Northern_Lights/cherry-studio
1
0
Fork 0
mirror of https://github.com/CherryHQ/cherry-studio.git synced 2026-01-12 00:49:14 +08:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

♻️ refactor: move hub prompts to dedicated prompts-code-mode.ts

Browse Source 
- Create src/renderer/src/config/prompts-code-mode.ts
- Move HUB_MODE_SYSTEM_PROMPT_BASE and AUTO_MODE_SYSTEM_PROMPT_BASE
- Move getHubModeSystemPrompt() and getAutoModeSystemPrompt()
- Extract shared buildToolsSection() helper
- Update parameterBuilder.ts import
This commit is contained in:
Vaayne 2025-12-31 11:51:25 +08:00
parent b58569fcec
commit 0482167476
3 changed files with 267 additions and 257 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/renderer/src/aiCore/prepareParams/parameterBuilder.ts
View File

@ -26,7 +26,7 @@ import {
isSupportedThinkingTokenModel,
isWebSearchModel
} from '@renderer/config/models'
import { getAutoModeSystemPrompt } from '@renderer/config/prompts'
import { getAutoModeSystemPrompt } from '@renderer/config/prompts-code-mode'
import { fetchAllActiveServerTools } from '@renderer/services/ApiService'
import { getDefaultModel } from '@renderer/services/AssistantService'
import store from '@renderer/store'

266
src/renderer/src/config/prompts-code-mode.ts Normal file
View File

@ -0,0 +1,266 @@
import { generateMcpToolFunctionName } from '@shared/mcp'
export interface ToolInfo {
name: string
serverName?: string
description?: string
}
/**
* Hub Mode System Prompt - For native MCP tool calling
* Used when model supports native function calling via MCP protocol
*/
const HUB_MODE_SYSTEM_PROMPT_BASE = `
## Hub MCP Tools – Code Execution Mode
You can discover and call MCP tools through the hub server using two meta-tools: **search** and **exec**.
### Critical Rules (Read First)
1. You MUST explicitly \`return\` the final value from your \`exec\` code. If you do not return a value, the result will be \`undefined\`.
2. All MCP tools are async functions. Always call them as \`await ToolName(params)\`.
3. Use the exact function names and parameter shapes returned by \`search\`.
4. You CANNOT call \`search\` or \`exec\` from inside \`exec\` code—use them only as MCP tools.
5. \`console.log\` output is NOT the result. Logs are separate; the final answer must come from \`return\`.
### Workflow
1. Call \`search\` with relevant keywords to discover tools.
2. Read the returned JavaScript function declarations and JSDoc to understand names and parameters.
3. Call \`exec\` with JavaScript code that uses the discovered tools and ends with an explicit \`return\`.
4. Use the \`exec\` result as your answer.
### What \`search\` Does
- Input: keyword string (comma-separated for OR-matching), plus optional \`limit\`.
- Output: JavaScript async function declarations with JSDoc showing exact function names, parameters, and return types.
### What \`exec\` Does
- Runs JavaScript code in an isolated async context (wrapped as \`(async () => { your code })())\`.
- All discovered tools are exposed as async functions: \`await ToolName(params)\`.
- Available helpers:
- \`parallel(...promises)\` → \`Promise.all(promises)\`
- \`settle(...promises)\` → \`Promise.allSettled(promises)\`
- \`console.log/info/warn/error/debug\`
- Returns JSON with: \`result\` (your returned value), \`logs\` (optional), \`error\` (optional), \`isError\` (optional).
### Example: Single Tool Call
\`\`\`javascript
// Step 1: search({ query: "browser,fetch" })
// Step 2: exec with:
const page = await CherryBrowser_fetch({ url: "https://example.com" })
return page
\`\`\`
### Example: Multiple Tools with Parallel
\`\`\`javascript
const [forecast, time] = await parallel(
Weather_getForecast({ city: "Paris" }),
Time_getLocalTime({ city: "Paris" })
)
return { city: "Paris", forecast, time }
\`\`\`
### Example: Handle Partial Failures with Settle
\`\`\`javascript
const results = await settle(
Weather_getForecast({ city: "Paris" }),
Weather_getForecast({ city: "Tokyo" })
)
const successful = results.filter(r => r.status === "fulfilled").map(r => r.value)
return { results, successful }
\`\`\`
### Example: Error Handling
\`\`\`javascript
try {
const user = await User_lookup({ email: "user@example.com" })
return { found: true, user }
} catch (error) {
return { found: false, error: String(error) }
}
\`\`\`
### Common Mistakes to Avoid
❌ **Forgetting to return** (result will be \`undefined\`):
\`\`\`javascript
const data = await SomeTool({ id: "123" })
// Missing return!
\`\`\`
✅ **Always return**:
\`\`\`javascript
const data = await SomeTool({ id: "123" })
return data
\`\`\`
❌ **Only logging, not returning**:
\`\`\`javascript
const data = await SomeTool({ id: "123" })
console.log(data) // Logs are NOT the result!
\`\`\`
❌ **Missing await**:
\`\`\`javascript
const data = SomeTool({ id: "123" }) // Returns Promise, not value!
return data
\`\`\`
❌ **Awaiting before parallel**:
\`\`\`javascript
await parallel(await ToolA(), await ToolB()) // Wrong: runs sequentially
\`\`\`
✅ **Pass promises directly to parallel**:
\`\`\`javascript
await parallel(ToolA(), ToolB()) // Correct: runs in parallel
\`\`\`
### Best Practices
- Always call \`search\` first to discover tools and confirm signatures.
- Always use an explicit \`return\` at the end of \`exec\` code.
- Use \`parallel\` for independent operations that can run at the same time.
- Use \`settle\` when some calls may fail but you still want partial results.
- Prefer a single \`exec\` call for multi-step flows.
- Treat \`console.*\` as debugging only, never as the primary result.
`
/**
* Auto Mode System Prompt - For XML tool_use format
* Used when model needs explicit XML format to invoke tools
* Only teaches search and exec tools
*/
const AUTO_MODE_SYSTEM_PROMPT_BASE = `
You can discover and invoke MCP tools through a hub using TWO meta-tools: \`search\` and \`exec\`.
## Tool Invocation Format
When you want to call a tool, output exactly one XML block:
<tool_use>
<name>{tool_name}</name>
<arguments>{json_arguments}</arguments>
</tool_use>
Rules:
- \`{tool_name}\` MUST be either \`search\` or \`exec\`
- \`<arguments>\` MUST contain valid JSON (no comments, no trailing commas)
- Do NOT include extra text before or after the \`<tool_use>\` block
## Available Tools
1. **search** - Discover MCP tools by keyword
\`\`\`json
{ "query": "keyword1,keyword2", "limit": 10 }
\`\`\`
Returns JavaScript function declarations with JSDoc showing names, parameters, and return types.
2. **exec** - Execute JavaScript that calls discovered tools
\`\`\`json
{ "code": "const r = await ToolName({...}); return r;" }
\`\`\`
**CRITICAL:** You MUST \`return\` the final value, or result will be \`undefined\`.
## Workflow
1. Call \`search\` with keywords to discover tools
2. Read the returned function signatures carefully
3. Call \`exec\` with JavaScript code that:
- Uses ONLY functions returned by \`search\`
- Calls them with \`await\`
- Ends with explicit \`return\`
4. Answer the user based on the result
## Example
User: "Calculate 15 * 7"
Assistant calls search:
<tool_use>
<name>search</name>
<arguments>{"query": "python,calculator"}</arguments>
</tool_use>
Hub returns function signature:
\`\`\`js
async function CherryPython_pythonExecute(params: { code: string }): Promise<unknown>
\`\`\`
Assistant calls exec:
<tool_use>
<name>exec</name>
<arguments>{"code": "const result = await CherryPython_pythonExecute({ code: '15 * 7' }); return result;"}</arguments>
</tool_use>
Hub returns: { "result": 105 }
Assistant answers: "15 × 7 = 105"
## Common Mistakes
❌ Forgetting to return (result will be undefined):
\`\`\`js
await SomeTool({ id: "123" })
\`\`\`
✅ Always return:
\`\`\`js
const data = await SomeTool({ id: "123" }); return data;
\`\`\`
❌ Calling exec before search - you must discover tools first
❌ Using functions not returned by search
`
function buildToolsSection(tools: ToolInfo[]): string {
const existingNames = new Set<string>()
return tools
.map((t) => {
const functionName = generateMcpToolFunctionName(t.serverName, t.name, existingNames)
const desc = t.description || ''
const normalizedDesc = desc.replace(/\s+/g, ' ').trim()
const truncatedDesc = normalizedDesc.length > 50 ? `${normalizedDesc.slice(0, 50)}...` : normalizedDesc
return `- ${functionName}: ${truncatedDesc}`
})
.join('\n')
}
/**
* Get system prompt for Hub Mode (native MCP tool calling)
*/
export function getHubModeSystemPrompt(tools: ToolInfo[] = []): string {
if (tools.length === 0) {
return ''
}
const toolsSection = buildToolsSection(tools)
return `${HUB_MODE_SYSTEM_PROMPT_BASE}
### Available Tools
${toolsSection}
`
}
/**
* Get system prompt for Auto Mode (XML tool_use format)
*/
export function getAutoModeSystemPrompt(tools: ToolInfo[] = []): string {
if (tools.length === 0) {
return ''
}
const toolsSection = buildToolsSection(tools)
return `${AUTO_MODE_SYSTEM_PROMPT_BASE}
## Discoverable Tools (use search to get full signatures)
${toolsSection}
`
}

256
src/renderer/src/config/prompts.ts
View File

@ -1,4 +1,3 @@
import { generateMcpToolFunctionName } from '@shared/mcp'
import dayjs from 'dayjs'
export const AGENT_PROMPT = `
@ -469,258 +468,3 @@ Example: [nytimes.com](https://nytimes.com/some-page).
If have multiple citations, please directly list them like this:
[www.nytimes.com](https://nytimes.com/some-page)[www.bbc.com](https://bbc.com/some-page)
`
const HUB_MODE_SYSTEM_PROMPT_BASE = `
## Hub MCP Tools – Code Execution Mode
You can discover and call MCP tools through the hub server using two meta-tools: **search** and **exec**.
### Critical Rules (Read First)
1. You MUST explicitly \`return\` the final value from your \`exec\` code. If you do not return a value, the result will be \`undefined\`.
2. All MCP tools are async functions. Always call them as \`await ToolName(params)\`.
3. Use the exact function names and parameter shapes returned by \`search\`.
4. You CANNOT call \`search\` or \`exec\` from inside \`exec\` code—use them only as MCP tools.
5. \`console.log\` output is NOT the result. Logs are separate; the final answer must come from \`return\`.
### Workflow
1. Call \`search\` with relevant keywords to discover tools.
2. Read the returned JavaScript function declarations and JSDoc to understand names and parameters.
3. Call \`exec\` with JavaScript code that uses the discovered tools and ends with an explicit \`return\`.
4. Use the \`exec\` result as your answer.
### What \`search\` Does
- Input: keyword string (comma-separated for OR-matching), plus optional \`limit\`.
- Output: JavaScript async function declarations with JSDoc showing exact function names, parameters, and return types.
### What \`exec\` Does
- Runs JavaScript code in an isolated async context (wrapped as \`(async () => { your code })())\`.
- All discovered tools are exposed as async functions: \`await ToolName(params)\`.
- Available helpers:
- \`parallel(...promises)\` → \`Promise.all(promises)\`
- \`settle(...promises)\` → \`Promise.allSettled(promises)\`
- \`console.log/info/warn/error/debug\`
- Returns JSON with: \`result\` (your returned value), \`logs\` (optional), \`error\` (optional), \`isError\` (optional).
### Example: Single Tool Call
\`\`\`javascript
// Step 1: search({ query: "browser,fetch" })
// Step 2: exec with:
const page = await CherryBrowser_fetch({ url: "https://example.com" })
return page
\`\`\`
### Example: Multiple Tools with Parallel
\`\`\`javascript
const [forecast, time] = await parallel(
Weather_getForecast({ city: "Paris" }),
Time_getLocalTime({ city: "Paris" })
)
return { city: "Paris", forecast, time }
\`\`\`
### Example: Handle Partial Failures with Settle
\`\`\`javascript
const results = await settle(
Weather_getForecast({ city: "Paris" }),
Weather_getForecast({ city: "Tokyo" })
)
const successful = results.filter(r => r.status === "fulfilled").map(r => r.value)
return { results, successful }
\`\`\`
### Example: Error Handling
\`\`\`javascript
try {
const user = await User_lookup({ email: "user@example.com" })
return { found: true, user }
} catch (error) {
return { found: false, error: String(error) }
}
\`\`\`
### Common Mistakes to Avoid
❌ **Forgetting to return** (result will be \`undefined\`):
\`\`\`javascript
const data = await SomeTool({ id: "123" })
// Missing return!
\`\`\`
✅ **Always return**:
\`\`\`javascript
const data = await SomeTool({ id: "123" })
return data
\`\`\`
❌ **Only logging, not returning**:
\`\`\`javascript
const data = await SomeTool({ id: "123" })
console.log(data) // Logs are NOT the result!
\`\`\`
❌ **Missing await**:
\`\`\`javascript
const data = SomeTool({ id: "123" }) // Returns Promise, not value!
return data
\`\`\`
❌ **Awaiting before parallel**:
\`\`\`javascript
await parallel(await ToolA(), await ToolB()) // Wrong: runs sequentially
\`\`\`
✅ **Pass promises directly to parallel**:
\`\`\`javascript
await parallel(ToolA(), ToolB()) // Correct: runs in parallel
\`\`\`
### Best Practices
- Always call \`search\` first to discover tools and confirm signatures.
- Always use an explicit \`return\` at the end of \`exec\` code.
- Use \`parallel\` for independent operations that can run at the same time.
- Use \`settle\` when some calls may fail but you still want partial results.
- Prefer a single \`exec\` call for multi-step flows.
- Treat \`console.*\` as debugging only, never as the primary result.
`
interface ToolInfo {
name: string
serverName?: string
description?: string
}
export function getHubModeSystemPrompt(tools: ToolInfo[] = []): string {
if (tools.length === 0) {
return ''
}
const existingNames = new Set<string>()
const toolsSection = tools
.map((t) => {
const functionName = generateMcpToolFunctionName(t.serverName, t.name, existingNames)
const desc = t.description || ''
const normalizedDesc = desc.replace(/\s+/g, ' ').trim()
const truncatedDesc = normalizedDesc.length > 50 ? `${normalizedDesc.slice(0, 50)}...` : normalizedDesc
return `- ${functionName}: ${truncatedDesc}`
})
.join('\n')
return `${HUB_MODE_SYSTEM_PROMPT_BASE}
### Available Tools
${toolsSection}
`
}
const AUTO_MODE_SYSTEM_PROMPT_BASE = `
You can discover and invoke MCP tools through a hub using TWO meta-tools: \`search\` and \`exec\`.
## Tool Invocation Format
When you want to call a tool, output exactly one XML block:
<tool_use>
<name>{tool_name}</name>
<arguments>{json_arguments}</arguments>
</tool_use>
Rules:
- \`{tool_name}\` MUST be either \`search\` or \`exec\`
- \`<arguments>\` MUST contain valid JSON (no comments, no trailing commas)
- Do NOT include extra text before or after the \`<tool_use>\` block
## Available Tools
1. **search** - Discover MCP tools by keyword
\`\`\`json
{ "query": "keyword1,keyword2", "limit": 10 }
\`\`\`
Returns JavaScript function declarations with JSDoc showing names, parameters, and return types.
2. **exec** - Execute JavaScript that calls discovered tools
\`\`\`json
{ "code": "const r = await ToolName({...}); return r;" }
\`\`\`
**CRITICAL:** You MUST \`return\` the final value, or result will be \`undefined\`.
## Workflow
1. Call \`search\` with keywords to discover tools
2. Read the returned function signatures carefully
3. Call \`exec\` with JavaScript code that:
- Uses ONLY functions returned by \`search\`
- Calls them with \`await\`
- Ends with explicit \`return\`
4. Answer the user based on the result
## Example
User: "Calculate 15 * 7"
Assistant calls search:
<tool_use>
<name>search</name>
<arguments>{"query": "python,calculator"}</arguments>
</tool_use>
Hub returns function signature:
\`\`\`js
async function CherryPython_pythonExecute(params: { code: string }): Promise<unknown>
\`\`\`
Assistant calls exec:
<tool_use>
<name>exec</name>
<arguments>{"code": "const result = await CherryPython_pythonExecute({ code: '15 * 7' }); return result;"}</arguments>
</tool_use>
Hub returns: { "result": 105 }
Assistant answers: "15 × 7 = 105"
## Common Mistakes
❌ Forgetting to return (result will be undefined):
\`\`\`js
await SomeTool({ id: "123" })
\`\`\`
✅ Always return:
\`\`\`js
const data = await SomeTool({ id: "123" }); return data;
\`\`\`
❌ Calling exec before search - you must discover tools first
❌ Using functions not returned by search
`
export function getAutoModeSystemPrompt(tools: ToolInfo[] = []): string {
if (tools.length === 0) {
return ''
}
const existingNames = new Set<string>()
const toolsSection = tools
.map((t) => {
const functionName = generateMcpToolFunctionName(t.serverName, t.name, existingNames)
const desc = t.description || ''
const normalizedDesc = desc.replace(/\s+/g, ' ').trim()
const truncatedDesc = normalizedDesc.length > 50 ? `${normalizedDesc.slice(0, 50)}...` : normalizedDesc
return `- ${functionName}: ${truncatedDesc}`
})
.join('\n')
return `${AUTO_MODE_SYSTEM_PROMPT_BASE}
## Discoverable Tools (use search to get full signatures)
${toolsSection}
`
}