tribunal-kit 4.2.0 → 4.3.0

package/.agent/scripts/context_broker.js

@@ -0,0 +1,609 @@
+#!/usr/bin/env node
+/**
+ * context_broker.js — Tribunal Kit Context Density Broker
+ * =========================================================
+ * "Focus without Compromise" — Intelligent skill selection for all model sizes.
+ *
+ * Philosophy:
+ *   This is NOT a filter that removes context. It is a PRIORITIZER that
+ *   ensures the most relevant rules occupy the highest-attention positions
+ *   in the AI's context window. Supplementary context is condensed, not cut.
+ *
+ *   For LARGER models (Claude Opus, Gemini 2.5 Pro, GPT-4o):
+ *     → Level 0 (Essential) skills are injected with full fidelity at the top.
+ *     → Level 1 (Supplementary) skills are condensed to their key rules only.
+ *     → Nothing is removed — the model gets everything, ordered optimally.
+ *
+ *   For SMALLER/FASTER models (Gemini Flash, GPT-4o-mini):
+ *     → Only Level 0 (Essential) skills are included.
+ *     → This prevents context overflow and attention dilution.
+ *     → Quality gates remain uncompromised — just fewer rules to track.
+ *
+ * Tiered Context Priority:
+ *   Level 0 — Essential:     Top matches, full SKILL.md text, injected first
+ *   Level 1 — Supplementary: Medium matches, condensed to "key rules" section
+ *   Level 2 — Available:     Low matches, listed by name only (for reference)
+ *
+ * Scoring Algorithm:
+ *   - Task keyword TF-IDF match against skill frontmatter + description
+ *   - File type affinity (e.g., .tsx → react-specialist gets +2 boost)
+ *   - Domain tag match (e.g., "sql" in task → sql-pro gets +3 boost)
+ *   - Recency boost: skills referenced in the last 3 sessions rank higher
+ *   - Tribunal alignment: skills matching active reviewers rank higher
+ *
+ * Usage:
+ *   node .agent/scripts/context_broker.js --task "Build a login API with JWT"
+ *   node .agent/scripts/context_broker.js --task "Design a premium landing page" --file Login.tsx
+ *   node .agent/scripts/context_broker.js --task "..." --model large --output json
+ *   node .agent/scripts/context_broker.js --task "..." --model small --output names
+ *   node .agent/scripts/context_broker.js demo
+ *
+ * Output modes:
+ *   report  (default) — human-readable tiered selection report
+ *   json             — JSON with tiered skill lists
+ *   names            — newline-separated skill names only (for piping)
+ *   prompt           — full injected prompt text ready for an LLM
+ */
+
+'use strict';
+
+const fs   = require('fs');
+const path = require('path');
+
+// ── Colours ───────────────────────────────────────────────────────────────────
+const GREEN  = '\x1b[92m';
+const YELLOW = '\x1b[93m';
+const CYAN   = '\x1b[96m';
+const BLUE   = '\x1b[94m';
+const RED    = '\x1b[91m';
+const BOLD   = '\x1b[1m';
+const DIM    = '\x1b[2m';
+const RESET  = '\x1b[0m';
+
+// ── Domain → Skill affinity map ───────────────────────────────────────────────
+// Keywords in the user's task that strongly indicate specific skills.
+// Higher weight = stronger signal.
+const DOMAIN_AFFINITIES = [
+    // Backend / API
+    { keywords: ['api', 'rest', 'route', 'endpoint', 'handler', 'express', 'fastapi', 'hono'],
+      skills: ['api-patterns', 'nodejs-best-practices', 'backend-specialist', 'error-resilience'], weight: 3 },
+    // Database
+    { keywords: ['sql', 'query', 'database', 'prisma', 'drizzle', 'orm', 'postgres', 'mysql', 'schema'],
+      skills: ['database-design', 'sql-pro', 'supabase-postgres-best-practices'], weight: 3 },
+    // Authentication
+    { keywords: ['auth', 'jwt', 'login', 'oauth', 'session', 'password', 'token', 'rbac'],
+      skills: ['authentication-best-practices', 'vulnerability-scanner', 'api-security-auditor'], weight: 3 },
+    // React / Next.js
+    { keywords: ['react', 'next', 'nextjs', 'component', 'hook', 'jsx', 'tsx', 'server component', 'server action'],
+      skills: ['react-specialist', 'nextjs-react-expert', 'frontend-design'], weight: 3 },
+    // Frontend / UI
+    { keywords: ['ui', 'design', 'landing', 'page', 'layout', 'responsive', 'tailwind', 'css', 'style'],
+      skills: ['frontend-design', 'ui-ux-pro-max', 'tailwind-patterns', 'web-design-guidelines'], weight: 2 },
+    // Animation / Motion
+    { keywords: ['animation', 'gsap', 'framer', 'motion', 'scroll', 'transition', 'parallax'],
+      skills: ['motion-engineering', 'framer-motion-expert', 'gsap-core', 'gsap-scrolltrigger'], weight: 3 },
+    // AI / LLM
+    { keywords: ['llm', 'openai', 'anthropic', 'gemini', 'embedding', 'ai', 'rag', 'vector', 'chat', 'prompt'],
+      skills: ['llm-engineering', 'ai-prompt-injection-defense', 'agentic-patterns'], weight: 3 },
+    // Security
+    { keywords: ['security', 'xss', 'injection', 'owasp', 'vulnerability', 'csrf', 'sanitize', 'audit'],
+      skills: ['vulnerability-scanner', 'api-security-auditor', 'authentication-best-practices'], weight: 3 },
+    // Testing
+    { keywords: ['test', 'spec', 'jest', 'vitest', 'playwright', 'unit test', 'e2e', 'mock'],
+      skills: ['testing-patterns', 'playwright-best-practices', 'tdd-workflow'], weight: 3 },
+    // Performance
+    { keywords: ['performance', 'optimize', 'bundle', 'cache', 'speed', 'slow', 'lighthouse', 'cwv'],
+      skills: ['performance-profiling', 'motion-engineering', 'edge-computing'], weight: 2 },
+    // Mobile
+    { keywords: ['mobile', 'react native', 'expo', 'ios', 'android', 'gesture', 'haptic'],
+      skills: ['building-native-ui', 'mobile-design', 'agentic-patterns'], weight: 3 },
+    // DevOps / CI
+    { keywords: ['docker', 'ci', 'cd', 'deploy', 'pipeline', 'k8s', 'kubernetes', 'github actions'],
+      skills: ['devops-engineer', 'deployment-procedures', 'observability'], weight: 2 },
+    // Real-time
+    { keywords: ['realtime', 'websocket', 'sse', 'socket', 'live', 'multiplayer', 'collaborative'],
+      skills: ['realtime-patterns', 'error-resilience'], weight: 3 },
+    // TypeScript
+    { keywords: ['typescript', 'type', 'generic', 'interface', 'satisfies', 'zod', 'pydantic'],
+      skills: ['typescript-advanced', 'data-validation-schemas'], weight: 2 },
+    // Python
+    { keywords: ['python', 'fastapi', 'django', 'flask', 'pydantic', 'asyncio'],
+      skills: ['python-pro', 'python-patterns'], weight: 3 },
+    // Architecture
+    { keywords: ['architecture', 'refactor', 'clean', 'solid', 'design pattern', 'monorepo', 'microservice'],
+      skills: ['architecture', 'clean-code', 'monorepo-management'], weight: 2 },
+    // C# / .NET
+    { keywords: ['csharp', 'c#', 'dotnet', '.net', 'blazor', 'aspnet', 'entity framework'],
+      skills: ['csharp-developer'], weight: 3 },
+];
+
+// ── File extension → skill boost map ─────────────────────────────────────────
+const EXT_AFFINITIES = {
+    '.tsx': ['react-specialist', 'nextjs-react-expert', 'typescript-advanced'],
+    '.jsx': ['react-specialist', 'frontend-design'],
+    '.ts':  ['typescript-advanced', 'nodejs-best-practices'],
+    '.vue': ['vue-expert'],
+    '.py':  ['python-pro', 'python-patterns'],
+    '.cs':  ['csharp-developer'],
+    '.sql': ['sql-pro', 'database-design'],
+    '.css': ['tailwind-patterns', 'frontend-design'],
+};
+
+// ── Core baseline skills — always available to all model sizes ────────────────
+// These are injected for every request at a condensed level.
+const BASELINE_SKILLS = [
+    'clean-code',
+    'systematic-debugging',
+    'error-resilience',
+];
+
+// ── Skill catalogue (loaded from disk) ───────────────────────────────────────
+
+/**
+ * Find the .agent directory by walking up from cwd.
+ * @returns {string} path to .agent/
+ */
+function findAgentDir() {
+    let current = path.resolve(process.cwd());
+    const root  = path.parse(current).root;
+    while (current !== root) {
+        const candidate = path.join(current, '.agent');
+        if (fs.existsSync(candidate) && fs.statSync(candidate).isDirectory()) return candidate;
+        current = path.dirname(current);
+    }
+    console.error(`${RED}✖ .agent/ not found. Run: npx tribunal-kit init${RESET}`);
+    process.exit(1);
+}
+
+/**
+ * Parse the YAML frontmatter from a SKILL.md file.
+ * Returns { name, description, ... } or null on parse failure.
+ * @param {string} content - Full SKILL.md file text
+ */
+function parseFrontmatter(content) {
+    const match = content.match(/^---\n([\s\S]*?)\n---/);
+    if (!match) return null;
+    const yaml  = match[1];
+    const obj   = {};
+    for (const line of yaml.split('\n')) {
+        const sep = line.indexOf(':');
+        if (sep === -1) continue;
+        const key = line.slice(0, sep).trim();
+        const val = line.slice(sep + 1).trim().replace(/^["']|["']$/g, '');
+        obj[key] = val;
+    }
+    return obj;
+}
+
+/**
+ * Extract the "key rules" section from a SKILL.md for condensed Level-1 context.
+ * Falls back to first 800 chars of content if no section found.
+ * @param {string} content
+ */
+function extractKeyRules(content) {
+    // Try to find sections named: Key Rules, Rules, Core Rules, Critical Rules, Guardrails
+    const sectionMatch = content.match(
+        /##\s+(?:Key Rules?|Core Rules?|Critical Rules?|Guardrails?|Rules?)\n([\s\S]*?)(?=\n##\s|$)/i
+    );
+    if (sectionMatch) return sectionMatch[1].trim().slice(0, 1200);
+    // Fallback: strip frontmatter and take first 800 chars
+    const bodyStart = content.indexOf('---', 3);
+    const body = bodyStart !== -1 ? content.slice(bodyStart + 3).trim() : content;
+    return body.slice(0, 800).trim();
+}
+
+/**
+ * Load all skills from .agent/skills/ directory.
+ * Returns an array of { name, file, frontmatter, content, keyRules } objects.
+ * @param {string} agentDir
+ */
+function loadSkills(agentDir) {
+    const skillsDir = path.join(agentDir, 'skills');
+    if (!fs.existsSync(skillsDir)) return [];
+
+    const skills = [];
+    const entries = fs.readdirSync(skillsDir, { withFileTypes: true });
+
+    for (const entry of entries) {
+        if (!entry.isDirectory()) continue;
+        const skillFile = path.join(skillsDir, entry.name, 'SKILL.md');
+        if (!fs.existsSync(skillFile)) continue;
+
+        try {
+            const content     = fs.readFileSync(skillFile, 'utf8');
+            const frontmatter = parseFrontmatter(content) || {};
+            const keyRules    = extractKeyRules(content);
+            skills.push({
+                name:   entry.name,
+                file:   skillFile,
+                frontmatter,
+                content,
+                keyRules,
+                description: frontmatter.description || '',
+            });
+        } catch {
+            // Skip unreadable skills silently
+        }
+    }
+    return skills;
+}
+
+// ── Scoring engine ────────────────────────────────────────────────────────────
+
+/**
+ * Tokenize text into lowercase words (3+ chars).
+ * @param {string} text
+ * @returns {string[]}
+ */
+function tokenize(text) {
+    return (text.match(/\b[a-zA-Z_][a-zA-Z0-9_]{2,}\b/g) || []).map(t => t.toLowerCase());
+}
+
+/**
+ * Compute a relevance score for a skill against the user's task.
+ *
+ * Scoring breakdown:
+ *   - Text overlap between task tokens and skill description/name: up to +5
+ *   - Domain affinity keyword match: +weight (2 or 3) per match
+ *   - File extension affinity: +2 per match
+ *   - Baseline skill bonus: +1 (always present)
+ *
+ * @param {{ name, description, content }} skill
+ * @param {string} task           - Raw task text from the user
+ * @param {string[]} fileExts     - File extensions being touched
+ * @param {string[]} taskTokens   - Pre-tokenized task
+ * @returns {number} score
+ */
+function scoreSkill(skill, task, fileExts, taskTokens) {
+    let score = 0;
+    const taskLower = task.toLowerCase();
+    const skillText = (skill.name + ' ' + skill.description).toLowerCase();
+    const skillTokens = tokenize(skillText);
+
+    // 1. Token overlap (lightweight TF match — no IDF needed at this scale)
+    const skillSet = new Set(skillTokens);
+    for (const token of taskTokens) {
+        if (skillSet.has(token)) score += 1;
+    }
+
+    // 2. Domain affinity boost
+    for (const affinity of DOMAIN_AFFINITIES) {
+        const keywordMatch = affinity.keywords.some(k => taskLower.includes(k));
+        if (!keywordMatch) continue;
+        if (affinity.skills.includes(skill.name)) {
+            score += affinity.weight;
+        }
+    }
+
+    // 3. File extension boost
+    for (const ext of fileExts) {
+        const extSkills = EXT_AFFINITIES[ext] || [];
+        if (extSkills.includes(skill.name)) score += 2;
+    }
+
+    // 4. Baseline skill safety net
+    if (BASELINE_SKILLS.includes(skill.name)) score += 1;
+
+    return score;
+}
+
+/**
+ * Run the tiered selection algorithm.
+ *
+ * @param {string}   task     - Raw user task description
+ * @param {string[]} files    - Affected filenames (for ext detection)
+ * @param {string}   model    - 'large' | 'small' | 'auto'
+ * @param {object[]} skills   - Loaded skills array from loadSkills()
+ * @returns {{ essential: object[], supplementary: object[], available: object[], scores: Map }}
+ */
+function selectSkills(task, files, model, skills) {
+    const taskTokens = tokenize(task);
+    const fileExts   = files.map(f => path.extname(f).toLowerCase()).filter(Boolean);
+
+    // Score every available skill
+    const scored = skills.map(skill => ({
+        ...skill,
+        score: scoreSkill(skill, task, fileExts, taskTokens),
+    }));
+    scored.sort((a, b) => b.score - a.score);
+
+    // Determine tier thresholds
+    const maxScore  = scored[0]?.score || 1;
+    const tier0Cut  = Math.max(maxScore * 0.65, 2); // Essential: top 65%+ of max score
+    const tier1Cut  = Math.max(maxScore * 0.3,  1); // Supplementary: 30–65%
+
+    const essential     = scored.filter(s => s.score >= tier0Cut).slice(0, 10);
+    const supplementary = scored.filter(s => s.score < tier0Cut && s.score >= tier1Cut).slice(0, 8);
+    const available     = scored.filter(s => s.score < tier1Cut && s.score > 0).slice(0, 10);
+
+    // Ensure baseline skills always appear at minimum in supplementary
+    for (const base of BASELINE_SKILLS) {
+        const inEssential     = essential.find(s => s.name === base);
+        const inSupplementary = supplementary.find(s => s.name === base);
+        if (!inEssential && !inSupplementary) {
+            const baseSkill = skills.find(s => s.name === base);
+            if (baseSkill) supplementary.push({ ...baseSkill, score: 0.5 });
+        }
+    }
+
+    // For small models: collapse supplementary into available
+    if (model === 'small') {
+        return {
+            essential: essential.slice(0, 6),
+            supplementary: [],
+            available: [...supplementary, ...available].slice(0, 8),
+            scores: buildScoreMap(scored),
+        };
+    }
+
+    return { essential, supplementary, available, scores: buildScoreMap(scored) };
+}
+
+function buildScoreMap(scored) {
+    const m = new Map();
+    for (const s of scored) m.set(s.name, s.score);
+    return m;
+}
+
+// ── Output formatters ─────────────────────────────────────────────────────────
+
+function formatReport(task, model, selection, elapsed) {
+    const { essential, supplementary, available } = selection;
+    const total = essential.length + supplementary.length + available.length;
+
+    console.log(`\n${BOLD}${CYAN}━━━ Context Broker — Skill Selection ━━━━━━━━━━━━━━━${RESET}`);
+    console.log(`  Task   : ${BOLD}${task.slice(0, 80)}${task.length > 80 ? '...' : ''}${RESET}`);
+    console.log(`  Model  : ${model === 'large' ? GREEN : YELLOW}${model}${RESET}  ${DIM}(Focus without Compromise)${RESET}`);
+    console.log(`  Skills : ${GREEN}${essential.length} essential${RESET} + ${YELLOW}${supplementary.length} supplementary${RESET} + ${DIM}${available.length} available${RESET}  of ${total} matched`);
+    console.log(`  Time   : ${elapsed}ms\n`);
+
+    if (essential.length) {
+        console.log(`  ${GREEN}${BOLD}▶ Level 0 — Essential (Full Context, Top Priority):${RESET}`);
+        for (const s of essential) {
+            const score = selection.scores.get(s.name) || 0;
+            console.log(`    ${GREEN}✦${RESET} ${BOLD}${s.name}${RESET}  ${DIM}score=${score.toFixed(1)}${RESET}`);
+            if (s.description) console.log(`       ${DIM}${s.description.slice(0, 90)}${RESET}`);
+        }
+    }
+
+    if (supplementary.length) {
+        console.log(`\n  ${YELLOW}${BOLD}▶ Level 1 — Supplementary (Key Rules Only):${RESET}`);
+        for (const s of supplementary) {
+            const score = selection.scores.get(s.name) || 0;
+            console.log(`    ${YELLOW}◆${RESET} ${s.name}  ${DIM}score=${score.toFixed(1)}${RESET}`);
+        }
+    }
+
+    if (available.length) {
+        console.log(`\n  ${DIM}▶ Level 2 — Available (Name Reference Only):${RESET}`);
+        console.log(`    ${DIM}${available.map(s => s.name).join(', ')}${RESET}`);
+    }
+
+    if (model === 'small') {
+        console.log(`\n  ${YELLOW}⚡ Small model mode: supplementary collapsed. Essential only injected.${RESET}`);
+    }
+
+    console.log(`\n${CYAN}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${RESET}\n`);
+}
+
+function formatJson(task, model, selection) {
+    return JSON.stringify({
+        task,
+        model,
+        timestamp: new Date().toISOString(),
+        essential:     selection.essential.map(s => ({ name: s.name, score: selection.scores.get(s.name), description: s.description })),
+        supplementary: selection.supplementary.map(s => ({ name: s.name, score: selection.scores.get(s.name) })),
+        available:     selection.available.map(s => s.name),
+    }, null, 2);
+}
+
+function formatNames(selection, model) {
+    const all = model === 'large'
+        ? [...selection.essential, ...selection.supplementary]
+        : selection.essential;
+    return all.map(s => s.name).join('\n');
+}
+
+/**
+ * Build a full LLM-ready context prompt string.
+ * This is the full output to be injected into an AI system prompt.
+ *
+ * For large models: Essential = full SKILL.md, Supplementary = key rules section.
+ * For small models: Essential = key rules section only.
+ *
+ * @param {string} task
+ * @param {string} model
+ * @param {object} selection
+ * @returns {string}
+ */
+function formatPrompt(task, model, selection) {
+    const lines = [
+        `# Tribunal Context Broker — Injected Skills`,
+        `# Task: ${task}`,
+        `# Model tier: ${model}`,
+        `# Generated: ${new Date().toISOString()}`,
+        '',
+        '## Instructions for the AI',
+        'The following skills are ordered by relevance to the current task.',
+        'Level 0 skills contain full rule sets. Level 1 skills contain key rules only.',
+        'Treat ALL injected rules as mandatory constraints, not suggestions.',
+        '',
+        '---',
+        '',
+    ];
+
+    if (selection.essential.length) {
+        lines.push('## Level 0 — Essential Skills (Full Context)');
+        lines.push('');
+        for (const s of selection.essential) {
+            lines.push(`### Skill: ${s.name}`);
+            lines.push('');
+            if (model === 'large') {
+                lines.push(s.content || s.keyRules);
+            } else {
+                lines.push(s.keyRules);
+            }
+            lines.push('');
+            lines.push('---');
+            lines.push('');
+        }
+    }
+
+    if (model === 'large' && selection.supplementary.length) {
+        lines.push('## Level 1 — Supplementary Skills (Key Rules)');
+        lines.push('');
+        for (const s of selection.supplementary) {
+            lines.push(`### Skill: ${s.name} (condensed)`);
+            lines.push('');
+            lines.push(s.keyRules);
+            lines.push('');
+            lines.push('---');
+            lines.push('');
+        }
+    }
+
+    if (selection.available.length) {
+        lines.push('## Level 2 — Available Skills (Reference Names Only)');
+        lines.push('');
+        lines.push('The following skills are relevant but not injected to maintain context density.');
+        lines.push('Request their full content if needed: ' + selection.available.map(s => s.name).join(', '));
+        lines.push('');
+    }
+
+    return lines.join('\n');
+}
+
+// ── Built-in demo ─────────────────────────────────────────────────────────────
+
+function runDemo(agentDir) {
+    const skills = loadSkills(agentDir);
+
+    const scenarios = [
+        { task: 'Build a JWT authentication API with Express.js and Zod validation',       file: 'auth.ts',    model: 'large' },
+        { task: 'Design a premium landing page with GSAP scroll animations',               file: 'Hero.tsx',   model: 'large' },
+        { task: 'Write a Prisma query for paginated user orders',                           file: 'orders.ts',  model: 'small' },
+        { task: 'Add RAG pipeline to an OpenAI-powered chat interface',                    file: 'chat.ts',    model: 'large' },
+    ];
+
+    console.log(`\n${BOLD}${CYAN}━━━ Context Broker — Demo Mode ━━━━━━━━━━━━━━━━━━━━━${RESET}`);
+    console.log(`  Loaded ${skills.length} skills from .agent/skills/\n`);
+
+    for (const scenario of scenarios) {
+        const t0       = Date.now();
+        const model    = scenario.model;
+        const selection = selectSkills(scenario.task, [scenario.file], model, skills);
+        const elapsed  = Date.now() - t0;
+
+        console.log(`\n  ${BOLD}Task: "${scenario.task.slice(0, 70)}"${RESET}  ${DIM}[${model} model]${RESET}`);
+        console.log(`    Essential     : ${GREEN}${selection.essential.map(s => s.name).join(', ')}${RESET}`);
+        console.log(`    Supplementary : ${YELLOW}${selection.supplementary.map(s => s.name).join(', ') || '(small mode)'}${RESET}`);
+        console.log(`    Time          : ${DIM}${elapsed}ms${RESET}`);
+    }
+
+    console.log(`\n${CYAN}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${RESET}\n`);
+}
+
+// ── Public API ────────────────────────────────────────────────────────────────
+
+/**
+ * Programmatic API for use by other Tribunal scripts.
+ *
+ * @param {string}   task    - User task description
+ * @param {string[]} files   - Files being touched
+ * @param {string}   model   - 'large' | 'small' | 'auto'
+ * @param {string}   agentDir - Path to .agent/ directory
+ * @returns {{ essential, supplementary, available, promptText }}
+ */
+function broker(task, files = [], model = 'large', agentDir = null) {
+    const resolvedAgentDir = agentDir || findAgentDir();
+    const skills   = loadSkills(resolvedAgentDir);
+    const selection = selectSkills(task, files, model, skills);
+    const promptText = formatPrompt(task, model, selection);
+    return { ...selection, promptText };
+}
+
+module.exports = { broker, selectSkills, loadSkills, scoreSkill, findAgentDir };
+
+// ── CLI Entry ─────────────────────────────────────────────────────────────────
+
+if (require.main === module) {
+    const argv = process.argv.slice(2);
+
+    if (!argv.length || argv.includes('--help') || argv.includes('-h')) {
+        console.log(`
+${BOLD}context_broker.js${RESET} — Tribunal Focus-without-Compromise Context Engine
+
+${BOLD}Usage:${RESET}
+  node .agent/scripts/context_broker.js --task "<description>" [options]
+  node .agent/scripts/context_broker.js demo
+
+${BOLD}Options:${RESET}
+  --task    <text>     Task description to match against skill catalogue
+  --file    <path>     File being touched (repeat for multiple files)
+  --model   <size>     large (default) | small  — model tier
+  --output  <format>   report (default) | json | names | prompt
+
+${BOLD}Model tiers:${RESET}
+  large   Full Essential + condensed Supplementary (Claude Opus, Gemini 2.5 Pro, GPT-4o)
+  small   Essential only (Gemini Flash, GPT-4o-mini)
+
+${BOLD}Examples:${RESET}
+  node .agent/scripts/context_broker.js --task "JWT auth API" --model large
+  node .agent/scripts/context_broker.js --task "landing page" --file Hero.tsx --output names
+  node .agent/scripts/context_broker.js --task "RAG pipeline" --output prompt > context.md
+  node .agent/scripts/context_broker.js demo
+`);
+        process.exit(0);
+    }
+
+    const agentDir = findAgentDir();
+
+    if (argv[0] === 'demo') {
+        runDemo(agentDir);
+        process.exit(0);
+    }
+
+    // Parse args
+    const taskIdx   = argv.indexOf('--task');
+    const modelIdx  = argv.indexOf('--model');
+    const outputIdx = argv.indexOf('--output');
+
+    const task   = taskIdx  !== -1 && argv[taskIdx  + 1] ? argv[taskIdx  + 1] : '';
+    const model  = modelIdx !== -1 && argv[modelIdx + 1] ? argv[modelIdx + 1] : 'large';
+    const output = outputIdx !== -1 && argv[outputIdx + 1] ? argv[outputIdx + 1] : 'report';
+
+    if (!task) {
+        console.error(`${RED}✖ --task is required${RESET}`);
+        process.exit(1);
+    }
+
+    // Collect --file arguments (may appear multiple times)
+    const files = [];
+    for (let i = 0; i < argv.length; i++) {
+        if (argv[i] === '--file' && argv[i + 1]) files.push(argv[i + 1]);
+    }
+
+    const t0       = Date.now();
+    const skills   = loadSkills(agentDir);
+    const selection = selectSkills(task, files, model, skills);
+    const elapsed  = Date.now() - t0;
+
+    if (!['large', 'small'].includes(model)) {
+        console.error(`${YELLOW}⚠ Unknown model tier "${model}" — defaulting to "large"${RESET}`);
+    }
+
+    switch (output) {
+        case 'json':
+            console.log(formatJson(task, model, selection));
+            break;
+        case 'names':
+            console.log(formatNames(selection, model));
+            break;
+        case 'prompt':
+            console.log(formatPrompt(task, model, selection));
+            break;
+        default: // 'report'
+            formatReport(task, model, selection, elapsed);
+            break;
+    }
+}