@zleap-ai/sag-cli 0.4.1 → 0.5.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zleap-ai/sag-cli",
-  "version": "0.4.1",
+  "version": "0.5.0",
   "description": "Command-line client and diagnostics for SAG knowledge bases",
   "type": "module",
   "bin": {
@@ -11,7 +11,8 @@
     "README.md",
     "README.en.md",
     "LICENSE",
-    "skill"
+    "skill",
+    "skills/sag-mcp"
   ],
   "scripts": {
     "build": "tsup",

package/skill/SKILL.md

@@ -8,6 +8,8 @@ metadata:
 
 # SAG Knowledge Base
 
+本 Skill 面向个人版/开源版 SAG 知识库；企业平台开放的 MCP 连接请使用随包发布的 `skills/sag-mcp` Skill。
+
 通过 `sag` CLI 探索和检索 SAG 知识库。
 
 ## 调用方式

package/skills/sag-mcp/SKILL.md

@@ -0,0 +1,17 @@
+---
+name: sag-mcp
+description: Use connected enterprise SAG MCP servers to search, read, and manage authorized knowledge.
+---
+
+# SAG MCP
+
+1. Identify the connected enterprise SAG MCP servers available in this task.
+2. When there is only one SAG MCP server, use it directly.
+3. When several SAG MCP servers are available, select the server that matches the user's stated organization or server context.
+4. If that context does not identify one server, ask the user which server to use before making a call.
+5. Read the current tool list before choosing a tool. Use only tools that the selected server lists.
+6. Read [the CLI reference](references/cli-reference.md) when you need tool discovery, call details, or input formatting.
+7. Read [the tool safety rules](references/tool-safety.md) before any action that can change knowledge or whose safety is unknown.
+8. Use returned evidence and identifiers accurately. State when the server has no result or does not authorize the requested action.
+
+Completion: answer from the selected server's returned information, or clearly state why no authorized result is available.

package/skills/sag-mcp/references/cli-reference.md

@@ -0,0 +1,9 @@
+# SAG MCP tool reference
+
+1. Open the selected server's tool list at the start of the task and again when the available tools may have changed.
+2. Select a listed tool whose description fits the user's requested outcome.
+3. Read that tool's input schema. Build one input object that contains only the required and relevant fields.
+4. Call the tool through the Agent's MCP interface with the selected server, tool name, and input object.
+5. Read the returned result before deciding whether another tool is needed. Preserve source references and identifiers supplied by the server.
+
+Do not invent tool names, input fields, source references, or results. Report missing permission, unavailable tools, and empty results plainly.

package/skills/sag-mcp/references/tool-safety.md

@@ -0,0 +1,10 @@
+# SAG MCP tool safety
+
+Use MCP tool annotations to decide whether a call needs confirmation.
+
+- Call a tool without confirmation only when `annotations.readOnlyHint === true`.
+- Ask for confirmation before a tool call when `annotations.destructiveHint === true`.
+- Ask for confirmation before a tool call when `annotations.readOnlyHint === false`.
+- Ask for confirmation before a tool call when annotations are missing or do not establish that it is read-only.
+
+Before asking, state the selected server, the tool, the intended effect, and the relevant input. Wait for the user's clear confirmation before making the call.