@yeaft/webchat-agent 0.1.665 → 0.1.667

package/unify/memory/extract.js

@@ -1,101 +0,0 @@
-/**
- * extract.js — Extract memory-worthy entries from conversation
- *
- * Called by consolidate.js during the Consolidate lifecycle.
- * Uses a single LLM call to identify facts, preferences, skills,
- * lessons, contexts, and relations from conversation messages.
- *
- * Reference: yeaft-unify-core-systems.md §3.1, yeaft-unify-design.md §6.1
- */
-
-import { MEMORY_KINDS } from './store.js';
-import { pickEffort } from '../effort.js';
-
-/**
- * Build the extraction prompt.
- * @param {object[]} messages — conversation messages to analyze
- * @returns {string}
- */
-function buildExtractionPrompt(messages) {
-  const conversation = messages.map(m => {
-    const prefix = m.role === 'user' ? 'User' : m.role === 'assistant' ? 'Assistant' : 'System';
-    return `[${prefix}]: ${m.content}`;
-  }).join('\n\n');
-
-  return `Analyze the following conversation and extract any memorable information worth saving to long-term memory.
-
-For each memory, provide:
-- **name**: A short slug-friendly name (e.g., "user-prefers-typescript", "project-uses-vue3")
-- **kind**: One of: ${MEMORY_KINDS.join(', ')}
-- **scope**: A tree path (e.g., "global", "tech/typescript", "work/project-name")
-- **tags**: Relevant keywords as an array
-- **importance**: "high", "normal", or "low"
-- **content**: 1-3 sentences describing the memory
-
-Memory kinds explained:
-- fact: Objective facts (project structure, tech stack)
-- preference: User preferences (coding style, tools)
-- skill: How to do something (patterns, techniques)
-- lesson: Lessons learned (bugs, pitfalls)
-- context: Temporal context (current OKR, progress)
-- relation: People and relationships (teammates, roles)
-
-Do NOT extract:
-- Specific code snippets (too large, will become stale)
-- Temporary debugging information
-- Trivial greetings or small talk
-
-Return a JSON array of memory objects. If nothing is worth remembering, return an empty array [].
-
-Conversation:
-${conversation}`;
-}
-
-/**
- * Extract memory entries from a set of conversation messages.
- *
- * @param {{ messages: object[], adapter: object, config: object }} params
- * @returns {Promise<object[]>} — extracted memory entries
- */
-export async function extractMemories({ messages, adapter, config }) {
-  if (!messages || messages.length === 0) return [];
-
-  const system = 'You are a memory extraction assistant. Analyze conversations and extract important facts, preferences, and lessons. Return ONLY a valid JSON array, no other text.';
-
-  const extractionPrompt = buildExtractionPrompt(messages);
-
-  try {
-    const result = await adapter.call({
-      model: config.model,
-      system,
-      messages: [{ role: 'user', content: extractionPrompt }],
-      maxTokens: 2048,
-      // task-327c: extract runs inside the consolidate pipeline — the
-      // JSON-structured output benefits from the same 'max' thinking tier.
-      effort: pickEffort({ scenario: 'consolidate' }),
-    });
-
-    const text = result.text.trim();
-
-    // Try to parse JSON array from the response
-    const jsonMatch = text.match(/\[[\s\S]*\]/);
-    if (!jsonMatch) return [];
-
-    const entries = JSON.parse(jsonMatch[0]);
-
-    // Validate and normalize entries
-    return entries
-      .filter(e => e && typeof e === 'object' && e.name && e.content)
-      .map(e => ({
-        name: String(e.name).slice(0, 80),
-        kind: MEMORY_KINDS.includes(e.kind) ? e.kind : 'fact',
-        scope: String(e.scope || 'global'),
-        tags: Array.isArray(e.tags) ? e.tags.map(String) : [],
-        importance: ['high', 'normal', 'low'].includes(e.importance) ? e.importance : 'normal',
-        content: String(e.content),
-      }));
-  } catch {
-    // LLM failure — return empty (non-critical operation)
-    return [];
-  }
-}

package/unify/memory/layout.js

@@ -1,358 +0,0 @@
-/**
- * layout.js — New-layout memory file helpers (tool-on-demand model)
- *
- * Directory layout:
- *   ~/.yeaft/memory/
- *     index.md                 — classification catalog (always injected into system prompt)
- *     user-preferences.md      — merged user preferences (default-injected)
- *     entries/*.md             — atomic entries (existing layout, read by memory_query)
- *     by-project/<slug>.md     — per-project narrative summaries
- *     by-topic/<slug>.md       — per-topic narrative summaries
- *     timeline/<YYYY-MM>.md    — monthly narrative digests
- *
- * Design:
- *   - index.md is a single human-readable file listing all classification files
- *     with a one-line summary and entry count per section.
- *   - Aggregate files (by-project / by-topic / timeline) are narrative prose,
- *     not raw entry concat — produced by Dream.
- *   - user-preferences.md is a deduped accumulation of preferences extracted
- *     from conversations.
- *   - Project header match: basename(cwd) is matched against by-project/<slug>.md
- *     filenames (case-insensitive substring match either direction).
- *
- * This module lives alongside store.js; it does NOT replace MemoryStore.
- * Old MEMORY.md / scopes.md continue to exist for backward compatibility
- * but are no longer maintained by Dream after this refactor.
- */
-
-import { existsSync, mkdirSync, readFileSync, readdirSync, writeFileSync, statSync } from 'fs';
-import { basename, join } from 'path';
-import { isPermissionError } from '../init.js';
-
-// ─── Constants ──────────────────────────────────────────────
-
-/** Classification category dirnames. */
-export const CATEGORY_DIRS = ['by-project', 'by-topic', 'timeline'];
-
-/** Classification single-file names living directly under memory/. */
-export const SINGLE_FILES = ['index.md', 'user-preferences.md'];
-
-/** Approximate character budget for prompt injection (~1.5k tokens). */
-export const PROMPT_INJECTION_CHAR_BUDGET = 6000;
-
-/** Character budget for project header excerpt (~300 tokens). */
-export const PROJECT_HEADER_CHAR_BUDGET = 1200;
-
-// ─── Path helpers ───────────────────────────────────────────
-
-/**
- * @param {string} yeaftDir — e.g. ~/.yeaft
- * @returns {string} — absolute path to memory root
- */
-export function memoryDir(yeaftDir) {
-  return join(yeaftDir, 'memory');
-}
-
-/**
- * Ensure the new-layout directory skeleton exists (idempotent).
- * @param {string} yeaftDir
- */
-export function ensureLayout(yeaftDir) {
-  const root = memoryDir(yeaftDir);
-  const dirs = [root, ...CATEGORY_DIRS.map(d => join(root, d)), join(root, 'entries')];
-  for (const d of dirs) {
-    try {
-      if (!existsSync(d)) mkdirSync(d, { recursive: true, mode: 0o755 });
-    } catch (err) {
-      if (!isPermissionError(err)) throw err;
-    }
-  }
-}
-
-// ─── File I/O ───────────────────────────────────────────────
-
-/**
- * Read a file under memory/ by relative path. Returns '' if missing.
- * @param {string} yeaftDir
- * @param {string} relPath — e.g. 'index.md', 'by-project/foo.md'
- * @returns {string}
- */
-export function readMemoryFile(yeaftDir, relPath) {
-  const fp = join(memoryDir(yeaftDir), relPath);
-  if (!existsSync(fp)) return '';
-  try {
-    return readFileSync(fp, 'utf8');
-  } catch (err) {
-    if (isPermissionError(err)) return '';
-    throw err;
-  }
-}
-
-/**
- * Write a file under memory/ by relative path. Creates parent dir if needed.
- * @param {string} yeaftDir
- * @param {string} relPath
- * @param {string} content
- */
-export function writeMemoryFile(yeaftDir, relPath, content) {
-  ensureLayout(yeaftDir);
-  const fp = join(memoryDir(yeaftDir), relPath);
-  try {
-    writeFileSync(fp, content, { encoding: 'utf8', mode: 0o644 });
-  } catch (err) {
-    if (!isPermissionError(err)) throw err;
-  }
-}
-
-/**
- * List all classification files (relative paths) that currently exist.
- * Includes single files (index.md, user-preferences.md) and per-category files.
- *
- * @param {string} yeaftDir
- * @returns {{ path: string, size: number }[]}
- */
-export function listClassificationFiles(yeaftDir) {
-  const root = memoryDir(yeaftDir);
-  const out = [];
-
-  for (const f of SINGLE_FILES) {
-    const fp = join(root, f);
-    if (existsSync(fp)) {
-      try {
-        out.push({ path: f, size: statSync(fp).size });
-      } catch { /* ignore */ }
-    }
-  }
-
-  for (const dir of CATEGORY_DIRS) {
-    const dp = join(root, dir);
-    if (!existsSync(dp)) continue;
-    try {
-      for (const f of readdirSync(dp)) {
-        if (!f.endsWith('.md')) continue;
-        const fp = join(dp, f);
-        try {
-          out.push({ path: `${dir}/${f}`, size: statSync(fp).size });
-        } catch { /* ignore */ }
-      }
-    } catch { /* ignore */ }
-  }
-
-  return out;
-}
-
-// ─── Project header matching ────────────────────────────────
-
-/**
- * Given a cwd, find the best-matching by-project/<slug>.md filename.
- * Rule: case-insensitive substring match between basename(cwd) and slug
- * (either direction). Returns the first match by sort order, or null.
- *
- * @param {string} yeaftDir
- * @param {string} cwd
- * @returns {string|null} — relative path like 'by-project/claude-web-chat.md', or null
- */
-export function findProjectFile(yeaftDir, cwd) {
-  if (!cwd) return null;
-  const base = basename(cwd).toLowerCase();
-  if (!base) return null;
-
-  const dir = join(memoryDir(yeaftDir), 'by-project');
-  if (!existsSync(dir)) return null;
-
-  let files;
-  try {
-    files = readdirSync(dir).filter(f => f.endsWith('.md')).sort();
-  } catch {
-    return null;
-  }
-
-  for (const f of files) {
-    const slug = f.slice(0, -3).toLowerCase();
-    if (slug === base) return `by-project/${f}`;
-  }
-  for (const f of files) {
-    const slug = f.slice(0, -3).toLowerCase();
-    if (slug.includes(base) || base.includes(slug)) return `by-project/${f}`;
-  }
-
-  return null;
-}
-
-/**
- * Return the leading excerpt of a file capped at charBudget.
- * Trims to the last complete line boundary to avoid mid-line cutoff.
- *
- * @param {string} text
- * @param {number} charBudget
- * @returns {string}
- */
-export function excerpt(text, charBudget) {
-  if (!text) return '';
-  if (text.length <= charBudget) return text;
-  const cut = text.slice(0, charBudget);
-  const lastNl = cut.lastIndexOf('\n');
-  return lastNl > charBudget * 0.6 ? cut.slice(0, lastNl) : cut;
-}
-
-// ─── Index rendering ────────────────────────────────────────
-
-/**
- * Render an auto-generated index.md from the current on-disk layout.
- *
- * Format:
- *   # Memory Index
- *
- *   ## Single files
- *   - user-preferences.md (NNN bytes) — user-written/Dream-merged preferences
- *
- *   ## by-project
- *   - by-project/foo.md (NNN bytes)
- *   - by-project/bar.md (NNN bytes)
- *
- *   ## by-topic ...
- *   ## timeline ...
- *   ## entries
- *   - N atomic entries (use memory_query to search)
- *
- * One-line summaries are pulled from the first non-empty, non-heading line
- * of each file.
- *
- * @param {string} yeaftDir
- * @param {number} entryCount — number of atomic entries (from MemoryStore)
- * @returns {string}
- */
-export function renderIndex(yeaftDir, entryCount) {
-  const root = memoryDir(yeaftDir);
-  const lines = ['# Memory Index', ''];
-
-  // Single files section
-  const singleLines = [];
-  for (const f of SINGLE_FILES) {
-    if (f === 'index.md') continue;
-    const fp = join(root, f);
-    if (!existsSync(fp)) continue;
-    const summary = firstSummary(fp);
-    const size = safeSize(fp);
-    singleLines.push(`- ${f} (${size} bytes)${summary ? ` — ${summary}` : ''}`);
-  }
-  if (singleLines.length) {
-    lines.push('## Single files', '', ...singleLines, '');
-  }
-
-  // Category sections
-  for (const cat of CATEGORY_DIRS) {
-    const dp = join(root, cat);
-    if (!existsSync(dp)) continue;
-    let files;
-    try {
-      files = readdirSync(dp).filter(f => f.endsWith('.md')).sort();
-    } catch {
-      continue;
-    }
-    if (!files.length) continue;
-    lines.push(`## ${cat}`, '');
-    for (const f of files) {
-      const fp = join(dp, f);
-      const summary = firstSummary(fp);
-      const size = safeSize(fp);
-      lines.push(`- ${cat}/${f} (${size} bytes)${summary ? ` — ${summary}` : ''}`);
-    }
-    lines.push('');
-  }
-
-  // Entries section (count only — atomic entries are searched via memory_query)
-  lines.push('## entries', '', `- ${entryCount} atomic entries (use memory_query to search)`, '');
-
-  lines.push(
-    '_Note: use the `memory_load` tool with one or more paths to load a classification',
-    'file in full, or `memory_query` to search atomic entries by keywords/tags._',
-  );
-
-  return lines.join('\n') + '\n';
-}
-
-/**
- * First non-heading non-empty line of a file, trimmed to 120 chars.
- */
-function firstSummary(fp) {
-  try {
-    const raw = readFileSync(fp, 'utf8');
-    for (const line of raw.split('\n')) {
-      const s = line.trim();
-      if (!s) continue;
-      if (s.startsWith('#')) continue;
-      if (s.startsWith('---')) continue;
-      return s.length > 120 ? s.slice(0, 120) + '…' : s;
-    }
-  } catch { /* ignore */ }
-  return '';
-}
-
-function safeSize(fp) {
-  try {
-    return statSync(fp).size;
-  } catch {
-    return 0;
-  }
-}
-
-// ─── Prompt injection builder ───────────────────────────────
-
-/**
- * Build the memory section to inject into the system prompt every turn.
- *
- * Content, in order:
- *   1. index.md (full text — auto-regenerated when missing/stale)
- *   2. user-preferences.md (full text)
- *   3. Project header excerpt (first PROJECT_HEADER_CHAR_BUDGET chars of
- *      the matching by-project/<slug>.md, if cwd matches)
- *
- * Total output is capped at PROMPT_INJECTION_CHAR_BUDGET; later sections
- * are dropped first if over budget.
- *
- * @param {{
- *   yeaftDir: string,
- *   cwd?: string,
- *   entryCount?: number,
- *   language?: 'en' | 'zh',
- * }} params
- * @returns {string}
- */
-export function buildMemoryInjection({ yeaftDir, cwd, entryCount = 0, language = 'en' }) {
-  if (!yeaftDir) return '';
-
-  const heading = language === 'zh' ? '## 记忆索引' : '## Memory Index';
-  const prefHeading = language === 'zh' ? '## 用户偏好' : '## User Preferences';
-  const projectHeading = language === 'zh' ? '## 当前项目摘要' : '## Current Project Summary';
-
-  let indexText = readMemoryFile(yeaftDir, 'index.md');
-  if (!indexText.trim()) {
-    // Auto-generate an index on the fly so the LLM always sees something useful.
-    indexText = renderIndex(yeaftDir, entryCount);
-  }
-
-  const prefText = readMemoryFile(yeaftDir, 'user-preferences.md');
-
-  let projectText = '';
-  const projectRel = cwd ? findProjectFile(yeaftDir, cwd) : null;
-  if (projectRel) {
-    projectText = excerpt(readMemoryFile(yeaftDir, projectRel), PROJECT_HEADER_CHAR_BUDGET);
-  }
-
-  const sections = [];
-  sections.push(`${heading}\n${indexText.trim()}`);
-  if (prefText.trim()) sections.push(`${prefHeading}\n${prefText.trim()}`);
-  if (projectText.trim()) sections.push(`${projectHeading} (${projectRel})\n${projectText.trim()}`);
-
-  // Enforce total char budget — drop from the end.
-  let combined = sections.join('\n\n');
-  while (combined.length > PROMPT_INJECTION_CHAR_BUDGET && sections.length > 1) {
-    sections.pop();
-    combined = sections.join('\n\n');
-  }
-  if (combined.length > PROMPT_INJECTION_CHAR_BUDGET) {
-    combined = excerpt(combined, PROMPT_INJECTION_CHAR_BUDGET);
-  }
-  return combined;
-}

package/unify/memory/schema.js

@@ -1,166 +0,0 @@
-/**
- * schema.js — task-334f R6 Memory schema constants.
- *
- * References:
- *   §Δ22 Memory 本质定性 (6 公理)
- *   §Δ23 Memory Entry Schema 扩字段
- *   §Δ25 Shard 语义初始分类
- *   §Δ26.3 Shard 软上限
- *
- * This module is **data only** — no I/O, no LLM calls. It is safe to import
- * from any layer (store, recall, tools) without circular hazards.
- */
-
-// ─── §Δ25.1 VP-memory default shard set ─────────────────────────
-export const VP_DEFAULT_SHARDS = Object.freeze([
-  'skill',
-  'relations',
-  'lessons',
-  'preferences',
-]);
-
-// ─── §Δ25.2 Feature-memory fixed 5 shards ───────────────────────
-export const FEATURE_SHARDS = Object.freeze([
-  'decision',
-  'progress',
-  'context',
-  'blocker',
-  'artifact',
-]);
-
-// ─── §Δ25.3 User-memory default shard set ───────────────────────
-export const USER_SHARDS = Object.freeze([
-  'profile',
-  'preferences',
-  'projects',
-  'goals',
-  'relations',
-]);
-
-/**
- * §Δ26.3 soft-cap table. `project-<slug>` is matched via the dedicated
- * helper `softCapFor()` below because its key is dynamic.
- *
- * Shape: { entries: number, bytes: number }
- */
-export const SOFT_CAPS = Object.freeze({
-  // VP
-  skill:       { entries: 80,  bytes: 64 * 1024 },
-  lessons:     { entries: 80,  bytes: 64 * 1024 },
-  preferences: { entries: 80,  bytes: 64 * 1024 },
-  relations:   { entries: 50,  bytes: 32 * 1024 },
-    // Feature (Δ26.3 feature-memory row)
-  decision: { entries: 40, bytes: 24 * 1024 },
-  progress: { entries: 40, bytes: 24 * 1024 },
-  context:  { entries: 40, bytes: 24 * 1024 },
-  blocker:  { entries: 40, bytes: 24 * 1024 },
-  artifact: { entries: 40, bytes: 24 * 1024 },
-  // User (Δ26.3 user-memory row — 60 entries / 48 KiB per shard)
-  profile:  { entries: 60, bytes: 48 * 1024 },
-  projects: { entries: 60, bytes: 48 * 1024 },
-  goals:    { entries: 60, bytes: 48 * 1024 },
-});
-
-/** Project shards are dynamic: `project-<slug>` → 150 entries / 128 KiB. */
-export const PROJECT_SHARD_SOFT_CAP = Object.freeze({
-  entries: 150,
-  bytes: 128 * 1024,
-});
-
-/** Default cap if callers ask for a shard not in the canonical set. */
-export const DEFAULT_SOFT_CAP = Object.freeze({
-  entries: 80,
-  bytes: 64 * 1024,
-});
-
-/**
- * Trigger threshold: when a dream sweep finds ≥ PROJECT_DERIVE_THRESHOLD
- * memory entries tagged with the same groupId, it may derive a
- * `project-<slug>` shard. (§Δ25.1)
- */
-export const PROJECT_DERIVE_THRESHOLD = 30;
-
-/**
- * Max number of VP-memory shards — re-compression kicks in when exceeded
- * (§Δ25.1: "最大 shard 数软上限 12").
- */
-export const MAX_VP_SHARDS = 12;
-
-/**
- * Return the soft cap for a given shard name.
- * Handles the dynamic `project-<slug>` case.
- */
-export function softCapFor(shardName) {
-  if (typeof shardName !== 'string' || !shardName) return DEFAULT_SOFT_CAP;
-  if (shardName.startsWith('project-')) return PROJECT_SHARD_SOFT_CAP;
-  return SOFT_CAPS[shardName] || DEFAULT_SOFT_CAP;
-}
-
-/**
- * Build a schema object suitable for `openShardStore(dir, schema)` (334o).
- *
- * @param {'vp'|'feature'|'user'} kind
- * @param {{ extraShards?: string[] }} [opts]  e.g. existing project shards
- * @returns {{ shards: string[], softCap: Record<string,{entries:number,bytes:number}>, defaultSoftCap: object }}
- */
-export function buildShardSchema(kind, opts = {}) {
-  let shards;
-  switch (kind) {
-    case 'vp':      shards = [...VP_DEFAULT_SHARDS]; break;
-    case 'feature': shards = [...FEATURE_SHARDS]; break;
-    case 'user':    shards = [...USER_SHARDS]; break;
-    default: throw new Error(`buildShardSchema: unknown kind "${kind}"`);
-  }
-  if (Array.isArray(opts.extraShards)) {
-    for (const s of opts.extraShards) {
-      if (typeof s === 'string' && s && !shards.includes(s)) shards.push(s);
-    }
-  }
-  const softCap = {};
-  for (const s of shards) softCap[s] = softCapFor(s);
-  return {
-    shards,
-    softCap,
-    defaultSoftCap: DEFAULT_SOFT_CAP,
-  };
-}
-
-/**
- * R6 memory-entry authored-by enum (§Δ23).
- * Free-form strings are allowed; these are canonical examples.
- */
-export const AUTHORED_BY = Object.freeze({
-  VP:             (vpId) => `vp:${vpId}`,
-  USER:           (uid)  => `user:${uid}`,
-  SUMMARY:        'system:summary-extractor',
-  DREAM:          'system:dream',
-});
-
-/**
- * Validate an R6 entry shape (schema-level, no I/O).
- * Throws on structural violation. Callers that want soft warnings should
- * wrap in try/catch.
- */
-export function validateR6Entry(entry) {
-  if (!entry || typeof entry !== 'object') throw new Error('entry must be an object');
-  if (!entry.id || typeof entry.id !== 'string') throw new Error('entry.id required');
-  if (!entry.shard || typeof entry.shard !== 'string') throw new Error('entry.shard required');
-  if (!entry.kind || typeof entry.kind !== 'string') throw new Error('entry.kind required');
-  // sourceRef is required except for identity/preference pure-declaration entries.
-  const needsSourceRef = !(entry.kind === 'identity' || entry.kind === 'preference');
-  if (needsSourceRef) {
-    if (!entry.sourceRef || typeof entry.sourceRef !== 'object') {
-      throw new Error('entry.sourceRef required for kind=' + entry.kind);
-    }
-    if (!Array.isArray(entry.sourceRef.msgIds) || entry.sourceRef.msgIds.length === 0) {
-      throw new Error('entry.sourceRef.msgIds required (non-empty array)');
-    }
-  }
-  if (entry.supersedes != null && !Array.isArray(entry.supersedes)) {
-    throw new Error('entry.supersedes must be an array when present');
-  }
-  if (entry.supersededBy != null && typeof entry.supersededBy !== 'string') {
-    throw new Error('entry.supersededBy must be a string when present');
-  }
-  return true;
-}