@yeaft/webchat-agent 0.1.530 → 0.1.532

package/connection/message-router.js

@@ -36,7 +36,7 @@ import { sendToServer, flushMessageBuffer } from './buffer.js';
 import { handleRestartAgent, handleUpgradeAgent } from './upgrade.js';
 import { loadMcpServers, updateMcpConfig } from '../mcp.js';
 import { getLlmConfig, updateLlmConfig, getUnifySettings, updateUnifySettings } from '../unify/config-api.js';
-import { handleUnifyChat, handleUnifyModeSwitch, handleUnifyModelSwitch, resetUnifySession, handleUnifyLoadHistory, handleUnifyMergeThread, handleUnifyForkThread, handleUnifyAbortThread, handleUnifyAbortAll, handleUnifyVpSubscribe, handleUnifyVpCreate, handleUnifyVpUpdate, handleUnifyVpDelete, handleUnifyVpRead } from '../unify/web-bridge.js';
+import { handleUnifyChat, handleUnifyModeSwitch, handleUnifyModelSwitch, resetUnifySession, handleUnifyLoadHistory, handleUnifyMergeThread, handleUnifyForkThread, handleUnifyAbortThread, handleUnifyAbortAll, handleUnifyVpSubscribe, handleUnifyVpCreate, handleUnifyVpUpdate, handleUnifyVpDelete, handleUnifyVpRead, handleUnifyTaskMessage, handleUnifyUserMemoryWrite, handleUnifyUserMemoryRemove } from '../unify/web-bridge.js';
 
 export async function handleMessage(msg) {
   switch (msg.type) {
@@ -420,6 +420,22 @@ export async function handleMessage(msg) {
       handleUnifyVpRead(msg);
       break;
 
+    // task-334h (R6 §Δ28 / §Δ31.6): task-scoped direct message echo.
+    // Replaces the withdrawn R3 `unify_task_private_chat`. Agent validates,
+    // stamps msgId + ts, and broadcasts the `task_message` mirror back.
+    case 'unify_task_message':
+      handleUnifyTaskMessage(msg);
+      break;
+
+    // task-334h (R6 §Δ29): user-memory skeleton. Payload schema + event
+    // names are wire-frozen here; real ingestion lands in task-334l.
+    case 'unify_user_memory_write':
+      handleUnifyUserMemoryWrite(msg);
+      break;
+    case 'unify_user_memory_remove':
+      handleUnifyUserMemoryRemove(msg);
+      break;
+
     // Expert roles definition (for ExpertPanel detail view)
     case 'get_expert_roles': {
       const { getExpertRolesDefinition } = await import('../expert-roles.js');

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.530",
+  "version": "0.1.532",
   "description": "Remote agent for Yeaft WebChat — connects worker machines to the central server",
   "main": "index.js",
   "type": "module",

package/unify/engine.js

@@ -310,6 +310,9 @@ export class Engine {
       memoryInjection,
       compactSummary,
       skillContent,
+      // task-334f: memory_trace tool is now registered (49 → 51 tools), so
+      // unlock the core_memory meta-line behind 334e's feature flag.
+      memoryTraceAvailable: true,
     });
   }
 

package/unify/memory/migrate-r5-to-r6.js

@@ -0,0 +1,61 @@
+/**
+ * migrate-r5-to-r6.js — task-334f §Δ23 migration stub.
+ *
+ * Legacy (R5): `~/.yeaft/memory/entries/<slug>.md` plus numeric shard files
+ *              `memory-001.md`, `memory-002.md`, ...
+ *
+ * R6:          `~/.yeaft/memory/vp/<vpId>/memory-<semantic>.md`
+ *              Semantic shards: skill / relations / lessons / preferences /
+ *                               project-<slug>
+ *
+ * This slice (334f) only DEFINES the API surface and a dry-run classifier.
+ * The actual batch migration runs in 334i; 334f does not mutate disk.
+ */
+
+import { existsSync, readdirSync, readFileSync } from 'fs';
+import { join } from 'path';
+import { parseEntry } from './store.js';
+import { classifyLegacyEntryToShard } from './shard-store.js';
+
+/**
+ * Produce a migration plan without applying it.
+ *
+ * @param {string} legacyEntriesDir  e.g. ~/.yeaft/memory/entries
+ * @returns {{
+ *   totalEntries: number,
+ *   plan: Array<{ slug: string, shard: string, kind: string, tags: string[] }>,
+ *   byShard: Record<string, number>
+ * }}
+ */
+export function planR5ToR6Migration(legacyEntriesDir) {
+  if (!legacyEntriesDir || !existsSync(legacyEntriesDir)) {
+    return { totalEntries: 0, plan: [], byShard: {} };
+  }
+  const files = readdirSync(legacyEntriesDir).filter(f => f.endsWith('.md'));
+  const plan = [];
+  const byShard = {};
+  for (const file of files) {
+    const raw = readFileSync(join(legacyEntriesDir, file), 'utf8');
+    const entry = parseEntry(raw);
+    if (!entry) continue;
+    const shard = classifyLegacyEntryToShard(entry);
+    plan.push({
+      slug: file.replace(/\.md$/, ''),
+      shard,
+      kind: entry.kind || 'fact',
+      tags: entry.tags || [],
+    });
+    byShard[shard] = (byShard[shard] || 0) + 1;
+  }
+  return { totalEntries: plan.length, plan, byShard };
+}
+
+/**
+ * Apply the migration. STUB — 334i will fill in the body writer. 334f keeps
+ * this function exported so downstream tests can assert the hook exists.
+ *
+ * @param {object} _opts  { legacyEntriesDir, targetDir, vpId, dryRun }
+ */
+export async function applyR5ToR6Migration(_opts) {
+  throw new Error('applyR5ToR6Migration: not yet implemented (task-334i)');
+}

package/unify/memory/recall-r6.js

@@ -0,0 +1,291 @@
+/**
+ * recall-r6.js — task-334f R6 4-step recall pipeline (§Δ24.2).
+ *
+ * Pipeline:
+ *   Step 1: Shard classifier   — fastModel OR keyword heuristic → top-1~2 shards
+ *   Step 2: In-shard candidate — scan only the selected shards, filter by kind/tags/pinned
+ *   Step 3: LLM rerank         — pick top-7 from candidates
+ *   Step 4: Inject body only   — return entries without sourceRef (§Δ23)
+ *
+ * The legacy R5 recall (agent/unify/memory/recall.js) remains for back-compat
+ * of callers still on MemoryStore. R6 callers should use `recallR6()` below.
+ */
+
+import { createHash } from 'crypto';
+import { pickEffort } from '../effort.js';
+import {
+  VP_DEFAULT_SHARDS,
+  TASK_SHARDS,
+  USER_SHARDS,
+} from './schema.js';
+
+const MAX_RECALL_RESULTS = 7;
+const MAX_CANDIDATES = 15;
+
+// ─── Step 1: Shard Classifier ───────────────────────────────────
+
+/**
+ * Keyword → shard heuristic. Zero-cost fallback when fastModel is unavailable.
+ *
+ * The lexicons are intentionally small: classifier output only needs to point
+ * at the most likely shard; LLM rerank in Step 3 catches misses. Empirically
+ * this covers > 70% of queries with zero LLM cost.
+ */
+const SHARD_LEXICON = {
+  skill:       ['code', 'api', 'library', 'framework', 'implement', 'debug', 'syntax', 'typescript', 'vue', 'pattern', '代码', '实现', '调试', '语法', '模式', '技术'],
+  lessons:     ['mistake', 'avoid', 'lesson', 'pitfall', 'gotcha', 'bug', 'regression', '教训', '避坑', '坑', '踩坑', '反模式'],
+  preferences: ['prefer', 'like', 'style', 'convention', 'favorite', '偏好', '风格', '习惯', '喜欢'],
+  relations:   ['colleague', 'partner', 'team', 'user', 'collaborator', 'vp', '同事', '队友', '协作', '关系'],
+  // Task-memory shards
+  decision:    ['decide', 'decision', 'chose', 'picked', 'resolved', '决定', '决策', '选择'],
+  progress:    ['done', 'progress', 'milestone', 'shipped', 'finished', '完成', '进度', '交付'],
+  context:     ['background', 'context', 'requirement', 'scope', '背景', '需求', '范围'],
+  blocker:     ['block', 'stuck', 'blocker', 'issue', 'waiting', '阻塞', '卡住', '等待'],
+  artifact:    ['pr', 'commit', 'doc', 'file', 'link', 'artifact', 'deliverable', 'commit', '产出', '文档'],
+  // User-memory shards
+  profile:     ['name', 'role', 'background', 'who', 'identity', '身份', '角色', '背景'],
+  projects:    ['project', 'repo', '项目', '仓库'],
+  goals:       ['goal', 'target', 'okr', 'plan', '目标', '计划'],
+};
+
+/**
+ * Pick top-N shards via keyword heuristic.
+ * @param {string} prompt
+ * @param {string[]} availableShards  all shards present in this store
+ * @param {number} [topN=2]
+ */
+export function classifyShardsByKeyword(prompt, availableShards, topN = 2) {
+  if (!prompt || !availableShards || availableShards.length === 0) return [];
+  const lower = prompt.toLowerCase();
+  const scores = new Map();
+  for (const shard of availableShards) {
+    // project-<slug> shards score via their slug
+    const lex = SHARD_LEXICON[shard] || [];
+    let score = 0;
+    for (const kw of lex) {
+      if (lower.includes(kw)) score += 1;
+    }
+    if (shard.startsWith('project-')) {
+      const slug = shard.slice('project-'.length);
+      if (slug && lower.includes(slug.toLowerCase())) score += 3;
+    }
+    if (score > 0) scores.set(shard, score);
+  }
+  const ranked = [...scores.entries()].sort((a, b) => b[1] - a[1]).map(([s]) => s);
+  if (ranked.length >= 1) return ranked.slice(0, topN);
+  // Fallback: first N defaults in the shard set
+  return availableShards.slice(0, topN);
+}
+
+/**
+ * Shard classifier — calls fastModel if adapter provided, else falls back to
+ * the keyword heuristic. Budget < 200 tokens (§Δ24.2).
+ */
+export async function classifyShards({
+  prompt,
+  availableShards,
+  adapter,
+  fastModel,
+  topN = 2,
+}) {
+  if (!adapter || !fastModel) {
+    return classifyShardsByKeyword(prompt, availableShards, topN);
+  }
+  const system = `You classify user queries into memory shards. Return ONLY a JSON array of up to ${topN} shard names from the provided list. No prose.`;
+  const user = `Available shards: ${JSON.stringify(availableShards)}
+Query: ${JSON.stringify(prompt)}
+Return JSON array of up to ${topN} most relevant shard names.`;
+  try {
+    const res = await adapter.call({
+      model: fastModel,
+      system,
+      messages: [{ role: 'user', content: user }],
+      maxTokens: 64,
+      effort: pickEffort({ scenario: 'recall' }),
+    });
+    const m = (res.text || '').match(/\[[\s\S]*?\]/);
+    if (!m) return classifyShardsByKeyword(prompt, availableShards, topN);
+    const arr = JSON.parse(m[0]);
+    const valid = arr.filter(s => typeof s === 'string' && availableShards.includes(s));
+    if (valid.length === 0) return classifyShardsByKeyword(prompt, availableShards, topN);
+    return valid.slice(0, topN);
+  } catch {
+    return classifyShardsByKeyword(prompt, availableShards, topN);
+  }
+}
+
+// ─── Step 2: In-shard candidate generation ──────────────────────
+
+function collectCandidates(memoryShardStore, { shards, kind, tags, pinned }) {
+  const filter = {};
+  if (shards && shards.length) filter.shard = shards.length === 1 ? shards[0] : shards;
+  if (kind) filter.kind = kind;
+  if (tags && tags.length) filter.tags = tags;
+  if (pinned !== undefined) filter.pinned = pinned;
+  const { results } = memoryShardStore.query(filter);
+  return results
+    // Drop superseded entries from the candidate pool (they stay on disk
+    // for memory_trace but should not compete for recall slots).
+    .filter(rec => !rec.supersededBy)
+    .slice(0, MAX_CANDIDATES);
+}
+
+// ─── Step 3: LLM Rerank ─────────────────────────────────────────
+
+async function llmRerank({ adapter, fastModel, prompt, candidates, memoryShardStore }) {
+  if (candidates.length <= MAX_RECALL_RESULTS) {
+    return candidates.map(c => c.id);
+  }
+  const lines = candidates.map((c, i) => {
+    return `${i + 1}. [id=${c.id}] shard=${c.shard} kind=${c.kind || '?'} tags=[${(c.tags || []).join(',')}]`;
+  }).join('\n');
+  const system = `You pick the most relevant memories for the user's prompt. Return ONLY a JSON array of entry ids (up to ${MAX_RECALL_RESULTS}). No prose.`;
+  const user = `User prompt: ${JSON.stringify(prompt)}
+
+Candidate memories:
+${lines}
+
+Return JSON array of up to ${MAX_RECALL_RESULTS} ids.`;
+  try {
+    const res = await adapter.call({
+      model: fastModel,
+      system,
+      messages: [{ role: 'user', content: user }],
+      maxTokens: 256,
+      effort: pickEffort({ scenario: 'recall' }),
+    });
+    const m = (res.text || '').match(/\[[\s\S]*\]/);
+    if (!m) return candidates.slice(0, MAX_RECALL_RESULTS).map(c => c.id);
+    const arr = JSON.parse(m[0]).filter(x => typeof x === 'string');
+    const valid = arr.filter(id => candidates.find(c => c.id === id));
+    if (valid.length === 0) return candidates.slice(0, MAX_RECALL_RESULTS).map(c => c.id);
+    return valid.slice(0, MAX_RECALL_RESULTS);
+  } catch {
+    return candidates.slice(0, MAX_RECALL_RESULTS).map(c => c.id);
+  }
+}
+
+// ─── Step 4: Inject (no sourceRef) ──────────────────────────────
+
+/**
+ * Produce the body-only injection payload (§Δ24.5). Prefix lines with
+ * `[mem:<shard>]` so the LLM knows the category without seeing the id.
+ */
+export function formatForInjection(entries) {
+  return entries.map(e => {
+    const prefix = `[mem:${e.shard}]`;
+    const body = (e.body || '').trim();
+    return `${prefix} ${body}`;
+  }).join('\n\n');
+}
+
+// ─── Fingerprint cache ──────────────────────────────────────────
+
+const _cache = new Map();
+const CACHE_TTL = 5 * 60 * 1000;
+
+function computeFingerprint({ shards, prompt, taskId }) {
+  const head = prompt.slice(0, 200);
+  const input = `${shards.join(',')}|${head}|${taskId || ''}`;
+  return createHash('sha256').update(input).digest('hex').slice(0, 16);
+}
+
+// ─── Entry point ────────────────────────────────────────────────
+
+/**
+ * Run the R6 4-step recall pipeline.
+ *
+ * @param {{
+ *   prompt: string,
+ *   memoryShardStore: object,           // openMemoryShardStore() handle
+ *   adapter?: object,                   // LLM adapter (null = keyword-only)
+ *   fastModel?: string,                 // fast model for classifier + rerank
+ *   availableShards?: string[],         // defaults to store stats
+ *   taskId?: string,
+ *   kind?: string, tags?: string[], pinned?: boolean,
+ * }} params
+ * @returns {Promise<{entries: object[], shards: string[], fingerprint: string, cached: boolean}>}
+ */
+export async function recallR6(params) {
+  const {
+    prompt,
+    memoryShardStore,
+    adapter,
+    fastModel,
+    availableShards,
+    taskId,
+    kind,
+    tags,
+    pinned,
+  } = params;
+
+  if (!prompt || !prompt.trim() || !memoryShardStore) {
+    return { entries: [], shards: [], fingerprint: '', cached: false };
+  }
+
+  const shardsFromStore = availableShards
+    || Object.keys(memoryShardStore.stats().shards);
+
+  if (shardsFromStore.length === 0) {
+    return { entries: [], shards: [], fingerprint: '', cached: false };
+  }
+
+  // Step 1
+  const chosenShards = await classifyShards({
+    prompt,
+    availableShards: shardsFromStore,
+    adapter,
+    fastModel,
+    topN: 2,
+  });
+
+  const fingerprint = computeFingerprint({ shards: chosenShards, prompt, taskId });
+  const cached = _cache.get(fingerprint);
+  if (cached && Date.now() - cached.t < CACHE_TTL) {
+    return { entries: cached.entries, shards: chosenShards, fingerprint, cached: true };
+  }
+
+  // Step 2
+  const candidates = collectCandidates(memoryShardStore, {
+    shards: chosenShards,
+    kind,
+    tags,
+    pinned,
+  });
+  if (candidates.length === 0) {
+    _cache.set(fingerprint, { entries: [], t: Date.now() });
+    return { entries: [], shards: chosenShards, fingerprint, cached: false };
+  }
+
+  // Step 3
+  const selectedIds = await llmRerank({
+    adapter, fastModel, prompt, candidates, memoryShardStore,
+  });
+
+  // Step 4 — load full bodies (minus sourceRef for injection).
+  const entries = [];
+  for (const id of selectedIds) {
+    const full = memoryShardStore.get(id);
+    if (!full) continue;
+    entries.push({
+      id: full.id,
+      shard: full.shard,
+      kind: full.kind,
+      body: full.body,
+      tags: full.tags || [],
+      // deliberately do NOT expose sourceRef here (§Δ23)
+    });
+  }
+  _cache.set(fingerprint, { entries, t: Date.now() });
+  return { entries, shards: chosenShards, fingerprint, cached: false };
+}
+
+export function clearR6RecallCache() {
+  _cache.clear();
+}
+
+export const R6_DEFAULTS = {
+  VP_DEFAULT_SHARDS,
+  TASK_SHARDS,
+  USER_SHARDS,
+};

package/unify/memory/schema.js

@@ -0,0 +1,166 @@
+/**
+ * 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 Task-memory fixed 5 shards ──────────────────────────
+export const TASK_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 },
+  // Task (Δ26.3 task-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'|'task'|'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 'task': shards = [...TASK_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;
+}