Agents have OpenSpace MCP tools available but never call them — all skills show 0 score

[shao-zhijie]

Bug/Question: Agents have OpenSpace MCP tools available but never call them — all skills show 0 score

Context

We're running OpenClaw v2026.4.24 with OpenSpace v1.26.0 as an MCP server (streamable-http on port 8083). The connection works fine — execute_task, search_skills, upload_skill, fix_skill tools are all injected into agent sessions.

The problem: Agents never actually call these tools. They bypass OpenSpace entirely and complete tasks using their local tools. The dashboard shows all skills at 0 score because nobody is using them.

Setup

// openclaw.json
{
  "mcp": {
    "servers": {
      "openspace": {
        "transport": "streamable-http",
        "url": "http://127.0.0.1:8083/mcp"
      }
    }
  }
}

• OpenClaw: v2026.4.24 (macOS arm64)
• OpenSpace: v1.26.0 (streamable-http, port 8083)
• MCP connection: ✅ working (tools injected into sessions)
• Tools available: openspace__execute_task, openspace__search_skills, openspace__upload_skill, openspace__fix_skill

What We've Tried

1. Added explicit rules in AGENTS.md:

   0. 🚨 强制使用 OpenSpace（最高优先级）
      - 任何任务，第一反应必须是：通过 OpenSpace 执行
      - execute_task 是默认动作，不是备选
      - 唯一例外：纯推理/纯回答/问候/闲聊
   

2. Verified tools are available — agents report seeing the tools in their system prompt

3. Tested manually — when I (the main agent) explicitly call openspace__execute_task, it works fine. The issue is that agents skip this step and go straight to local execution.

Questions

1. Is there a recommended way to make agents "default to" OpenSpace? Our agents see the tools but treat them as optional, not as the primary execution path.

2. Should execute_task be called automatically for certain task types? Or must every agent be explicitly prompted to use it?

3. Is there a "routing layer" or "auto-dispatch" feature that can intercept tasks and route them to OpenSpace without the agent deciding?

4. For multi-agent setups (main + consulting + fund-god + vibe-coding), does each agent need its own OpenSpace config, or do they share the same MCP server instance?

Related Issues

• Why it is smoothly to execute the task in the openspace lonely, however, when it comes to the openclaw with openspace, the mcp server always exceeds the limited time #73 — MCP timeout when using OpenClaw with OpenSpace
• [Question] Integration: OpenClaw (host) + OpenSpace (worker) single-shot task architecture #68 — Integration architecture question (single-shot tasks)
• The dashboard is showing stale data，it doesn't reflect the recent changes I made to the local skills. #70 — Dashboard showing stale data

Expected Behavior

When an agent receives a task that requires tools/search/execution, it should automatically call openspace__execute_task as the first action, not fall back to it after trying local tools.

Actual Behavior

Agents use local tools (exec, read, write, web_search, etc.) to complete tasks directly. OpenSpace tools are available but never invoked. Dashboard shows 0 score for all skills.

[shao-zhijie]

总结：0 分是 LLM 智能判断的正常行为 ✅

我们在生产环境（OpenClaw 5.x + OpenSpace）上长期运行了 OpenSpace，发现 skills 一直 0 分不是 bug，是 LLM 的预期行为：

为什么会 0 分？

• LLM 收到任务后，会先判断'这个任务是否需要 skill'
• 当 LLM 认为纯推理/纯回答就够用时，不会调用 skill
• 只有当 LLM 看到'需要外部工具、需要多步骤、需要专业知识库'时，才会触发 skill selection
• 所以大多数闲聊/greeting 类任务，skills 打分是 0

我们观察到的模式：

• 简单对话：0 个 skill 被调用 → 0 分正常
• 复杂分析：5-10 个 skill 被评分 → 部分被选中
• 持续学习：随着 skill 库变丰富，调用率会上升

建议：

• 不要把 '0 分' 当作 bug
• 可以加一个 skill_invocation_rate 指标来评估 LLM 实际使用 skill 的频率
• 期待 OpenSpace 未来能加 LLM 主动调用 skill 的 prompt hint

我们这边已经用 execute_task + 自定义 skill selection 逻辑解决了，issue 可以作为 design discussion 保留，标记为 not-a-bug。