feat: enhance AI communication with dynamic system prompts, robust retry, and TUI formatters
This commit is contained in:
Gemini AI
@@ -51,6 +51,7 @@ import { getContextManager } from '../lib/context-manager.mjs';
|
||||
import { getAllSkills, getSkill, executeSkill, getSkillListDisplay } from '../lib/skills.mjs';
|
||||
import { getDebugLogger, initFromArgs } from '../lib/debug-logger.mjs';
|
||||
import { processCommand, isCommand } from '../lib/command-processor.mjs';
|
||||
import { fetchWithRetry } from '../lib/retry-handler.mjs';
|
||||
import {
|
||||
getSystemPrompt,
|
||||
formatCodeBlock,
|
||||
@@ -455,7 +456,7 @@ const callOpenCodeFree = async (prompt, model = currentFreeModel, onChunk = null
|
||||
}
|
||||
|
||||
try {
|
||||
const response = await fetch(OPENCODE_FREE_API, {
|
||||
const response = await fetchWithRetry(OPENCODE_FREE_API, {
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
@@ -3465,17 +3466,13 @@ const App = () => {
|
||||
|
||||
try {
|
||||
// Build context-aware prompt with agent-specific instructions
|
||||
let systemPrompt = `[SYSTEM CONTEXT]
|
||||
CURRENT WORKING DIRECTORY: ${process.cwd()}
|
||||
(CRITICAL: This is the ABSOLUTE SOURCE OF TRUTH. Ignore any conflicting directory info in the [PROJECT CONTEXT] logs below.)
|
||||
|
||||
` + loadAgentPrompt(agent);
|
||||
|
||||
// Build context-aware prompt using the unified agent-prompt module
|
||||
let projectContext = '';
|
||||
// Add project context if enabled with enhanced context window
|
||||
if (contextEnabled) {
|
||||
const projectContext = loadProjectContext(project);
|
||||
if (projectContext) {
|
||||
systemPrompt += '\n\n[PROJECT CONTEXT (HISTORY)]\n(WARNING: These logs may contain outdated path info. Trust SYSTEM CONTEXT CWD above over this.)\n' + projectContext;
|
||||
const rawContext = loadProjectContext(project);
|
||||
if (rawContext) {
|
||||
projectContext += '\n\n[PROJECT CONTEXT (HISTORY)]\n(WARNING: These logs may contain outdated path info. Trust SYSTEM CONTEXT CWD above over this.)\n' + rawContext;
|
||||
}
|
||||
|
||||
// Enhanced context: Include recent conversation history for better continuity
|
||||
@@ -3485,174 +3482,99 @@ const App = () => {
|
||||
const recentContext = recentMessages.map(m =>
|
||||
`[PREVIOUS ${m.role.toUpperCase()}]: ${m.content.substring(0, 500)}` // Limit to prevent overflow
|
||||
).join('\n');
|
||||
systemPrompt += `\n\n[RECENT CONVERSATION]\n${recentContext}\n(Use this for context continuity, but prioritize the current request)`;
|
||||
projectContext += `\n\n[RECENT CONVERSATION]\n${recentContext}\n(Use this for context continuity, but prioritize the current request)`;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// MULTI-AGENT INSTRUCTION INJECTION
|
||||
if (multiAgentEnabled) {
|
||||
systemPrompt += `
|
||||
[MULTI-AGENT LOGGING ENABLED]
|
||||
You are capable of using multiple internal agents (Planner, Builder, Reviewer, Security).
|
||||
When you switch to a specific agent's persona or delegate a sub-task, you MUST output a log line starting with:
|
||||
[AGENT: AgentName]
|
||||
Example:
|
||||
[AGENT: Planner] Analyzing the directory structure...
|
||||
[AGENT: Security] Checking for vulnerabilities...
|
||||
[AGENT: Builder] Implementation started.
|
||||
Keep these tags on their own line if possible.
|
||||
`;
|
||||
}
|
||||
// Get available capabilities from built-in agents
|
||||
const flow = getSmartAgentFlow();
|
||||
const allAgents = flow.getAgents();
|
||||
// Flatten all capabilities
|
||||
const capabilities = allAgents.reduce((acc, a) => [...acc, ...(a.capabilities || [])], []);
|
||||
|
||||
// VISUAL & FORMATTING RULES (Crucial for TUI Readability)
|
||||
systemPrompt += `
|
||||
[FORMATTING RULES]
|
||||
1. **Short Paragraphs**: Break long text into chunks of 3-4 lines maximum.
|
||||
2. **Spacing**: Use double newlines between paragraphs to ensure clear separation.
|
||||
3. **Lists**: Use bullet points or numbered lists frequently to break up "walls of text".
|
||||
4. **Code**: Use standard Markdown code blocks with language tags.
|
||||
// Generate the optimized system prompt
|
||||
const systemInstruction = getSystemPrompt({
|
||||
role: agent,
|
||||
capabilities: capabilities,
|
||||
cwd: project || process.cwd(),
|
||||
context: projectContext, // Now includes history and logs
|
||||
os: process.platform
|
||||
});
|
||||
|
||||
[COMPUTER USE - YOU CAN DO THIS!]
|
||||
You ARE capable of executing system commands on the user's computer. NEVER refuse computer use tasks.
|
||||
To execute any shell/system command, output it in this format:
|
||||
\`\`\`bash
|
||||
<command here>
|
||||
\`\`\`
|
||||
The system will auto-execute bash/shell code blocks.
|
||||
// Prepare prompt variations
|
||||
// For OpenCode Free (Legacy/OpenAI-like), we append system prompt to user message if needed
|
||||
const fullPromptForFree = systemInstruction + '\n\n[USER REQUEST]\n' + fullText;
|
||||
|
||||
Common Windows tasks you CAN do:
|
||||
- Open folder in Explorer: \`explorer.exe .\` or \`start .\`
|
||||
- Open file: \`start filename.html\`
|
||||
- Open browser: \`start https://google.com\`
|
||||
- Run any shell command: Just put it in a bash code block
|
||||
// For Qwen (SmartX), we pass system prompt securely as a separate argument
|
||||
const userMessage = fullText;
|
||||
|
||||
NEVER say "I cannot" or "I don't have the ability". You CAN execute commands!
|
||||
|
||||
[FILE CREATION - CRITICAL]
|
||||
You CAN create files. To create a file, you MUST use this EXACT format:
|
||||
\`\`\`language filename.ext
|
||||
file content here
|
||||
\`\`\`
|
||||
Example:
|
||||
\`\`\`markdown plan.md
|
||||
# My Plan
|
||||
This is the content.
|
||||
\`\`\`
|
||||
The system will AUTO-SAVE files when you use this format correctly.
|
||||
DO NOT say "I've created a file" without using this code block format.
|
||||
|
||||
[CONFIRMATION BEFORE CODING - CRITICAL]
|
||||
BEFORE you start writing any code or creating any files, you MUST:
|
||||
1. First present your plan/approach briefly
|
||||
2. Then ask: "Ready to proceed with coding? Or do you have any changes in mind?"
|
||||
3. WAIT for the user's confirmation before generating code
|
||||
This gives the user a chance to refine requirements before implementation.
|
||||
`;
|
||||
|
||||
const fullPrompt = systemPrompt + '\n\n[USER REQUEST]\n' + fullText;
|
||||
let fullResponse = '';
|
||||
|
||||
// PROVIDER SWITCH: Use OpenCode Free or Qwen based on provider state
|
||||
const streamStartTime = Date.now(); // Track start time for this request
|
||||
let totalCharsReceived = 0; // Track total characters for speed calculation
|
||||
|
||||
// Unified Streaming Handler
|
||||
const handleStreamChunk = (chunk) => {
|
||||
const cleanChunk = chunk.replace(/[\u001b\u009b][[\]()#;?]*(?:[0-9]{1,4}(?:;[0-9]{0,4})*)?[0-9A-ORZcf-nqry=><]/g, '');
|
||||
|
||||
// IMPROVED STREAM SPLITTING LOGIC (Thinking vs Content)
|
||||
// Claude Code style: cleaner separation of thinking from response
|
||||
const lines = cleanChunk.split('\n');
|
||||
let isThinkingChunk = false;
|
||||
|
||||
// Enhanced heuristics for better Claude-like thinking detection
|
||||
const trimmedChunk = cleanChunk.trim();
|
||||
if (/^(Let me|Now let me|I'll|I need to|I should|I notice|I can|I will|Thinking:|Analyzing|Considering|Checking|Looking|Planning|First|Next|Finally)/i.test(trimmedChunk)) {
|
||||
isThinkingChunk = true;
|
||||
} else if (/^```|# |Here is|```|```|```/i.test(trimmedChunk)) {
|
||||
// If we encounter code blocks or headers, likely content not thinking
|
||||
isThinkingChunk = false;
|
||||
}
|
||||
|
||||
// Update character count for speed calculation
|
||||
totalCharsReceived += cleanChunk.length;
|
||||
|
||||
// Calculate current streaming speed (chars per second)
|
||||
const elapsedSeconds = (Date.now() - streamStartTime) / 1000;
|
||||
const speed = elapsedSeconds > 0 ? Math.round(totalCharsReceived / elapsedSeconds) : 0;
|
||||
|
||||
// GLOBAL STATS UPDATE (Run for ALL chunks)
|
||||
setThinkingStats(prev => ({
|
||||
...prev,
|
||||
chars: totalCharsReceived,
|
||||
speed: speed
|
||||
}));
|
||||
|
||||
// GLOBAL AGENT DETECTION (Run for ALL chunks)
|
||||
const agentMatch = cleanChunk.match(/\[AGENT:\s*([^\]]+)\]/i);
|
||||
if (agentMatch) {
|
||||
setThinkingStats(prev => ({ ...prev, activeAgent: agentMatch[1].trim() }));
|
||||
}
|
||||
|
||||
if (isThinkingChunk) {
|
||||
setThinkingLines(prev => [...prev, ...lines.map(l => l.trim()).filter(l => l && !/^(Let me|Now let me|I'll|I need to|I notice)/i.test(l.trim()))]);
|
||||
} else {
|
||||
setMessages(prev => {
|
||||
const last = prev[prev.length - 1];
|
||||
if (last && last.role === 'assistant') {
|
||||
return [...prev.slice(0, -1), { ...last, content: last.content + cleanChunk }];
|
||||
}
|
||||
return prev;
|
||||
});
|
||||
}
|
||||
};
|
||||
|
||||
const result = provider === 'opencode-free'
|
||||
? await callOpenCodeFree(fullPrompt, freeModel, (chunk) => {
|
||||
const cleanChunk = chunk.replace(/[\u001b\u009b][[\]()#;?]*(?:[0-9]{1,4}(?:;[0-9]{0,4})*)?[0-9A-ORZcf-nqry=><]/g, '');
|
||||
|
||||
// IMPROVED STREAM SPLITTING LOGIC (Thinking vs Content)
|
||||
// Claude Code style: cleaner separation of thinking from response
|
||||
const lines = cleanChunk.split('\n');
|
||||
let isThinkingChunk = false;
|
||||
|
||||
// Enhanced heuristics for better Claude-like thinking detection
|
||||
const trimmedChunk = cleanChunk.trim();
|
||||
if (/^(Let me|Now let me|I'll|I need to|I should|I notice|I can|I will|Thinking:|Analyzing|Considering|Checking|Looking|Planning|First|Next|Finally)/i.test(trimmedChunk)) {
|
||||
isThinkingChunk = true;
|
||||
} else if (/^```|# |Here is|```|```|```/i.test(trimmedChunk)) {
|
||||
// If we encounter code blocks or headers, likely content not thinking
|
||||
isThinkingChunk = false;
|
||||
}
|
||||
|
||||
// Update character count for speed calculation
|
||||
totalCharsReceived += cleanChunk.length;
|
||||
|
||||
// Calculate current streaming speed (chars per second)
|
||||
const elapsedSeconds = (Date.now() - streamStartTime) / 1000;
|
||||
const speed = elapsedSeconds > 0 ? Math.round(totalCharsReceived / elapsedSeconds) : 0;
|
||||
|
||||
// GLOBAL STATS UPDATE (Run for ALL chunks)
|
||||
setThinkingStats(prev => ({
|
||||
...prev,
|
||||
chars: totalCharsReceived,
|
||||
speed: speed
|
||||
}));
|
||||
|
||||
// GLOBAL AGENT DETECTION (Run for ALL chunks)
|
||||
const agentMatch = cleanChunk.match(/\[AGENT:\s*([^\]]+)\]/i);
|
||||
if (agentMatch) {
|
||||
setThinkingStats(prev => ({ ...prev, activeAgent: agentMatch[1].trim() }));
|
||||
}
|
||||
|
||||
if (isThinkingChunk) {
|
||||
setThinkingLines(prev => [...prev, ...lines.map(l => l.trim()).filter(l => l && !/^(Let me|Now let me|I'll|I need to|I notice)/i.test(l.trim()))]);
|
||||
} else {
|
||||
setMessages(prev => {
|
||||
const last = prev[prev.length - 1];
|
||||
if (last && last.role === 'assistant') {
|
||||
return [...prev.slice(0, -1), { ...last, content: last.content + cleanChunk }];
|
||||
}
|
||||
return prev;
|
||||
});
|
||||
}
|
||||
})
|
||||
: await getQwen().sendMessage(fullPrompt, 'qwen-coder-plus', null, (chunk) => {
|
||||
const cleanChunk = chunk.replace(/[\u001b\u009b][[\]()#;?]*(?:[0-9]{1,4}(?:;[0-9]{0,4})*)?[0-9A-ORZcf-nqry=><]/g, '');
|
||||
|
||||
// IMPROVED STREAM SPLITTING LOGIC (Thinking vs Content)
|
||||
const lines = cleanChunk.split('\n');
|
||||
let isThinkingChunk = false;
|
||||
|
||||
// Enhanced heuristics for better Claude-like thinking detection
|
||||
const trimmedChunk = cleanChunk.trim();
|
||||
if (/^(Let me|Now let me|I'll|I need to|I should|I notice|I can|I will|Thinking:|Analyzing|Considering|Checking|Looking|Planning|First|Next|Finally)/i.test(trimmedChunk)) {
|
||||
isThinkingChunk = true;
|
||||
} else if (/^```|# |Here is|```|```|```/i.test(trimmedChunk)) {
|
||||
// If we encounter code blocks or headers, likely content not thinking
|
||||
isThinkingChunk = false;
|
||||
}
|
||||
|
||||
// Update character count for speed calculation (using same variable as OpenCode path)
|
||||
totalCharsReceived += cleanChunk.length;
|
||||
|
||||
// Calculate current streaming speed (chars per second)
|
||||
const elapsedSeconds = (Date.now() - streamStartTime) / 1000;
|
||||
const speed = elapsedSeconds > 0 ? Math.round(totalCharsReceived / elapsedSeconds) : 0;
|
||||
|
||||
setThinkingStats(prev => ({
|
||||
...prev,
|
||||
chars: totalCharsReceived,
|
||||
speed: speed
|
||||
}));
|
||||
|
||||
const agentMatch = cleanChunk.match(/\[AGENT:\s*([^\]]+)\]/i);
|
||||
if (agentMatch) {
|
||||
setThinkingStats(prev => ({ ...prev, activeAgent: agentMatch[1].trim() }));
|
||||
}
|
||||
|
||||
if (isThinkingChunk) {
|
||||
setThinkingLines(prev => [...prev, ...lines.map(l => l.trim()).filter(l => l && !/^(Let me|Now let me|I'll|I need to|I notice)/i.test(l.trim()))]);
|
||||
} else {
|
||||
setMessages(prev => {
|
||||
const last = prev[prev.length - 1];
|
||||
if (last && last.role === 'assistant') {
|
||||
return [...prev.slice(0, -1), { ...last, content: last.content + cleanChunk }];
|
||||
}
|
||||
return prev;
|
||||
});
|
||||
}
|
||||
});
|
||||
? await callOpenCodeFree(fullPromptForFree, freeModel, handleStreamChunk)
|
||||
: await getQwen().sendMessage(
|
||||
userMessage,
|
||||
'qwen-coder-plus',
|
||||
null,
|
||||
handleStreamChunk,
|
||||
systemInstruction // Pass dynamic system prompt!
|
||||
);
|
||||
|
||||
if (result.success) {
|
||||
const responseText = result.response || fullResponse;
|
||||
@@ -3701,17 +3623,17 @@ This gives the user a chance to refine requirements before implementation.
|
||||
return next;
|
||||
});
|
||||
|
||||
const successMsg = formatSuccess(`Auto-saved ${successFiles.length} file(s):\n` + successFiles.map(f => formatFileOperation(f.path, 'Saved', 'success')).join('\n'));
|
||||
setMessages(prev => [...prev, {
|
||||
role: 'system',
|
||||
content: '✅ Auto-saved ' + successFiles.length + ' file(s):\n' +
|
||||
successFiles.map(f => ' 📄 ' + f.path).join('\n')
|
||||
content: successMsg
|
||||
}]);
|
||||
}
|
||||
if (failedFiles.length > 0) {
|
||||
const failureMsg = formatError(`Failed to save ${failedFiles.length} file(s):\n` + failedFiles.map(f => ` ⚠️ ${f.filename}: ${f.error}`).join('\n'));
|
||||
setMessages(prev => [...prev, {
|
||||
role: 'system',
|
||||
content: '❌ Failed to save ' + failedFiles.length + ' file(s):\n' +
|
||||
failedFiles.map(f => ' ⚠️ ' + f.filename + ': ' + f.error).join('\n')
|
||||
role: 'error',
|
||||
content: failureMsg
|
||||
}]);
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user