drop
+- ❌ Missing from current code → candidate for restoration
+
+### Step 4: Decision matrix
+
+| Stash type | Action |
+|------------|--------|
+| Lock files only | Drop immediately |
+| Source code, all features merged | Drop |
+| Source code, some features missing | Selective restore (cherry-pick specific hunks) or apply + resolve conflicts |
+| Source code, all features missing | `git stash apply stash@{N}` then test |
+
+### Pitfalls
+
+- **Don't blindly pop stashes on an active branch** — always `apply` first (preserves stash), test, then `drop` if good.
+- **Registry mirror changes in lock files** (npmmirror.com, mirrors.tencentyun.com) are local environment artifacts, not valuable code. Drop them.
+- **`peer: true` removal** in lock files = npm re-resolved peer deps as direct deps. Not meaningful.
+- **5+ day old stashes** with source changes are likely abandoned experiments. Check if the user still needs them before restoring.
+- **Merge conflicts** are common after 5+ days — upstream moves fast. Expect to resolve manually.
+
+### Restoring selectively
+
+If only some hunks from a stash are needed:
+
+```bash
+# Apply but don't drop
+git stash apply stash@{N}
+
+# Review conflicts
+git diff
+
+# Or use interactive checkout for specific files
+git checkout stash@{N} -- path/to/specific/file.py
+```

autonomous-ai-agents/hermes-agent/references/qqbot-media-api.md

@@ -0,0 +1,61 @@
+# QQ Bot Rich Media API Reference
+
+Source: https://bot.q.qq.com/wiki/develop/api-v2/server-inter/message/send-receive/rich-media.html
+
+## Endpoints
+
+| Scope | Endpoint | Method |
+|-------|----------|--------|
+| 单聊 | `/v2/users/{openid}/files` | POST |
+| 群聊 | `/v2/groups/{group_openid}/files` | POST |
+
+## Parameters
+
+| Field | Type | Required | Description |
+|-------|------|----------|-------------|
+| file_type | int | Yes | 1=图片, 2=视频, 3=语音, 4=文件 |
+| url | string | Yes* | 媒体资源 URL（*url 或 file_data 二选一） |
+| file_data | string | No | base64 二进制数据 |
+| srv_send_msg | bool | Yes | true=直接发送（占用主动消息频次），false=仅上传获取 file_info |
+| file_name | string | No | 文件名（file_type=4 时建议传） |
+
+## Supported Formats
+
+- 图片: png/jpg
+- 视频: mp4
+- 语音: silk/wav/mp3/flac
+- 文件: 无限制（群场景暂不开放 file_type=4）
+
+## Response
+
+```json
+{
+  "file_uuid": "...",
+  "file_info": "...",   // 用于发送消息接口的 media 字段
+  "ttl": 3600           // 剩余秒数，0=长期有效
+}
+```
+
+## Sending with file_info
+
+```json
+{
+  "msg_type": 7,
+  "media": {"file_info": "<file_info_from_upload>"},
+  "msg_seq": 1
+}
+```
+
+## Notes
+
+- `file_info` 不受目标端影响，可复用到多个群/用户
+- 用 `/v2/users/{openid}/files` 上传的仅能发单聊，群上传的仅能发群聊
+- 建议 `srv_send_msg=false`，先获取 file_info 再发送
+- 文件大小限制：~100MB（分块上传），~10MB（inline base64）
+
+## Hermes Adapter Implementation
+
+- Gateway adapter: `gateway/platforms/qqbot/adapter.py` → `_send_media()` (line ~2690)
+- Chunked upload: `gateway/platforms/qqbot/chunked_upload.py`
+- Media types defined in `gateway/platforms/qqbot/constants.py` (MEDIA_TYPE_IMAGE=1, VIDEO=2, VOICE=3, FILE=4)
+- send_message tool gap: `_send_qqbot()` in `tools/send_message_tool.py:1677` is text-only

autonomous-ai-agents/hermes-agent/references/search-tools.md

@@ -0,0 +1,96 @@
+# Search Tools for Hermes Agent
+
+## Current Environment
+
+| Tool | Type | Usage |
+|------|------|-------|
+| **web** (built-in) | Web Search & Scraping | `web_search`, `web_extract` tools |
+| **session_search** | Session history | `session_search(query="...")` |
+| **mmx search** | MiniMax CLI | `mmx search query "关键词"` |
+
+## Popular Search APIs for AI Agents
+
+| Tool | Best For | Pricing | MCP Server |
+|------|----------|---------|------------|
+| **Tavily Search** | AI-native search, structured results | Free tier available | `tavily-mcp` |
+| **SerpAPI** | Google results scraping | Paid (100 free/month) | `serpapi-mcp` |
+| **Brave Search** | Privacy-focused, no tracking | Free tier (2000/month) | `brave-search-mcp` |
+| **Perplexity API** | AI search with citations | Paid | API only |
+| **Firecrawl** | Web scraping + extraction | Free tier | `firecrawl-mcp` |
+| **Jina Reader** | URL to Markdown conversion | Free tier | API only |
+| **SearXNG** | Self-hosted meta search | Free (self-hosted) | `searxng-mcp` |
+
+## Configuration Examples
+
+### Tavily Search (Recommended for AI Agents)
+
+```yaml
+# ~/.hermes/config.yaml
+mcp_servers:
+  tavily:
+    command: npx
+    args: ["-y", "tavily-mcp@latest"]
+    env:
+      TAVILY_API_KEY: "tvly-xxxxx"
+```
+
+Get API key: https://tavily.com (free tier: 1000 searches/month)
+
+### Brave Search
+
+```yaml
+mcp_servers:
+  brave-search:
+    command: npx
+    args: ["-y", "@anthropic/brave-search-mcp@latest"]
+    env:
+      BRAVE_API_KEY: "BSAxxxxx"
+```
+
+Get API key: https://brave.com/search/api/ (free tier: 2000 queries/month)
+
+### Firecrawl (for Web Scraping)
+
+```yaml
+mcp_servers:
+  firecrawl:
+    command: npx
+    args: ["-y", "firecrawl-mcp"]
+    env:
+      FIRECRAWL_API_KEY: "fc-xxxxx"
+```
+
+## When to Use Which
+
+| Scenario | Recommended Tool |
+|----------|------------------|
+| Quick factual lookup | `mmx search` or built-in `web_search` |
+| Structured data extraction | Tavily Search |
+| Google-specific results | SerpAPI |
+| Privacy-sensitive search | Brave Search |
+| Deep web scraping | Firecrawl |
+| Convert page to markdown | Jina Reader |
+| Previous conversation context | `session_search` |
+
+## Pitfalls
+
+- **mcp_servers: {}** is currently empty in this environment
+- Tool changes require `/reset` (new session) to take effect
+- Free tiers have rate limits; monitor usage for production use
+- Some MCP servers require Node.js (`npx` command)
+- **mmx search Token Plan 限制**：`mmx search query` 需要 MiniMax Token Plan 支持 `coding-plan-search` 模型。报错 "your current token plan not support model" 表示 plan 不含搜索能力，需升级或换替代方案
+- **服务器出站网络受限**：某些云服务器/容器的出站网络被限制，`urllib` 报 `Network is unreachable` (errno 101)。此时 curl、Python requests、mimo_search.py 均无法访问外部。排查：`curl -s -o /dev/null -w "%{http_code}" https://www.baidu.com`
+- **Playwright/Chrome 沙盒问题**：服务器上 Chrome 需要 `--no-sandbox` 参数才能启动。Hermes 的 browser 工具可能因沙盒限制失败。排查：检查 Chrome stderr 中是否有 "No usable sandbox" 错误
+- **mimo_search.py 环境变量**：脚本依赖 `XIAOMI_API_KEY` 环境变量，但 Hermes config.yaml 中的 key 不会自动 export。需手动 `export XIAOMI_API_KEY=...` 或从 config 提取
+
+## 搜索降级策略（按优先级）
+
+当用户请求搜索时，按以下顺序尝试：
+
+1. **mmx search** — 最简单，但需 Token Plan 支持
+2. **mimo_search.py** — MiMo 原生搜索，需 XIAOMI_API_KEY 环境变量 + 网络可达
+3. **web_search 工具** — Hermes 内置，需 web toolset 启用
+4. **Playwright 浏览器** — 通用兜底，但需 Chrome 可启动
+5. **curl + 搜索引擎** — 最基础，需服务器出站网络可达
+
+全部失败时，告知用户具体失败原因（Token Plan 限制 / 网络不通 / 浏览器沙盒问题），并建议用户在本地自行搜索。

autonomous-ai-agents/hermes-agent/references/skill-management-pitfalls.md

@@ -0,0 +1,64 @@
+# Skill Management Pitfalls
+
+Learned from attempting to optimize the skill library based on SkillRouter paper findings.
+
+## Pitfall 1: "Same Output" ≠ "Functionally Overlapping"
+
+**Wrong:** Deleted `pptx-generator` (python-pptx) because `powerpoint` (pptxgenjs) also makes .pptx files.
+**Right:** Different tech stacks = different fallback options. python-pptx is pure Python, pptxgenjs needs Node.js. Keep both.
+
+**Rule:** Two skills overlap only when they use the same tools AND serve the same user intent. Same output format is not enough.
+
+## Pitfall 2: Don't Cross-Reference in Descriptions
+
+**Wrong:** In arxiv's description: "需要多源学术搜索优先用 sn-search-academic"
+**Right:** Each skill describes itself only. No competitive recommendations.
+
+**Why:** Creates circular dependencies. If skill A recommends B, and B recommends A, the LLM loops.
+
+## Pitfall 3: Don't Expose Implementation Details
+
+**Wrong:** In sn-infographic description: "需要 SN_API_KEY"
+**Right:** "需要 SenseNova API"
+
+**Rule:** Descriptions should express user-facing capabilities, not internal tool/API names.
+
+## Pitfall 4: Check Hard Dependencies Before Deleting
+
+**Wrong:** Marked sn-research-planning for deletion because it was "never called."
+**Right:** sn-deep-research calls it via `skill_view("sn-research-planning")` at runtime. Deleting breaks the pipeline.
+
+**How to check:**
+```python
+# Search all SKILL.md files for references to the target skill name
+# Only HARD dependencies count: skill_view("target-name") or "读取 target-name"
+# "Related skills" mentions are SOFT and don't block deletion
+```
+
+## Pitfall 5: "Never skill_view'd" ≠ "Unused"
+
+Skills can be auto-loaded via the system prompt's "MUST load" instruction without explicit `skill_view()` calls. Session data only shows explicit tool calls.
+
+**Better metric:** Check if the skill is referenced as a runtime dependency by other skills.
+
+## Pitfall 6: Don't Batch Recommendations Without Verification
+
+**Wrong:** Generated all 87 recommendations at once, sent to user, then had to fix multiple errors.
+**Right:** Verify each category before sending. Check dependencies. Then send once.
+
+**User feedback:** "你这建议就不能确认好之后再发给我吗" (Can you verify before sending?)
+
+## Description Quality Formula
+
+Good skill description = **What it does** + **Trigger words** + **Negative boundary**
+
+Example:
+```
+"深度调研全流程编排器（入口 skill）。自动完成：规划→分维度取证→综合→成稿。
+触发词：深度研究/调研/全面研究/调研报告/deep research。
+不用于：单点事实问答、一句话摘要。"
+```
+
+- What: 深度调研全流程编排器
+- Triggers: 深度研究/调研/全面研究
+- Boundary: 不用于单点事实问答

autonomous-ai-agents/hermes-agent/references/web-search-from-server.md

@@ -0,0 +1,51 @@
+# 服务器环境网络搜索指南
+
+## Playwright 浏览器启动（服务器/容器环境）
+
+服务器环境（Ubuntu 23.10+、容器、VM）需要 `--no-sandbox` 参数：
+
+```python
+from playwright.sync_api import sync_playwright
+
+with sync_playwright() as p:
+    browser = p.chromium.launch(
+        headless=True, 
+        args=['--no-sandbox', '--disable-setuid-sandbox']
+    )
+    page = browser.new_page()
+    page.goto('https://example.com', timeout=30000)
+    # ... 操作页面
+    browser.close()
+```
+
+注意：Hermes 内置的 `browser_navigate` 工具不支持传递启动参数，必须直接用 Playwright API。
+
+## 搜索引擎反爬情况（2026年测试）
+
+| 搜索引擎 | 状态 | 备注 |
+|----------|------|------|
+| 百度 | ❌ 验证码 | 滑块验证，无法绕过 |
+| 搜狗 | ❌ 验证码 | 图片点选验证 |
+| 必应(cn.bing.com) | ⚠️ 可用但质量差 | 中文搜索结果常偏离关键词 |
+| Google | ❌ 超时 | 服务器网络不可达 |
+| DuckDuckGo | ❌ 超时 | 服务器网络不可达 |
+
+**结论**：服务器环境下，主流搜索引擎基本不可用。必应是唯一能返回结果的，但质量不稳定。
+
+## mmx search 限制
+
+`mmx search` 需要 Token Plan 支持 `coding-plan-search` 模型。如果报错：
+```
+your current token plan not support model, coding-plan-search
+```
+说明当前计划不支持搜索功能，需要升级或使用其他方式。
+
+## MiMo 模型搜索能力
+
+MiMo (mimo-v2.5-pro) 可以声明 `web_search` 工具调用，但**实际上不会真正联网搜索**。它只能基于训练数据回答，无法获取实时信息。
+
+## 替代方案优先级
+
+1. Playwright + 必应（唯一可行的浏览器方案）
+2. 直接访问目标网站（如培训机构官网）
+3. 用户自行搜索后提供信息