@covibes/zeroshot 5.3.0 → 5.4.1

package/src/agent/agent-context-builder.js

@@ -1,335 +0,0 @@
-/**
- * AgentContextBuilder - Build agent execution context from ledger
- *
- * Provides:
- * - Context assembly from multiple message sources
- * - Context strategy evaluation (topics, limits, since timestamps)
- * - Prompt injection and formatting
- * - Token-based truncation
- * - Defensive context overflow prevention
- */
-
-// Defensive limit: 500,000 chars ≈ 125k tokens (safe buffer below 200k limit)
-// Prevents "Prompt is too long" errors that kill tasks
-const MAX_CONTEXT_CHARS = 500000;
-
-/**
- * Generate an example object from a JSON schema
- * Used to show models a concrete example of expected output
- *
- * @param {object} schema - JSON schema
- * @returns {object|null} Example object or null if generation fails
- */
-function generateExampleFromSchema(schema) {
-  if (!schema || schema.type !== 'object' || !schema.properties) {
-    return null;
-  }
-
-  const example = {};
-
-  for (const [key, propSchema] of Object.entries(schema.properties)) {
-    if (propSchema.enum && propSchema.enum.length > 0) {
-      // Use first enum value as example
-      example[key] = propSchema.enum[0];
-    } else if (propSchema.type === 'string') {
-      example[key] = propSchema.description || `${key} value`;
-    } else if (propSchema.type === 'boolean') {
-      example[key] = true;
-    } else if (propSchema.type === 'number' || propSchema.type === 'integer') {
-      example[key] = 0;
-    } else if (propSchema.type === 'array') {
-      if (propSchema.items?.type === 'string') {
-        example[key] = [];
-      } else {
-        example[key] = [];
-      }
-    } else if (propSchema.type === 'object') {
-      example[key] = generateExampleFromSchema(propSchema) || {};
-    }
-  }
-
-  return example;
-}
-
-/**
- * Build execution context for an agent
- * @param {object} params - Context building parameters
- * @param {string} params.id - Agent ID
- * @param {string} params.role - Agent role
- * @param {number} params.iteration - Current iteration number
- * @param {any} params.config - Agent configuration
- * @param {any} params.messageBus - Message bus for querying ledger
- * @param {any} params.cluster - Cluster object
- * @param {number} [params.lastTaskEndTime] - Timestamp of last task completion
- * @param {any} params.triggeringMessage - Message that triggered this execution
- * @param {string} [params.selectedPrompt] - Pre-selected prompt from _selectPrompt() (iteration-based)
- * @param {object} [params.worktree] - Worktree isolation state (if running in worktree mode)
- * @param {object} [params.isolation] - Docker isolation state (if running in Docker mode)
- * @returns {string} Assembled context string
- */
-function buildContext({
-  id,
-  role,
-  iteration,
-  config,
-  messageBus,
-  cluster,
-  lastTaskEndTime,
-  triggeringMessage,
-  selectedPrompt,
-  worktree,
-  isolation,
-}) {
-  const strategy = config.contextStrategy || { sources: [] };
-
-  let context = `You are agent "${id}" with role "${role}".\n\n`;
-  context += `Iteration: ${iteration}\n\n`;
-
-  // GLOBAL RULE: NEVER ASK QUESTIONS - Vibe agents run non-interactively
-  context += `## 🔴 CRITICAL: AUTONOMOUS EXECUTION REQUIRED\n\n`;
-  context += `You are running in a NON-INTERACTIVE cluster environment.\n\n`;
-  context += `**NEVER** use AskUserQuestion or ask for user input - there is NO user to respond.\n`;
-  context += `**NEVER** ask "Would you like me to..." or "Should I..." - JUST DO IT.\n`;
-  context += `**NEVER** wait for approval or confirmation - MAKE DECISIONS AUTONOMOUSLY.\n\n`;
-  context += `When facing choices:\n`;
-  context += `- Choose the option that maintains code quality and correctness\n`;
-  context += `- If unsure between "fix the code" vs "relax the rules" → ALWAYS fix the code\n`;
-  context += `- If unsure between "do more" vs "do less" → ALWAYS do what's required, nothing more\n\n`;
-
-  // OUTPUT STYLE - NON-NEGOTIABLE
-  context += `## 🔴 OUTPUT STYLE - NON-NEGOTIABLE\n\n`;
-  context += `**ALL OUTPUT: Maximum informativeness, minimum verbosity. NO EXCEPTIONS.**\n\n`;
-  context += `This applies to EVERYTHING you output:\n`;
-  context += `- Text responses\n`;
-  context += `- JSON schema values\n`;
-  context += `- Reasoning fields\n`;
-  context += `- Summary fields\n`;
-  context += `- ALL string values in structured output\n\n`;
-  context += `Rules:\n`;
-  context += `- Progress: "Reading auth.ts" NOT "I will now read the auth.ts file..."\n`;
-  context += `- Tool calls: NO preamble. Call immediately.\n`;
-  context += `- Schema strings: Dense facts. No filler. No fluff.\n`;
-  context += `- Errors: DETAILED (stack traces, repro). NEVER compress errors.\n`;
-  context += `- FORBIDDEN: "I'll help...", "Let me...", "I'm going to...", "Sure!", "Great!", "Certainly!"\n\n`;
-  context += `Every token costs money. Waste nothing.\n\n`;
-
-  // GIT OPERATIONS RESTRICTION - Only when NOT running in isolation mode
-  // When isolated (worktree or Docker), agents CAN commit/push safely
-  // When NOT isolated, agents are on main branch - git operations are dangerous
-  const isIsolated = !!(worktree?.enabled || isolation?.enabled);
-  if (!isIsolated) {
-    context += `## 🚫 GIT OPERATIONS - FORBIDDEN\n\n`;
-    context += `NEVER commit, push, or create PRs. You only modify files.\n`;
-    context += `The git-pusher agent handles ALL git operations AFTER validators approve.\n\n`;
-    context += `- ❌ NEVER run: git add, git commit, git push, gh pr create\n`;
-    context += `- ❌ NEVER suggest committing changes\n`;
-    context += `- ✅ Only modify files and publish your completion message when done\n\n`;
-  }
-
-  // Add prompt from config (system prompt, instructions, output format)
-  // If selectedPrompt is provided (iteration-based), use it directly
-  // Otherwise fall back to legacy config.prompt handling
-  const promptText =
-    selectedPrompt || (typeof config.prompt === 'string' ? config.prompt : config.prompt?.system);
-
-  if (promptText) {
-    context += `## Instructions\n\n${promptText}\n\n`;
-  } else if (config.prompt && typeof config.prompt !== 'string' && !config.prompt?.system) {
-    // FAIL HARD: prompt exists but format is unrecognized (and no selectedPrompt provided)
-    throw new Error(
-      `Agent "${id}" has invalid prompt format. ` +
-        `Expected string or object with .system property, got: ${JSON.stringify(config.prompt).slice(0, 100)}...`
-    );
-  }
-
-  // Output format schema (if configured via legacy format)
-  if (config.prompt?.outputFormat) {
-    context += `## Output Schema (REQUIRED)\n\n`;
-    context += `\`\`\`json\n${JSON.stringify(config.prompt.outputFormat.example, null, 2)}\n\`\`\`\n\n`;
-    context += `STRING VALUES IN THIS SCHEMA: Dense. Factual. No filler words. No pleasantries.\n`;
-    if (config.prompt.outputFormat.rules) {
-      for (const rule of config.prompt.outputFormat.rules) {
-        context += `- ${rule}\n`;
-      }
-    }
-    context += '\n';
-  }
-
-  // AUTO-INJECT JSON OUTPUT INSTRUCTIONS when jsonSchema is defined
-  // This ensures ALL agents with structured output schemas get explicit "output ONLY JSON" instructions
-  // Critical for less capable models (Codex, Gemini) that output prose without explicit instructions
-  if (config.jsonSchema && config.outputFormat === 'json') {
-    context += `## 🔴 OUTPUT FORMAT - JSON ONLY\n\n`;
-    context += `Your response must be ONLY valid JSON. No other text before or after.\n`;
-    context += `Start with { and end with }. Nothing else.\n\n`;
-    context += `Required schema:\n`;
-    context += `\`\`\`json\n${JSON.stringify(config.jsonSchema, null, 2)}\n\`\`\`\n\n`;
-
-    // Generate example from schema
-    const example = generateExampleFromSchema(config.jsonSchema);
-    if (example) {
-      context += `Example output:\n`;
-      context += `\`\`\`json\n${JSON.stringify(example, null, 2)}\n\`\`\`\n\n`;
-    }
-
-    context += `CRITICAL RULES:\n`;
-    context += `- Output ONLY the JSON object - no explanation, no thinking, no preamble\n`;
-    context += `- Use EXACTLY the enum values specified (case-sensitive)\n`;
-    context += `- Include ALL required fields\n\n`;
-  }
-
-  // Add sources
-  for (const source of strategy.sources) {
-    // Resolve special 'since' values
-    let sinceTimestamp = source.since;
-    if (source.since === 'cluster_start') {
-      sinceTimestamp = cluster.createdAt;
-    } else if (source.since === 'last_task_end') {
-      // Use timestamp of last task completion, or cluster start if no tasks completed yet
-      sinceTimestamp = lastTaskEndTime || cluster.createdAt;
-    }
-
-    const messages = messageBus.query({
-      cluster_id: cluster.id,
-      topic: source.topic,
-      sender: source.sender,
-      since: sinceTimestamp,
-      limit: source.limit,
-    });
-
-    if (messages.length > 0) {
-      context += `\n## Messages from topic: ${source.topic}\n\n`;
-      for (const msg of messages) {
-        context += `[${new Date(msg.timestamp).toISOString()}] ${msg.sender}:\n`;
-        if (msg.content?.text) {
-          context += `${msg.content.text}\n`;
-        }
-        if (msg.content?.data) {
-          context += `Data: ${JSON.stringify(msg.content.data, null, 2)}\n`;
-        }
-        context += '\n';
-      }
-    }
-  }
-
-  // Add triggering message
-  context += `\n## Triggering Message\n\n`;
-  context += `Topic: ${triggeringMessage.topic}\n`;
-  context += `Sender: ${triggeringMessage.sender}\n`;
-  if (triggeringMessage.content?.text) {
-    context += `\n${triggeringMessage.content.text}\n`;
-  }
-
-  // DEFENSIVE TRUNCATION - Prevent context overflow errors
-  // Strategy: Keep ISSUE_OPENED (original task) + most recent messages
-  // Truncate from MIDDLE (oldest context messages) if too long
-  const originalLength = context.length;
-
-  if (originalLength > MAX_CONTEXT_CHARS) {
-    console.log(
-      `[Context] Context too large (${originalLength} chars), truncating to prevent overflow...`
-    );
-
-    // Split context into sections
-    const lines = context.split('\n');
-
-    // Find critical sections that must be preserved
-    let issueOpenedStart = -1;
-    let issueOpenedEnd = -1;
-    let triggeringStart = -1;
-
-    for (let i = 0; i < lines.length; i++) {
-      if (lines[i].includes('## Messages from topic: ISSUE_OPENED')) {
-        issueOpenedStart = i;
-      }
-      if (issueOpenedStart !== -1 && issueOpenedEnd === -1 && lines[i].startsWith('## ')) {
-        issueOpenedEnd = i;
-      }
-      if (lines[i].includes('## Triggering Message')) {
-        triggeringStart = i;
-        break;
-      }
-    }
-
-    // Build truncated context:
-    // 1. Header (agent info, instructions, output format)
-    // 2. ISSUE_OPENED message (original task - NEVER truncate)
-    // 3. Most recent N messages (whatever fits in budget)
-    // 4. Triggering message (current event)
-
-    const headerEnd = issueOpenedStart !== -1 ? issueOpenedStart : triggeringStart;
-    const header = lines.slice(0, headerEnd).join('\n');
-
-    const issueOpened =
-      issueOpenedStart !== -1 && issueOpenedEnd !== -1
-        ? lines.slice(issueOpenedStart, issueOpenedEnd).join('\n')
-        : '';
-
-    const triggeringMsg = lines.slice(triggeringStart).join('\n');
-
-    // Calculate remaining budget for recent messages
-    const fixedSize = header.length + issueOpened.length + triggeringMsg.length;
-    const budgetForRecent = MAX_CONTEXT_CHARS - fixedSize - 200; // 200 char buffer for markers
-
-    // Collect recent messages (from end backwards until budget exhausted)
-    const recentLines = [];
-    let recentSize = 0;
-
-    const middleStart = issueOpenedEnd !== -1 ? issueOpenedEnd : headerEnd;
-    const middleEnd = triggeringStart;
-    const middleLines = lines.slice(middleStart, middleEnd);
-
-    for (let i = middleLines.length - 1; i >= 0; i--) {
-      const line = middleLines[i];
-      const lineSize = line.length + 1; // +1 for newline
-
-      if (recentSize + lineSize > budgetForRecent) {
-        break; // Budget exhausted
-      }
-
-      recentLines.unshift(line);
-      recentSize += lineSize;
-    }
-
-    // Assemble truncated context
-    const parts = [header];
-
-    if (issueOpened) {
-      parts.push(issueOpened);
-    }
-
-    if (recentLines.length < middleLines.length) {
-      // Some messages were truncated
-      const truncatedCount = middleLines.length - recentLines.length;
-      parts.push(
-        `\n[...${truncatedCount} earlier context messages truncated to prevent overflow...]\n`
-      );
-    }
-
-    if (recentLines.length > 0) {
-      parts.push(recentLines.join('\n'));
-    }
-
-    parts.push(triggeringMsg);
-
-    context = parts.join('\n');
-
-    const truncatedLength = context.length;
-    console.log(
-      `[Context] Truncated from ${originalLength} to ${truncatedLength} chars (${Math.round((truncatedLength / originalLength) * 100)}% retained)`
-    );
-  }
-
-  // Legacy maxTokens check (for backward compatibility with agent configs)
-  const maxTokens = strategy.maxTokens || 100000;
-  const maxChars = maxTokens * 4;
-  if (context.length > maxChars) {
-    context = context.slice(0, maxChars) + '\n\n[Context truncated...]';
-  }
-
-  return context;
-}
-
-module.exports = {
-  buildContext,
-};