tribunal-kit 4.4.0 → 4.4.2

package/.agent/scripts/skill_evolution.js

@@ -1,644 +1,637 @@
-#!/usr/bin/env node
-/**
- * skill_evolution.js — Tribunal Kit Skill Evolution Forge
- * =========================================================
- * Analyzes the delta between what the AI proposed and what the developer
- * actually committed, then distills those decisions into evolving
- * project-specific SKILL idioms — WITHOUT sending full files to any LLM.
- *
- * Core Strategy: Semantic Delta Extraction
- *   1. Read the raw git diff of staged/recent changes
- *   2. Strip trivial noise (whitespace, comments, import renames)
- *   3. Score remaining lines for "Architectural Weight"
- *   4. Only high-weight deltas reach the LLM reflection prompt
- *   5. LLM returns structured YAML idiom entries (not prose)
- *   6. Idioms are merged into .agent/skills/project-idioms/SKILL.md
- *
- * Usage:
- *   node .agent/scripts/skill_evolution.js digest
- *   node .agent/scripts/skill_evolution.js digest --dry-run
- *   node .agent/scripts/skill_evolution.js show
- *   node .agent/scripts/skill_evolution.js reset
- *   node .agent/scripts/skill_evolution.js status
- */
-
-'use strict';
-
-const fs      = require('fs');
-const path    = require('path');
-const https   = require('https');
-const { execSync } = require('child_process');
-const readline = require('readline');
-
-// ── Colours ──────────────────────────────────────────────────────────────────
-const GREEN  = '\x1b[92m';
-const YELLOW = '\x1b[93m';
-const CYAN   = '\x1b[96m';
-const RED    = '\x1b[91m';
-const BLUE   = '\x1b[94m';
-const BOLD   = '\x1b[1m';
-const DIM    = '\x1b[2m';
-const RESET  = '\x1b[0m';
-
-// ── Find .agent directory ─────────────────────────────────────────────────────
-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(`${RED}✖ Error: '.agent' directory not found. Please run 'npx tribunal-kit init' first.${RESET}`);
-    process.exit(1);
-}
-
-const AGENT_DIR   = findAgentDir();
-const SKILL_DIR   = path.join(AGENT_DIR, 'skills', 'project-idioms');
-const SKILL_FILE  = path.join(SKILL_DIR, 'SKILL.md');
-const HISTORY_DIR = path.join(AGENT_DIR, 'history', 'skill-evolution');
-const LOG_FILE    = path.join(HISTORY_DIR, 'digest-log.json');
-
-// ── Architectural Weight Patterns ────────────────────────────────────────────
-const HIGH_WEIGHT_PATTERNS = [
-    /\bclass\b/, /\binterface\b/, /\btype\s+\w+\s*=/, /\bextends\b/, /\bimplements\b/,
-    /\bthrow\b/, /\bcatch\b/, /\btry\b/,
-    /\bprisma\.\w+\(/, /\bsupabase\./, /\bfetch\(/, /\baxios\./,
-    /\bReturnType\b/, /\bPromise</, /\basync\s+function/, /\bawait\b/,
-    /\bexport\s+(default\s+)?(class|function|const)/, /\bmodule\.exports\b/,
-    /\bRouter\b|\bapp\.(get|post|put|delete|patch)\(/,
-    /\buse[A-Z]\w+\(/, /\bcreateContext\(/,
-    /\bz\.object\(/, /\bPrisma\b|\bdrizzle\b/,
-    /\benv\.\w+/, /\bprocess\.env\./,
-];
-
-const NOISE_PATTERNS = [
-    /^\s*$/, /^\s*(\/\/|#|\/\*).*$/, /^\s*\*/,
-    /^\s*import\s+\{[^}]+\}\s+from\s+['"](?!\.)/, /^\s*(console\.(log|warn|error)|print\()/,
-    /^\s*\w+\s*[:,]?\s*$/,
-];
-
-function architecturalWeight(line) {
-    const code = line.replace(/^[+-]/, '').trim();
-    for (const p of NOISE_PATTERNS) { if (p.test(code)) return 0; }
-    for (const p of HIGH_WEIGHT_PATTERNS) { if (p.test(code)) return 2; }
-    return 1;
-}
-
-// ── Levenshtein Semantic Deduplication ───────────────────────────────────────
-// FIX: Replaces over-aggressive substring matching (.includes) with normalised
-// edit-distance similarity. A new idiom must differ by >= 20% from all existing
-// ones to be accepted. Threshold 0.80 = 80% similar → considered a duplicate.
-
-function levenshtein(a, b) {
-    const m = a.length, n = b.length;
-    // Allocate DP table with base cases pre-filled
-    const dp = Array.from({ length: m + 1 }, (_, i) =>
-        Array.from({ length: n + 1 }, (_, j) => i === 0 ? j : j === 0 ? i : 0)
-    );
-    for (let i = 1; i <= m; i++) {
-        for (let j = 1; j <= n; j++) {
-            dp[i][j] = a[i - 1] === b[j - 1]
-                ? dp[i - 1][j - 1]
-                : 1 + Math.min(dp[i - 1][j], dp[i][j - 1], dp[i - 1][j - 1]);
-        }
-    }
-    return dp[m][n];
-}
-
-function normalizedSimilarity(a, b) {
-    if (!a.length && !b.length) return 1.0;
-    const maxLen = Math.max(a.length, b.length);
-    if (maxLen === 0) return 1.0;
-    return 1 - levenshtein(a, b) / maxLen;
-}
-
-/**
- * Returns true if newPattern is semantically similar to any existing idiom.
- * Uses Levenshtein normalised similarity with a configurable threshold.
- * @param {string}   newPattern
- * @param {object[]} existingIdioms - array of { pattern } objects
- * @param {number}   threshold      - 0.0–1.0 (default 0.80 = 80% similar = duplicate)
- */
-function isDuplicateIdiom(newPattern, existingIdioms, threshold = 0.80) {
-    const newLow = newPattern.toLowerCase();
-    return existingIdioms.some(ex => {
-        const exLow = (ex.pattern || '').toLowerCase();
-        return normalizedSimilarity(newLow, exLow) >= threshold;
-    });
-}
-
-function semanticDelta(diffText, minWeight = 2) {
-    const lines = diffText.split('\n');
-    const kept = [];
-    let currentHunkHasHigh = false;
-    let hunkLines = [];
-
-    for (const line of lines) {
-        if (line.startsWith('---') || line.startsWith('+++') || line.startsWith('diff --git')) {
-            kept.push(line); continue;
-        }
-        if (line.startsWith('@@')) {
-            if (currentHunkHasHigh) kept.push(...hunkLines);
-            currentHunkHasHigh = false;
-            hunkLines = [line]; continue;
-        }
-        if (line.startsWith('+') || line.startsWith('-')) {
-            const w = architecturalWeight(line);
-            hunkLines.push(line);
-            if (w >= minWeight) currentHunkHasHigh = true;
-        } else {
-            hunkLines.push(line);
-        }
-    }
-    if (currentHunkHasHigh) kept.push(...hunkLines);
-
-    let result = kept.join('\n');
-    result = result.replace(/\n( ?\n){3,}/g, '\n\n');
-    return result.trim();
-}
-
-// ── Git helpers ────────────────────────────────────────────────────────────────
-function getGitDiff(mode = 'staged') {
-    try {
-        let cmd;
-        if (mode === 'staged')    cmd = 'git diff --cached --unified=3';
-        else if (mode === 'head') cmd = 'git diff HEAD~1 HEAD --unified=3';
-        else                      cmd = 'git diff --unified=3';
-        return execSync(cmd, { encoding: 'utf8', timeout: 10000, stdio: ['pipe', 'pipe', 'pipe'] });
-    } catch { return ''; }
-}
-
-function countTokensEstimate(text) {
-    return Math.max(1, Math.floor(text.length / 4));
-}
-
-// ── Idiom management ──────────────────────────────────────────────────────────
-function loadExistingIdioms() {
-    if (!fs.existsSync(SKILL_FILE)) return [];
-    const content = fs.readFileSync(SKILL_FILE, 'utf8');
-    const idioms = [];
-    const pattern = /\|\s*(\d+)\s*\|\s*`([^`]+)`\s*\|\s*([^|]+)\|\s*([^|]+)\|\s*([^|]+)\|/g;
-    let m;
-    while ((m = pattern.exec(content)) !== null) {
-        idioms.push({
-            id: parseInt(m[1], 10), pattern: m[2].trim(),
-            reason: m[3].trim(), domain: m[4].trim(), since: m[5].trim(),
-        });
-    }
-    return idioms;
-}
-
-function nextIdiomId(idioms) {
-    if (!idioms.length) return 1;
-    return Math.max(...idioms.map(i => i.id)) + 1;
-}
-
-function renderSkillMd(idioms, digestCount) {
-    const now = new Date().toISOString().slice(0, 10);
-    const rows = idioms.map(i =>
-        `| ${i.id} | \`${i.pattern}\` | ${i.reason} | ${i.domain} | ${i.since} |`
-    );
-    const table = rows.length ? rows.join('\n') : '_No idioms recorded yet._';
-
-    return `---
-name: project-idioms
-description: >
-  Auto-evolved skill containing project-specific architectural idioms.
-  Generated by skill_evolution.js — do not edit manually. Commit this
-  file to share your Engineering Culture across the team.
-version: auto
-last-updated: ${now}
-digest-cycles: ${digestCount}
-pattern: generator
----
-
-# Project Idioms — Auto-Evolved Skill
-
-> **Authority Level: ABSOLUTE**
-> These idioms were extracted from the developer's own code decisions.
-> They override generic agent defaults. Every agent MUST respect them.
-
----
-
-## How Idioms Are Born
-
-1. Developer commits code that differs from the AI proposal.
-2. \`skill_evolution.js digest\` extracts architectural deltas only.
-3. A minimal LLM reflection prompt (< 500 tokens) identifies the "WHY."
-4. The idiom is recorded here with a stable pattern + reason pair.
-
----
-
-## Recorded Idioms
-
-| ID | Pattern | Why This Project Uses It | Domain | Since |
-|:---|:--------|:-------------------------|:-------|:------|
-${table}
-
----
-
-## Enforcement Rules for All Agents
-
-\`\`\`
-□ Before proposing code: scan this skill's idiom table
-□ If your proposal contradicts an idiom → flag it explicitly
-□ Never override an idiom silently — always ask the developer first
-□ When citing an idiom: "Per Project Idiom #N: [pattern] — [reason]"
-\`\`\`
-
----
-
-## Digest History
-
-Last digest: \`${now}\`
-Total cycles: \`${digestCount}\`
-
-Run \`node .agent/scripts/skill_evolution.js status\` to see the full log.
-`;
-}
-
-function generateReflectionPrompt(delta) {
-    return `You are analyzing a code delta from a developer who changed an AI-proposed solution.
-Your only job: identify the ARCHITECTURAL IDIOM this change reveals about their project.
-
-Rules:
-- Return ONLY a YAML list of idioms. No prose. No explanation outside YAML.
-- Each idiom: pattern (code signature), reason (1 sentence WHY), domain (backend/frontend/database/general)
-- Ignore whitespace, comment, import changes — only architectural choices
-- If no meaningful idiom can be extracted, return: "idioms: []"
-- Maximum 3 idioms per delta.
-
-Delta:
-\`\`\`
-${delta.slice(0, 1500)}
-\`\`\`
-
-Output format (YAML only):
-idioms:
-  - pattern: "<code pattern or convention>"
-    reason: "<why this project uses this pattern>"
-    domain: "<backend|frontend|database|security|performance|general>"
-`;
-}
-
-function parseLlmYamlResponse(response) {
-    const idioms = [];
-    let inIdioms = false;
-    let current = {};
-
-    for (const line of response.split('\n')) {
-        const stripped = line.trim();
-        if (stripped === 'idioms:') { inIdioms = true; continue; }
-        if (!inIdioms) continue;
-        if (stripped.startsWith('- pattern:')) {
-            if (current.pattern) idioms.push(current);
-            current = { pattern: stripped.split(':', 2)[1].trim().replace(/^"|"$/g, '') };
-        } else if (stripped.startsWith('reason:') && current.pattern) {
-            current.reason = stripped.split(':', 2)[1].trim().replace(/^"|"$/g, '');
-        } else if (stripped.startsWith('domain:') && current.pattern) {
-            current.domain = stripped.split(':', 2)[1].trim().replace(/^"|"$/g, '');
-        }
-    }
-    if (current.pattern) idioms.push(current);
-    return idioms;
-}
-
-// ── Log helpers ────────────────────────────────────────────────────────────────
-function loadLog() {
-    fs.mkdirSync(HISTORY_DIR, { recursive: true });
-    if (fs.existsSync(LOG_FILE)) {
-        try { return JSON.parse(fs.readFileSync(LOG_FILE, 'utf8')); } catch { /* fallthrough */ }
-    }
-    return { cycles: [], total_tokens_saved: 0, total_idioms: 0 };
-}
-
-function saveLog(log) {
-    fs.mkdirSync(HISTORY_DIR, { recursive: true });
-    fs.writeFileSync(LOG_FILE, JSON.stringify(log, null, 2), 'utf8');
-}
-
-// ── Auto-LLM API Integration ─────────────────────────────────────────────────
-// GENERATE: Eliminates the manual copy-paste loop by auto-detecting an API key
-// and calling the LLM directly. Falls back to manual mode if no key is found.
-// Supported providers (checked in order): Anthropic → OpenAI → Gemini.
-
-function detectLlmProvider() {
-    if (process.env.ANTHROPIC_API_KEY) return { provider: 'anthropic', key: process.env.ANTHROPIC_API_KEY };
-    if (process.env.OPENAI_API_KEY)    return { provider: 'openai',    key: process.env.OPENAI_API_KEY };
-    if (process.env.GEMINI_API_KEY)    return { provider: 'gemini',    key: process.env.GEMINI_API_KEY };
-    return null;
-}
-
-/**
- * Call an LLM API with the reflection prompt. Returns the raw text response.
- * Uses only built-in Node.js `https` — zero external dependencies.
- *
- * @param {string} prompt    - The reflection prompt to send
- * @param {string} provider  - 'anthropic' | 'openai' | 'gemini'
- * @param {string} apiKey    - The API key
- * @returns {Promise<string|null>} LLM response text or null on failure
- */
-async function callLlmApi(prompt, provider, apiKey) {
-    const timeout = 30000; // 30s max — skill evolution is non-blocking
-
-    function httpsPost(hostname, path, headers, body) {
-        return new Promise((resolve, reject) => {
-            const data = JSON.stringify(body);
-            const req = https.request(
-                { method: 'POST', hostname, path, headers: { ...headers, 'Content-Length': Buffer.byteLength(data) } },
-                (res) => {
-                    let raw = '';
-                    res.on('data', c => { raw += c; });
-                    res.on('end', () => resolve(raw));
-                    res.on('error', reject);
-                }
-            );
-            req.on('error', reject);
-            req.setTimeout(timeout, () => { req.destroy(new Error('LLM API timeout')); });
-            req.write(data);
-            req.end();
-        });
-    }
-
-    try {
-        if (provider === 'anthropic') {
-            const raw = await httpsPost(
-                'api.anthropic.com',
-                '/v1/messages',
-                {
-                    'Content-Type': 'application/json',
-                    'x-api-key': apiKey,
-                    'anthropic-version': '2023-06-01',
-                },
-                {
-                    model: 'claude-3-haiku-20240307',  // Fastest/cheapest — idiom extraction
-                    max_tokens: 512,
-                    messages: [{ role: 'user', content: prompt }],
-                }
-            );
-            const json = JSON.parse(raw);
-            return json?.content?.[0]?.text ?? null;
-        }
-
-        if (provider === 'openai') {
-            const raw = await httpsPost(
-                'api.openai.com',
-                '/v1/chat/completions',
-                {
-                    'Content-Type': 'application/json',
-                    'Authorization': `Bearer ${apiKey}`,
-                },
-                {
-                    model: 'gpt-4o-mini',  // Cheapest capable model for YAML extraction
-                    max_tokens: 512,
-                    messages: [{ role: 'user', content: prompt }],
-                    temperature: 0.1,
-                }
-            );
-            const json = JSON.parse(raw);
-            return json?.choices?.[0]?.message?.content ?? null;
-        }
-
-        if (provider === 'gemini') {
-            const raw = await httpsPost(
-                'generativelanguage.googleapis.com',
-                `/v1beta/models/gemini-1.5-flash:generateContent?key=${apiKey}`,
-                { 'Content-Type': 'application/json' },
-                {
-                    contents: [{ parts: [{ text: prompt }] }],
-                    generationConfig: { maxOutputTokens: 512, temperature: 0.1 },
-                }
-            );
-            const json = JSON.parse(raw);
-            return json?.candidates?.[0]?.content?.parts?.[0]?.text ?? null;
-        }
-    } catch (e) {
-        return null; // Network/parse failure — caller falls back to manual mode
-    }
-    return null;
-}
-
-// ── Commands ──────────────────────────────────────────────────────────────────
-async function cmdDigest(args) {
-    const dryRun  = args.includes('--dry-run');
-    const diffMode = args.includes('--head') ? 'head' : 'staged';
-
-    console.log(`\n${BOLD}${CYAN}━━━ Skill Evolution — Digest Cycle ━━━━━━━━━━━━━━━━${RESET}`);
-    if (dryRun) console.log(`  ${YELLOW}DRY RUN — no files will be written${RESET}\n`);
-
-    console.log(`  ${DIM}[1/5] Fetching git diff (${diffMode})...${RESET}`);
-    const rawDiff = getGitDiff(diffMode);
-    if (!rawDiff.trim()) {
-        console.log(`  ${YELLOW}⚠ No diff found. Commit or stage changes first.${RESET}`);
-        console.log(`  ${DIM}Tip: Use --head to diff against the last commit.${RESET}\n`);
-        return;
-    }
-
-    const rawTokens = countTokensEstimate(rawDiff);
-    console.log(`  ${DIM}   Raw diff: ~${rawTokens} tokens (${rawDiff.length} chars)${RESET}`);
-
-    console.log(`  ${DIM}[2/5] Extracting architectural delta (Semantic Filter)...${RESET}`);
-    const delta = semanticDelta(rawDiff, 2);
-    if (!delta.trim()) {
-        console.log(`  ${GREEN}✔ Delta is 100% trivial (whitespace/comments/imports only).${RESET}`);
-        console.log(`  ${DIM}  No LLM call needed. Zero tokens consumed.${RESET}\n`);
-        return;
-    }
-
-    const deltaTokens = countTokensEstimate(delta);
-    const savedTokens = rawTokens - deltaTokens;
-    const savedPct = Math.floor((savedTokens / Math.max(rawTokens, 1)) * 100);
-    console.log(`  ${GREEN}✔ Filtered to ~${deltaTokens} tokens  (${savedPct}% reduction, saved ~${savedTokens} tokens)${RESET}`);
-
-    console.log(`\n  ${BOLD}Architectural Delta Preview:${RESET}`);
-    const previewLines = delta.split('\n').slice(0, 20);
-    for (const line of previewLines) {
-        if (line.startsWith('+'))       console.log(`    ${GREEN}${line}${RESET}`);
-        else if (line.startsWith('-'))  console.log(`    ${RED}${line}${RESET}`);
-        else if (line.startsWith('@@')) console.log(`    ${BLUE}${line}${RESET}`);
-        else                            console.log(`    ${DIM}${line}${RESET}`);
-    }
-    if (delta.split('\n').length > 20) console.log(`    ${DIM}... (${delta.split('\n').length - 20} more lines)${RESET}`);
-
-    if (dryRun) {
-        console.log(`\n  ${YELLOW}[DRY RUN] Would send ${deltaTokens} tokens to LLM for reflection.${RESET}`);
-        console.log(`  ${DIM}Run without --dry-run to complete the digest.${RESET}\n`);
-        return;
-    }
-
-    // GENERATE: Auto-LLM call. Tries API first, falls back to manual paste if no key.
-    const reflectionPrompt = generateReflectionPrompt(delta);
-    let llmResponse = '';
-
-    let llmCreds = detectLlmProvider();
-    if (llmCreds) {
-        console.log(`  ${DIM}[3/5] LLM Reflection — auto-calling ${llmCreds.provider} API...${RESET}`);
-        const autoResponse = await callLlmApi(reflectionPrompt, llmCreds.provider, llmCreds.key);
-        if (autoResponse) {
-            llmResponse = autoResponse;
-            console.log(`  ${GREEN}✔ Auto-response received (${llmCreds.provider}) — ${llmResponse.split('\n').length} lines${RESET}`);
-        } else {
-            console.log(`  ${YELLOW}⚠ API call failed — falling back to manual mode${RESET}`);
-            llmCreds = null; // triggers manual fallback below
-        }
-    }
-
-    if (!llmCreds || !llmResponse) {
-        // Manual fallback: copy-paste mode (no API key configured)
-        console.log(`\n  ${DIM}[3/5] LLM Reflection — copy the prompt below and paste the response${RESET}`);
-        console.log(`  ${DIM}  Tip: Set ANTHROPIC_API_KEY, OPENAI_API_KEY, or GEMINI_API_KEY to automate this step.${RESET}`);
-        console.log(`\n  ${BOLD}${'─'.repeat(60)}${RESET}`);
-        console.log(reflectionPrompt);
-        console.log(`  ${BOLD}${'─'.repeat(60)}${RESET}`);
-        console.log(`\n  ${BOLD}Paste LLM response below (type END_RESPONSE when done):${RESET}`);
-
-        const rl = readline.createInterface({ input: process.stdin, output: process.stdout });
-        const responseLines = [];
-        await new Promise(resolve => {
-            const listener = (line) => {
-                if (line.trim() === 'END_RESPONSE') { rl.removeListener('line', listener); resolve(); }
-                else responseLines.push(line);
-            };
-            rl.on('line', listener);
-        });
-        rl.close();
-        llmResponse = responseLines.join('\n');
-    }
-
-    console.log(`\n  ${DIM}[4/5] Parsing idioms...${RESET}`);
-    const newIdioms = parseLlmYamlResponse(llmResponse);
-    if (!newIdioms.length) {
-        console.log(`  ${YELLOW}⚠ No idioms extracted from LLM response.${RESET}`);
-        console.log(`  ${DIM}  The LLM may have returned idioms: [] — no architectural pattern detected.${RESET}\n`);
-        return;
-    }
-
-    console.log(`  ${GREEN}✔ Extracted ${newIdioms.length} idiom(s)${RESET}`);
-    for (const idiom of newIdioms) {
-        console.log(`    ${CYAN}• ${idiom.pattern || '?'}${RESET}  — ${idiom.reason || ''}`);
-    }
-
-    console.log(`\n  ${DIM}[5/5] Merging into project-idioms/SKILL.md...${RESET}`);
-    const existing = loadExistingIdioms();
-    const log = loadLog();
-    let nextId = nextIdiomId(existing);
-    const today = new Date().toISOString().slice(0, 10);
-    const merged = [...existing];
-    let added = 0;
-
-    for (const idiom of newIdioms) {
-        // FIX: Use Levenshtein normalised similarity (threshold 0.80) instead of
-        // substring .includes() which was over-aggressive and blocked valid idioms.
-        if (isDuplicateIdiom(idiom.pattern || '', existing)) {
-            console.log(`  ${DIM}  Skipped near-duplicate: ${idiom.pattern}${RESET}`);
-            continue;
-        }
-        merged.push({
-            id: nextId, pattern: idiom.pattern || '?',
-            reason: idiom.reason || 'No reason provided.',
-            domain: idiom.domain || 'general', since: today,
-        });
-        nextId++;
-        added++;
-    }
-
-    if (added === 0) {
-        console.log(`  ${YELLOW}⚠ All extracted idioms were duplicates. SKILL.md unchanged.${RESET}\n`);
-        return;
-    }
-
-    log.total_idioms = merged.length;
-    const skillMd = renderSkillMd(merged, (log.cycles || []).length + 1);
-    fs.mkdirSync(SKILL_DIR, { recursive: true });
-    fs.writeFileSync(SKILL_FILE, skillMd, 'utf8');
-
-    log.cycles = log.cycles || [];
-    log.cycles.push({
-        timestamp: new Date().toISOString().slice(0, 19),
-        raw_tokens: rawTokens, delta_tokens: deltaTokens,
-        tokens_saved: savedTokens, idioms_added: added,
-    });
-    log.total_tokens_saved = (log.total_tokens_saved || 0) + savedTokens;
-    saveLog(log);
-
-    console.log(`\n  ${GREEN}✔ ${added} new idiom(s) added to SKILL.md${RESET}`);
-    console.log(`  ${DIM}   File: ${SKILL_FILE}${RESET}`);
-    console.log(`  ${DIM}   Total idioms: ${merged.length}${RESET}`);
-    console.log(`  ${DIM}   Lifetime tokens saved: ${log.total_tokens_saved}${RESET}\n`);
-    console.log(`  ${CYAN}Commit SKILL.md to share your Engineering Culture with the team.${RESET}\n`);
-}
-
-function cmdShow() {
-    if (!fs.existsSync(SKILL_FILE)) { console.log(`${YELLOW}No project-idioms skill found. Run 'digest' first.${RESET}`); return; }
-    console.log(fs.readFileSync(SKILL_FILE, 'utf8'));
-}
-
-function cmdReset() {
-    if (fs.existsSync(SKILL_FILE)) { fs.unlinkSync(SKILL_FILE); console.log(`${GREEN}✔ project-idioms/SKILL.md deleted.${RESET}`); }
-    if (fs.existsSync(LOG_FILE))   { fs.unlinkSync(LOG_FILE);   console.log(`${GREEN}✔ Digest log cleared.${RESET}`); }
-    console.log(`${DIM}Run 'digest' to start a fresh evolution cycle.${RESET}`);
-}
-
-function cmdStatus() {
-    const log = loadLog();
-    const cycles = log.cycles || [];
-    const totalSaved = log.total_tokens_saved || 0;
-    const totalIdioms = log.total_idioms || 0;
-    const idiomsExist = fs.existsSync(SKILL_FILE);
-
-    console.log(`\n${BOLD}${CYAN}━━━ Skill Evolution Status ━━━━━━━━━━━━━━━━━━━━━━━━${RESET}`);
-    console.log(`  Digest cycles    : ${BOLD}${cycles.length}${RESET}`);
-    console.log(`  Total idioms     : ${BOLD}${totalIdioms}${RESET}`);
-    console.log(`  Tokens saved     : ${GREEN}${totalSaved.toLocaleString()} tokens${RESET}  (≈ $${(totalSaved / 1_000_000 * 3).toFixed(4)} at $3/M)`);
-    console.log(`  SKILL.md exists  : ${idiomsExist ? '✔' : '✗'}`);
-
-    if (cycles.length) {
-        console.log(`\n  ${BOLD}Last 5 digest cycles:${RESET}`);
-        for (const cycle of cycles.slice(-5).reverse()) {
-            const ts = (cycle.timestamp || '?').slice(0, 16);
-            const deltaT = cycle.delta_tokens || 0;
-            const saved = cycle.tokens_saved || 0;
-            const addedCount = cycle.idioms_added || 0;
-            const pct = Math.floor((saved / Math.max(cycle.raw_tokens || 1, 1)) * 100);
-            console.log(`    ${DIM}${ts}${RESET}  delta=${deltaT}tok  saved=${saved}tok (${pct}%)  idioms+=${addedCount}`);
-        }
-    }
-    console.log(`${CYAN}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${RESET}\n`);
-}
-
-// ── Main ──────────────────────────────────────────────────────────────────────
-const COMMANDS = { digest: cmdDigest, show: cmdShow, reset: cmdReset, status: cmdStatus };
-
-async function main() {
-    const argv = process.argv.slice(2);
-    if (!argv.length || ['-h', '--help', 'help'].includes(argv[0])) {
-        console.log(`
-${BOLD}skill_evolution.js${RESET} — Tribunal Skill Evolution Forge
-
-${BOLD}Commands:${RESET}
-  digest [--dry-run] [--head]   Analyze latest git diff and evolve SKILL.md
-                                  --dry-run  : preview without writing
-                                  --head     : diff last commit instead of staged
-  show                           Print current project-idioms/SKILL.md
-  status                         Show digest history and token savings
-  reset                          Clear all idioms and start fresh
-
-${BOLD}Token Budget:${RESET}
-  Raw diff -> Semantic Filter -> Only architectural lines -> LLM
-  Typical savings: 70–90% of tokens. Most trivial commits = 0 tokens.
-`);
-        return;
-    }
-
-    const cmd = argv[0];
-    const rest = argv.slice(1);
-    if (!COMMANDS[cmd]) { console.log(`${RED}✖ Unknown command: '${cmd}'${RESET}`); process.exit(1); }
-    await COMMANDS[cmd](rest);
-}
-
-module.exports = { semanticDelta, architecturalWeight, parseLlmYamlResponse, loadExistingIdioms };
-
-if (require.main === module) {
-    main().catch(err => { console.error(err); process.exit(1); });
-}
+#!/usr/bin/env node
+/**
+ * skill_evolution.js — Tribunal Kit Skill Evolution Forge
+ * =========================================================
+ * Analyzes the delta between what the AI proposed and what the developer
+ * actually committed, then distills those decisions into evolving
+ * project-specific SKILL idioms — WITHOUT sending full files to any LLM.
+ *
+ * Core Strategy: Semantic Delta Extraction
+ *   1. Read the raw git diff of staged/recent changes
+ *   2. Strip trivial noise (whitespace, comments, import renames)
+ *   3. Score remaining lines for "Architectural Weight"
+ *   4. Only high-weight deltas reach the LLM reflection prompt
+ *   5. LLM returns structured YAML idiom entries (not prose)
+ *   6. Idioms are merged into .agent/skills/project-idioms/SKILL.md
+ *
+ * Usage:
+ *   node .agent/scripts/skill_evolution.js digest
+ *   node .agent/scripts/skill_evolution.js digest --dry-run
+ *   node .agent/scripts/skill_evolution.js show
+ *   node .agent/scripts/skill_evolution.js reset
+ *   node .agent/scripts/skill_evolution.js status
+ */
+
+'use strict';
+
+const fs      = require('fs');
+const path    = require('path');
+const https   = require('https');
+const { execSync } = require('child_process');
+const readline = require('readline');
+
+// ── Colours ──────────────────────────────────────────────────────────────────
+const { GREEN, YELLOW, CYAN, RED, BLUE, BOLD, DIM, RESET } = require('./_colors');
+
+// ── Find .agent directory ─────────────────────────────────────────────────────
+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(`${RED}✖ Error: '.agent' directory not found. Please run 'npx tribunal-kit init' first.${RESET}`);
+    process.exit(1);
+}
+
+const AGENT_DIR   = findAgentDir();
+const SKILL_DIR   = path.join(AGENT_DIR, 'skills', 'project-idioms');
+const SKILL_FILE  = path.join(SKILL_DIR, 'SKILL.md');
+const HISTORY_DIR = path.join(AGENT_DIR, 'history', 'skill-evolution');
+const LOG_FILE    = path.join(HISTORY_DIR, 'digest-log.json');
+
+// ── Architectural Weight Patterns ────────────────────────────────────────────
+const HIGH_WEIGHT_PATTERNS = [
+    /\bclass\b/, /\binterface\b/, /\btype\s+\w+\s*=/, /\bextends\b/, /\bimplements\b/,
+    /\bthrow\b/, /\bcatch\b/, /\btry\b/,
+    /\bprisma\.\w+\(/, /\bsupabase\./, /\bfetch\(/, /\baxios\./,
+    /\bReturnType\b/, /\bPromise</, /\basync\s+function/, /\bawait\b/,
+    /\bexport\s+(default\s+)?(class|function|const)/, /\bmodule\.exports\b/,
+    /\bRouter\b|\bapp\.(get|post|put|delete|patch)\(/,
+    /\buse[A-Z]\w+\(/, /\bcreateContext\(/,
+    /\bz\.object\(/, /\bPrisma\b|\bdrizzle\b/,
+    /\benv\.\w+/, /\bprocess\.env\./,
+];
+
+const NOISE_PATTERNS = [
+    /^\s*$/, /^\s*(\/\/|#|\/\*).*$/, /^\s*\*/,
+    /^\s*import\s+\{[^}]+\}\s+from\s+['"](?!\.)/, /^\s*(console\.(log|warn|error)|print\()/,
+    /^\s*\w+\s*[:,]?\s*$/,
+];
+
+function architecturalWeight(line) {
+    const code = line.replace(/^[+-]/, '').trim();
+    for (const p of NOISE_PATTERNS) { if (p.test(code)) return 0; }
+    for (const p of HIGH_WEIGHT_PATTERNS) { if (p.test(code)) return 2; }
+    return 1;
+}
+
+// ── Levenshtein Semantic Deduplication ───────────────────────────────────────
+// FIX: Replaces over-aggressive substring matching (.includes) with normalised
+// edit-distance similarity. A new idiom must differ by >= 20% from all existing
+// ones to be accepted. Threshold 0.80 = 80% similar → considered a duplicate.
+
+function levenshtein(a, b) {
+    const m = a.length, n = b.length;
+    // Allocate DP table with base cases pre-filled
+    const dp = Array.from({ length: m + 1 }, (_, i) =>
+        Array.from({ length: n + 1 }, (_, j) => i === 0 ? j : j === 0 ? i : 0)
+    );
+    for (let i = 1; i <= m; i++) {
+        for (let j = 1; j <= n; j++) {
+            dp[i][j] = a[i - 1] === b[j - 1]
+                ? dp[i - 1][j - 1]
+                : 1 + Math.min(dp[i - 1][j], dp[i][j - 1], dp[i - 1][j - 1]);
+        }
+    }
+    return dp[m][n];
+}
+
+function normalizedSimilarity(a, b) {
+    if (!a.length && !b.length) return 1.0;
+    const maxLen = Math.max(a.length, b.length);
+    if (maxLen === 0) return 1.0;
+    return 1 - levenshtein(a, b) / maxLen;
+}
+
+/**
+ * Returns true if newPattern is semantically similar to any existing idiom.
+ * Uses Levenshtein normalised similarity with a configurable threshold.
+ * @param {string}   newPattern
+ * @param {object[]} existingIdioms - array of { pattern } objects
+ * @param {number}   threshold      - 0.0–1.0 (default 0.80 = 80% similar = duplicate)
+ */
+function isDuplicateIdiom(newPattern, existingIdioms, threshold = 0.80) {
+    const newLow = newPattern.toLowerCase();
+    return existingIdioms.some(ex => {
+        const exLow = (ex.pattern || '').toLowerCase();
+        return normalizedSimilarity(newLow, exLow) >= threshold;
+    });
+}
+
+function semanticDelta(diffText, minWeight = 2) {
+    const lines = diffText.split('\n');
+    const kept = [];
+    let currentHunkHasHigh = false;
+    let hunkLines = [];
+
+    for (const line of lines) {
+        if (line.startsWith('---') || line.startsWith('+++') || line.startsWith('diff --git')) {
+            kept.push(line); continue;
+        }
+        if (line.startsWith('@@')) {
+            if (currentHunkHasHigh) kept.push(...hunkLines);
+            currentHunkHasHigh = false;
+            hunkLines = [line]; continue;
+        }
+        if (line.startsWith('+') || line.startsWith('-')) {
+            const w = architecturalWeight(line);
+            hunkLines.push(line);
+            if (w >= minWeight) currentHunkHasHigh = true;
+        } else {
+            hunkLines.push(line);
+        }
+    }
+    if (currentHunkHasHigh) kept.push(...hunkLines);
+
+    let result = kept.join('\n');
+    result = result.replace(/\n( ?\n){3,}/g, '\n\n');
+    return result.trim();
+}
+
+// ── Git helpers ────────────────────────────────────────────────────────────────
+function getGitDiff(mode = 'staged') {
+    try {
+        let cmd;
+        if (mode === 'staged')    cmd = 'git diff --cached --unified=3';
+        else if (mode === 'head') cmd = 'git diff HEAD~1 HEAD --unified=3';
+        else                      cmd = 'git diff --unified=3';
+        return execSync(cmd, { encoding: 'utf8', timeout: 10000, stdio: ['pipe', 'pipe', 'pipe'] });
+    } catch { return ''; }
+}
+
+function countTokensEstimate(text) {
+    return Math.max(1, Math.floor(text.length / 4));
+}
+
+// ── Idiom management ──────────────────────────────────────────────────────────
+function loadExistingIdioms() {
+    if (!fs.existsSync(SKILL_FILE)) return [];
+    const content = fs.readFileSync(SKILL_FILE, 'utf8');
+    const idioms = [];
+    const pattern = /\|\s*(\d+)\s*\|\s*`([^`]+)`\s*\|\s*([^|]+)\|\s*([^|]+)\|\s*([^|]+)\|/g;
+    let m;
+    while ((m = pattern.exec(content)) !== null) {
+        idioms.push({
+            id: parseInt(m[1], 10), pattern: m[2].trim(),
+            reason: m[3].trim(), domain: m[4].trim(), since: m[5].trim(),
+        });
+    }
+    return idioms;
+}
+
+function nextIdiomId(idioms) {
+    if (!idioms.length) return 1;
+    return Math.max(...idioms.map(i => i.id)) + 1;
+}
+
+function renderSkillMd(idioms, digestCount) {
+    const now = new Date().toISOString().slice(0, 10);
+    const rows = idioms.map(i =>
+        `| ${i.id} | \`${i.pattern}\` | ${i.reason} | ${i.domain} | ${i.since} |`
+    );
+    const table = rows.length ? rows.join('\n') : '_No idioms recorded yet._';
+
+    return `---
+name: project-idioms
+description: >
+  Auto-evolved skill containing project-specific architectural idioms.
+  Generated by skill_evolution.js — do not edit manually. Commit this
+  file to share your Engineering Culture across the team.
+version: auto
+last-updated: ${now}
+digest-cycles: ${digestCount}
+pattern: generator
+---
+
+# Project Idioms — Auto-Evolved Skill
+
+> **Authority Level: ABSOLUTE**
+> These idioms were extracted from the developer's own code decisions.
+> They override generic agent defaults. Every agent MUST respect them.
+
+---
+
+## How Idioms Are Born
+
+1. Developer commits code that differs from the AI proposal.
+2. \`skill_evolution.js digest\` extracts architectural deltas only.
+3. A minimal LLM reflection prompt (< 500 tokens) identifies the "WHY."
+4. The idiom is recorded here with a stable pattern + reason pair.
+
+---
+
+## Recorded Idioms
+
+| ID | Pattern | Why This Project Uses It | Domain | Since |
+|:---|:--------|:-------------------------|:-------|:------|
+${table}
+
+---
+
+## Enforcement Rules for All Agents
+
+\`\`\`
+□ Before proposing code: scan this skill's idiom table
+□ If your proposal contradicts an idiom → flag it explicitly
+□ Never override an idiom silently — always ask the developer first
+□ When citing an idiom: "Per Project Idiom #N: [pattern] — [reason]"
+\`\`\`
+
+---
+
+## Digest History
+
+Last digest: \`${now}\`
+Total cycles: \`${digestCount}\`
+
+Run \`node .agent/scripts/skill_evolution.js status\` to see the full log.
+`;
+}
+
+function generateReflectionPrompt(delta) {
+    return `You are analyzing a code delta from a developer who changed an AI-proposed solution.
+Your only job: identify the ARCHITECTURAL IDIOM this change reveals about their project.
+
+Rules:
+- Return ONLY a YAML list of idioms. No prose. No explanation outside YAML.
+- Each idiom: pattern (code signature), reason (1 sentence WHY), domain (backend/frontend/database/general)
+- Ignore whitespace, comment, import changes — only architectural choices
+- If no meaningful idiom can be extracted, return: "idioms: []"
+- Maximum 3 idioms per delta.
+
+Delta:
+\`\`\`
+${delta.slice(0, 1500)}
+\`\`\`
+
+Output format (YAML only):
+idioms:
+  - pattern: "<code pattern or convention>"
+    reason: "<why this project uses this pattern>"
+    domain: "<backend|frontend|database|security|performance|general>"
+`;
+}
+
+function parseLlmYamlResponse(response) {
+    const idioms = [];
+    let inIdioms = false;
+    let current = {};
+
+    for (const line of response.split('\n')) {
+        const stripped = line.trim();
+        if (stripped === 'idioms:') { inIdioms = true; continue; }
+        if (!inIdioms) continue;
+        if (stripped.startsWith('- pattern:')) {
+            if (current.pattern) idioms.push(current);
+            current = { pattern: stripped.split(':', 2)[1].trim().replace(/^"|"$/g, '') };
+        } else if (stripped.startsWith('reason:') && current.pattern) {
+            current.reason = stripped.split(':', 2)[1].trim().replace(/^"|"$/g, '');
+        } else if (stripped.startsWith('domain:') && current.pattern) {
+            current.domain = stripped.split(':', 2)[1].trim().replace(/^"|"$/g, '');
+        }
+    }
+    if (current.pattern) idioms.push(current);
+    return idioms;
+}
+
+// ── Log helpers ────────────────────────────────────────────────────────────────
+function loadLog() {
+    fs.mkdirSync(HISTORY_DIR, { recursive: true });
+    if (fs.existsSync(LOG_FILE)) {
+        try { return JSON.parse(fs.readFileSync(LOG_FILE, 'utf8')); } catch { /* fallthrough */ }
+    }
+    return { cycles: [], total_tokens_saved: 0, total_idioms: 0 };
+}
+
+function saveLog(log) {
+    fs.mkdirSync(HISTORY_DIR, { recursive: true });
+    fs.writeFileSync(LOG_FILE, JSON.stringify(log, null, 2), 'utf8');
+}
+
+// ── Auto-LLM API Integration ─────────────────────────────────────────────────
+// GENERATE: Eliminates the manual copy-paste loop by auto-detecting an API key
+// and calling the LLM directly. Falls back to manual mode if no key is found.
+// Supported providers (checked in order): Anthropic → OpenAI → Gemini.
+
+function detectLlmProvider() {
+    if (process.env.ANTHROPIC_API_KEY) return { provider: 'anthropic', key: process.env.ANTHROPIC_API_KEY };
+    if (process.env.OPENAI_API_KEY)    return { provider: 'openai',    key: process.env.OPENAI_API_KEY };
+    if (process.env.GEMINI_API_KEY)    return { provider: 'gemini',    key: process.env.GEMINI_API_KEY };
+    return null;
+}
+
+/**
+ * Call an LLM API with the reflection prompt. Returns the raw text response.
+ * Uses only built-in Node.js `https` — zero external dependencies.
+ *
+ * @param {string} prompt    - The reflection prompt to send
+ * @param {string} provider  - 'anthropic' | 'openai' | 'gemini'
+ * @param {string} apiKey    - The API key
+ * @returns {Promise<string|null>} LLM response text or null on failure
+ */
+async function callLlmApi(prompt, provider, apiKey) {
+    const timeout = 30000; // 30s max — skill evolution is non-blocking
+
+    function httpsPost(hostname, path, headers, body) {
+        return new Promise((resolve, reject) => {
+            const data = JSON.stringify(body);
+            const req = https.request(
+                { method: 'POST', hostname, path, headers: { ...headers, 'Content-Length': Buffer.byteLength(data) } },
+                (res) => {
+                    let raw = '';
+                    res.on('data', c => { raw += c; });
+                    res.on('end', () => resolve(raw));
+                    res.on('error', reject);
+                }
+            );
+            req.on('error', reject);
+            req.setTimeout(timeout, () => { req.destroy(new Error('LLM API timeout')); });
+            req.write(data);
+            req.end();
+        });
+    }
+
+    try {
+        if (provider === 'anthropic') {
+            const raw = await httpsPost(
+                'api.anthropic.com',
+                '/v1/messages',
+                {
+                    'Content-Type': 'application/json',
+                    'x-api-key': apiKey,
+                    'anthropic-version': '2023-06-01',
+                },
+                {
+                    model: 'claude-3-haiku-20240307',  // Fastest/cheapest — idiom extraction
+                    max_tokens: 512,
+                    messages: [{ role: 'user', content: prompt }],
+                }
+            );
+            const json = JSON.parse(raw);
+            return json?.content?.[0]?.text ?? null;
+        }
+
+        if (provider === 'openai') {
+            const raw = await httpsPost(
+                'api.openai.com',
+                '/v1/chat/completions',
+                {
+                    'Content-Type': 'application/json',
+                    'Authorization': `Bearer ${apiKey}`,
+                },
+                {
+                    model: 'gpt-4o-mini',  // Cheapest capable model for YAML extraction
+                    max_tokens: 512,
+                    messages: [{ role: 'user', content: prompt }],
+                    temperature: 0.1,
+                }
+            );
+            const json = JSON.parse(raw);
+            return json?.choices?.[0]?.message?.content ?? null;
+        }
+
+        if (provider === 'gemini') {
+            const raw = await httpsPost(
+                'generativelanguage.googleapis.com',
+                `/v1beta/models/gemini-1.5-flash:generateContent?key=${apiKey}`,
+                { 'Content-Type': 'application/json' },
+                {
+                    contents: [{ parts: [{ text: prompt }] }],
+                    generationConfig: { maxOutputTokens: 512, temperature: 0.1 },
+                }
+            );
+            const json = JSON.parse(raw);
+            return json?.candidates?.[0]?.content?.parts?.[0]?.text ?? null;
+        }
+    } catch {
+        return null; // Network/parse failure — caller falls back to manual mode
+    }
+    return null;
+}
+
+// ── Commands ──────────────────────────────────────────────────────────────────
+async function cmdDigest(args) {
+    const dryRun  = args.includes('--dry-run');
+    const diffMode = args.includes('--head') ? 'head' : 'staged';
+
+    console.log(`\n${BOLD}${CYAN}━━━ Skill Evolution — Digest Cycle ━━━━━━━━━━━━━━━━${RESET}`);
+    if (dryRun) console.log(`  ${YELLOW}DRY RUN — no files will be written${RESET}\n`);
+
+    console.log(`  ${DIM}[1/5] Fetching git diff (${diffMode})...${RESET}`);
+    const rawDiff = getGitDiff(diffMode);
+    if (!rawDiff.trim()) {
+        console.log(`  ${YELLOW}⚠ No diff found. Commit or stage changes first.${RESET}`);
+        console.log(`  ${DIM}Tip: Use --head to diff against the last commit.${RESET}\n`);
+        return;
+    }
+
+    const rawTokens = countTokensEstimate(rawDiff);
+    console.log(`  ${DIM}   Raw diff: ~${rawTokens} tokens (${rawDiff.length} chars)${RESET}`);
+
+    console.log(`  ${DIM}[2/5] Extracting architectural delta (Semantic Filter)...${RESET}`);
+    const delta = semanticDelta(rawDiff, 2);
+    if (!delta.trim()) {
+        console.log(`  ${GREEN}✔ Delta is 100% trivial (whitespace/comments/imports only).${RESET}`);
+        console.log(`  ${DIM}  No LLM call needed. Zero tokens consumed.${RESET}\n`);
+        return;
+    }
+
+    const deltaTokens = countTokensEstimate(delta);
+    const savedTokens = rawTokens - deltaTokens;
+    const savedPct = Math.floor((savedTokens / Math.max(rawTokens, 1)) * 100);
+    console.log(`  ${GREEN}✔ Filtered to ~${deltaTokens} tokens  (${savedPct}% reduction, saved ~${savedTokens} tokens)${RESET}`);
+
+    console.log(`\n  ${BOLD}Architectural Delta Preview:${RESET}`);
+    const previewLines = delta.split('\n').slice(0, 20);
+    for (const line of previewLines) {
+        if (line.startsWith('+'))       console.log(`    ${GREEN}${line}${RESET}`);
+        else if (line.startsWith('-'))  console.log(`    ${RED}${line}${RESET}`);
+        else if (line.startsWith('@@')) console.log(`    ${BLUE}${line}${RESET}`);
+        else                            console.log(`    ${DIM}${line}${RESET}`);
+    }
+    if (delta.split('\n').length > 20) console.log(`    ${DIM}... (${delta.split('\n').length - 20} more lines)${RESET}`);
+
+    if (dryRun) {
+        console.log(`\n  ${YELLOW}[DRY RUN] Would send ${deltaTokens} tokens to LLM for reflection.${RESET}`);
+        console.log(`  ${DIM}Run without --dry-run to complete the digest.${RESET}\n`);
+        return;
+    }
+
+    // GENERATE: Auto-LLM call. Tries API first, falls back to manual paste if no key.
+    const reflectionPrompt = generateReflectionPrompt(delta);
+    let llmResponse = '';
+
+    let llmCreds = detectLlmProvider();
+    if (llmCreds) {
+        console.log(`  ${DIM}[3/5] LLM Reflection — auto-calling ${llmCreds.provider} API...${RESET}`);
+        const autoResponse = await callLlmApi(reflectionPrompt, llmCreds.provider, llmCreds.key);
+        if (autoResponse) {
+            llmResponse = autoResponse;
+            console.log(`  ${GREEN}✔ Auto-response received (${llmCreds.provider}) — ${llmResponse.split('\n').length} lines${RESET}`);
+        } else {
+            console.log(`  ${YELLOW}⚠ API call failed — falling back to manual mode${RESET}`);
+            llmCreds = null; // triggers manual fallback below
+        }
+    }
+
+    if (!llmCreds || !llmResponse) {
+        // Manual fallback: copy-paste mode (no API key configured)
+        console.log(`\n  ${DIM}[3/5] LLM Reflection — copy the prompt below and paste the response${RESET}`);
+        console.log(`  ${DIM}  Tip: Set ANTHROPIC_API_KEY, OPENAI_API_KEY, or GEMINI_API_KEY to automate this step.${RESET}`);
+        console.log(`\n  ${BOLD}${'─'.repeat(60)}${RESET}`);
+        console.log(reflectionPrompt);
+        console.log(`  ${BOLD}${'─'.repeat(60)}${RESET}`);
+        console.log(`\n  ${BOLD}Paste LLM response below (type END_RESPONSE when done):${RESET}`);
+
+        const rl = readline.createInterface({ input: process.stdin, output: process.stdout });
+        const responseLines = [];
+        await new Promise(resolve => {
+            const listener = (line) => {
+                if (line.trim() === 'END_RESPONSE') { rl.removeListener('line', listener); resolve(); }
+                else responseLines.push(line);
+            };
+            rl.on('line', listener);
+        });
+        rl.close();
+        llmResponse = responseLines.join('\n');
+    }
+
+    console.log(`\n  ${DIM}[4/5] Parsing idioms...${RESET}`);
+    const newIdioms = parseLlmYamlResponse(llmResponse);
+    if (!newIdioms.length) {
+        console.log(`  ${YELLOW}⚠ No idioms extracted from LLM response.${RESET}`);
+        console.log(`  ${DIM}  The LLM may have returned idioms: [] — no architectural pattern detected.${RESET}\n`);
+        return;
+    }
+
+    console.log(`  ${GREEN}✔ Extracted ${newIdioms.length} idiom(s)${RESET}`);
+    for (const idiom of newIdioms) {
+        console.log(`    ${CYAN}• ${idiom.pattern || '?'}${RESET}  — ${idiom.reason || ''}`);
+    }
+
+    console.log(`\n  ${DIM}[5/5] Merging into project-idioms/SKILL.md...${RESET}`);
+    const existing = loadExistingIdioms();
+    const log = loadLog();
+    let nextId = nextIdiomId(existing);
+    const today = new Date().toISOString().slice(0, 10);
+    const merged = [...existing];
+    let added = 0;
+
+    for (const idiom of newIdioms) {
+        // FIX: Use Levenshtein normalised similarity (threshold 0.80) instead of
+        // substring .includes() which was over-aggressive and blocked valid idioms.
+        if (isDuplicateIdiom(idiom.pattern || '', existing)) {
+            console.log(`  ${DIM}  Skipped near-duplicate: ${idiom.pattern}${RESET}`);
+            continue;
+        }
+        merged.push({
+            id: nextId, pattern: idiom.pattern || '?',
+            reason: idiom.reason || 'No reason provided.',
+            domain: idiom.domain || 'general', since: today,
+        });
+        nextId++;
+        added++;
+    }
+
+    if (added === 0) {
+        console.log(`  ${YELLOW}⚠ All extracted idioms were duplicates. SKILL.md unchanged.${RESET}\n`);
+        return;
+    }
+
+    log.total_idioms = merged.length;
+    const skillMd = renderSkillMd(merged, (log.cycles || []).length + 1);
+    fs.mkdirSync(SKILL_DIR, { recursive: true });
+    fs.writeFileSync(SKILL_FILE, skillMd, 'utf8');
+
+    log.cycles = log.cycles || [];
+    log.cycles.push({
+        timestamp: new Date().toISOString().slice(0, 19),
+        raw_tokens: rawTokens, delta_tokens: deltaTokens,
+        tokens_saved: savedTokens, idioms_added: added,
+    });
+    log.total_tokens_saved = (log.total_tokens_saved || 0) + savedTokens;
+    saveLog(log);
+
+    console.log(`\n  ${GREEN}✔ ${added} new idiom(s) added to SKILL.md${RESET}`);
+    console.log(`  ${DIM}   File: ${SKILL_FILE}${RESET}`);
+    console.log(`  ${DIM}   Total idioms: ${merged.length}${RESET}`);
+    console.log(`  ${DIM}   Lifetime tokens saved: ${log.total_tokens_saved}${RESET}\n`);
+    console.log(`  ${CYAN}Commit SKILL.md to share your Engineering Culture with the team.${RESET}\n`);
+}
+
+function cmdShow() {
+    if (!fs.existsSync(SKILL_FILE)) { console.log(`${YELLOW}No project-idioms skill found. Run 'digest' first.${RESET}`); return; }
+    console.log(fs.readFileSync(SKILL_FILE, 'utf8'));
+}
+
+function cmdReset() {
+    if (fs.existsSync(SKILL_FILE)) { fs.unlinkSync(SKILL_FILE); console.log(`${GREEN}✔ project-idioms/SKILL.md deleted.${RESET}`); }
+    if (fs.existsSync(LOG_FILE))   { fs.unlinkSync(LOG_FILE);   console.log(`${GREEN}✔ Digest log cleared.${RESET}`); }
+    console.log(`${DIM}Run 'digest' to start a fresh evolution cycle.${RESET}`);
+}
+
+function cmdStatus() {
+    const log = loadLog();
+    const cycles = log.cycles || [];
+    const totalSaved = log.total_tokens_saved || 0;
+    const totalIdioms = log.total_idioms || 0;
+    const idiomsExist = fs.existsSync(SKILL_FILE);
+
+    console.log(`\n${BOLD}${CYAN}━━━ Skill Evolution Status ━━━━━━━━━━━━━━━━━━━━━━━━${RESET}`);
+    console.log(`  Digest cycles    : ${BOLD}${cycles.length}${RESET}`);
+    console.log(`  Total idioms     : ${BOLD}${totalIdioms}${RESET}`);
+    console.log(`  Tokens saved     : ${GREEN}${totalSaved.toLocaleString()} tokens${RESET}  (≈ $${(totalSaved / 1_000_000 * 3).toFixed(4)} at $3/M)`);
+    console.log(`  SKILL.md exists  : ${idiomsExist ? '✔' : '✗'}`);
+
+    if (cycles.length) {
+        console.log(`\n  ${BOLD}Last 5 digest cycles:${RESET}`);
+        for (const cycle of cycles.slice(-5).reverse()) {
+            const ts = (cycle.timestamp || '?').slice(0, 16);
+            const deltaT = cycle.delta_tokens || 0;
+            const saved = cycle.tokens_saved || 0;
+            const addedCount = cycle.idioms_added || 0;
+            const pct = Math.floor((saved / Math.max(cycle.raw_tokens || 1, 1)) * 100);
+            console.log(`    ${DIM}${ts}${RESET}  delta=${deltaT}tok  saved=${saved}tok (${pct}%)  idioms+=${addedCount}`);
+        }
+    }
+    console.log(`${CYAN}━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━${RESET}\n`);
+}
+
+// ── Main ──────────────────────────────────────────────────────────────────────
+const COMMANDS = { digest: cmdDigest, show: cmdShow, reset: cmdReset, status: cmdStatus };
+
+async function main() {
+    const argv = process.argv.slice(2);
+    if (!argv.length || ['-h', '--help', 'help'].includes(argv[0])) {
+        console.log(`
+${BOLD}skill_evolution.js${RESET} — Tribunal Skill Evolution Forge
+
+${BOLD}Commands:${RESET}
+  digest [--dry-run] [--head]   Analyze latest git diff and evolve SKILL.md
+                                  --dry-run  : preview without writing
+                                  --head     : diff last commit instead of staged
+  show                           Print current project-idioms/SKILL.md
+  status                         Show digest history and token savings
+  reset                          Clear all idioms and start fresh
+
+${BOLD}Token Budget:${RESET}
+  Raw diff -> Semantic Filter -> Only architectural lines -> LLM
+  Typical savings: 70–90% of tokens. Most trivial commits = 0 tokens.
+`);
+        return;
+    }
+
+    const cmd = argv[0];
+    const rest = argv.slice(1);
+    if (!COMMANDS[cmd]) { console.log(`${RED}✖ Unknown command: '${cmd}'${RESET}`); process.exit(1); }
+    await COMMANDS[cmd](rest);
+}
+
+module.exports = { semanticDelta, architecturalWeight, parseLlmYamlResponse, loadExistingIdioms };
+
+if (require.main === module) {
+    main().catch(err => { console.error(err); process.exit(1); });
+}