@a9i5k4/dsh-auto-memory 2.5.3 → 3.0.0

package/lib/ws-overview-rank.js

@@ -12,8 +12,8 @@
  */
 import path from 'node:path'
 
-/** 日志文件名日期识别(与 index.js readWorkspaceMemory 的 DATE_RE 同源)。 */
-const DATE_RE = /\d{4}-\d{2}-\d{2}/
+/** 日期词干须整串匹配,与 index.js 的 DATE_RE 格式对齐,避免临时/备份文件参与排序(issue #51)。 */
+const DATE_RE = /^(\d{4})-(\d{2})-(\d{2})$/
 
 /**
  * 按记忆活跃度(最新日志 mtime)降序排列工作区。

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@a9i5k4/dsh-auto-memory",
   "description": "Proactive associative memory for DSH: zero-prompt recall injected before the model speaks, three-layer auto-consolidation, skill crystallization, and Astra-style context management - handoff ledgers, PLAN whiteboard, water-level sensing. Local-first, model-agnostic, zero deps. 主动联想记忆+Astra 式上下文管理:自动唤回/自动沉淀/技能固化/交接账本与白板跨窗口续命/水位感知。",
-  "version": "2.5.3",
+  "version": "3.0.0",
   "type": "module",
   "main": "lib/index.js",
   "exports": {

package/python/m7_embedding_v1.py

@@ -13,10 +13,10 @@ Pure re-implementation of the frozen policy for the production sidecar path
                 -> configHash; mismatch = stale = full rebuild
 
 Providers:
-  hash-pre-v1   deterministic stdlib-only embedding (sha256-seeded bag of
+  hash-v1   deterministic stdlib-only embedding (sha256-seeded bag of
                 token-trigram dims). Zero dependencies, zero network. Used
                 by CI and offline protocol tests. NOT a quality provider.
-  bge-m3-pre-v1 real model via transformers (lazy import; requires the
+  bge-m3-v1 real model via transformers (lazy import; requires the
                 pinned local snapshot dir passed in the embedding config).
 
 No DSH file reads; no writes except what the worker explicitly passes in.
@@ -25,9 +25,9 @@ import hashlib
 import json
 import re
 
-PROVIDER_REAL = 'bge-m3-pre-v1'
+PROVIDER_REAL = 'bge-m3-v1'
 PROVIDER_REAL_INT8 = 'bge-m3-onnx-int8-v1'
-PROVIDER_HASH = 'hash-pre-v1'
+PROVIDER_HASH = 'hash-v1'
 CHUNK_POLICY_VERSION = 'm7_chunk_v1'
 CHUNK_MAX_TOKENS = 512
 QUERY_MAX_TOKENS = 256
@@ -74,7 +74,7 @@ def config_hash(provider, model_revision, dimension):
 
 def chunk_id_for(memory_id, record_digest, ordinal):
     return 'chk_' + sha_hex(
-        ('m7-chunk-pre-v1\u0000' + memory_id + '\u0000' + record_digest +
+        ('m7-chunk-v1\u0000' + memory_id + '\u0000' + record_digest +
          '\u0000' + str(ordinal)).encode('utf-8'))[:32]
 
 

package/python/worker_semantic_v1.py

@@ -19,7 +19,7 @@ index_sync rejection matrix, same atomic derived-corpus persistence. Adds:
 
 Embedding backend is selected by an optional JSON config file passed via
 the DSH_M7_EMBEDDING_CONFIG environment variable (no CLI change, no JS
-change): {"provider":"bge-m3-pre-v1"|"hash-pre-v1", "modelDir":...,
+change): {"provider":"bge-m3-v1"|"hash-v1", "modelDir":...,
 "modelRevision":..., "dimension":1024, "torchThreads":16}.
 Without the env var the worker degrades to fake-worker behavior (protocol
 alive, embedding not ready) - never crashes, never changes ack semantics.
@@ -64,7 +64,7 @@ def wsref_of(workspace_key):
     artifact of differing miv values)."""
     canon = canonical_workspace_key(workspace_key)
     return 'wsr_' + hashlib.sha256(
-        ('evidence-wsref-pre-v1\u0000' + canon).encode('utf-8')).hexdigest()[:32]
+        ('evidence-wsref-v1\u0000' + canon).encode('utf-8')).hexdigest()[:32]
 
 
 def _tokenize(text, stopwords=frozenset()):
@@ -640,7 +640,7 @@ class SemanticWorker(base.Worker):
         if not cands:
             return None
         activation_id = 'act_' + base.first32(
-            base.sha_str('m7-semantic-activation-pre-v1\u0000' + obs))
+            base.sha_str('m7-semantic-activation-v1\u0000' + obs))
         created = req.get('sentAt', 0)
         ttl = int(pol['ttlSteps'])
         return {
@@ -690,12 +690,23 @@ class SemanticWorker(base.Worker):
         if act is None:
             return None
         reasons = ','.join(str(x) for x in (out.get('reasonCodes') or []))
+
+        def _num(x):
+            try:
+                return float(x or 0.0)
+            except (TypeError, ValueError):
+                return 0.0
+
+        # 2026-09-14 P1-⑮:reason 串此前只有 lane/decision/reasonCodes,注入侧看不到
+        # 0-1 相似度(intent 概率 / 稠密分 / 融合 margin)。数值段放在 reasonCodes **之前**,
+        # 这样 160 字符截断只会砍掉代码列表,不会砍掉分值。
         act['threshold'] = {
             'policyVersion': featv2.ACTIVATION_POLICY_VERSION,
             'score': round(score, 6),
             'threshold': float(th.get('tauHi', 0.45)),
-            'reason': ('fv2 lane=%s %s %s' % (
-                feats.get('lane'), out.get('decision'), reasons))[:160],
+            'reason': ('fv2 lane=%s %s intent=%.2f dense=%.2f margin=%.2f %s' % (
+                feats.get('lane'), out.get('decision'), score,
+                _num(feats.get('denseTop')), _num(feats.get('margin')), reasons))[:160],
         }
         act['level'] = 'excerpt'
         return act
@@ -1356,7 +1367,7 @@ def main():
     args, _unknown = ap.parse_known_args()
     if args.selftest:
         base.run_selftest()
-        w = SemanticWorker('ep', '', {'provider': 'hash-pre-v1',
+        w = SemanticWorker('ep', '', {'provider': 'hash-v1',
                                       'dimension': 64})
         view = w.embedding_view()
         assert view['enabled'] is True and view['ready'] is False

package/python/worker_v1.py

@@ -22,12 +22,23 @@ import os
 import re
 import sys
 import tempfile
+import time
 
 PROTOCOL = 'm7_wire_v1'
 NAMESPACE = 'dsh-auto-memory'
 INDEX_POLICY = 'index_sync_v1'
 MAX_LINE_BYTES = 256 * 1024
 MAX_RECORDS_PER_PAGE = 64
+# ★2026-09-17 修「索引永久不就绪 → 语义唤回全程降级」：JS 侧契约是「新 miv latest-wins；
+# 旧 in-flight sync abort/cancel」，但 abort 只作用于 JS 的 fetch，Python 从不收到取消帧；
+# 而 active_sync 原先只在 reject_sync / commit 两处清零 ⇒ 一次「begin 之后没收 page/commit」
+# 的同步会把槽永久占住，此后每次 index_begin 都被回 sync-in-progress。
+# 实测取证：engine-switch-state.json = failed:2527 / error:"sync-in-progress" / startedAt:0；
+# 诊断日志 index-not-ready 累计 22945 行（sync-in-progress 占 664），自 2026-08-26 起持续；
+# worker 于 00:49:50 重启后仍有 1593 行 ⇒ 重启只是暂时缓解，状态会重新形成。
+# 本常量 = 「距最后一次活动多久仍无进展即视为持有者已死」的阈值；取 150s > JS 的 syncTimeoutMs
+# (120s)，保证 JS 先放弃、worker 后回收。
+SYNC_TAKEOVER_IDLE_SECONDS = 150
 STDERR_BUDGET = 64
 
 JS_TYPES = frozenset(['health', 'context_push', 'index_sync_begin', 'index_sync_page',
@@ -136,7 +147,7 @@ class Worker:
         self.active_sync = None     # at most one in-flight index sync (JS serializes anyway)
         self.derived = {}           # (workspaceRef, scope) -> current committed entry ONLY (old versions discarded)
         self.counts = {'frames': 0, 'errors': 0, 'acks': 0, 'activations': 0, 'commits': 0,
-                       'cancels': 0, 'close_sessions': 0}
+                       'cancels': 0, 'close_sessions': 0, 'sync_takeovers': 0}
 
     # ---------- outbound ----------
 
@@ -246,7 +257,7 @@ class Worker:
                 continue
             if not RE_HEX64.match(fdig) or not RE_HEX64.match(rdig):
                 continue
-            activation_id = 'act_' + first32(sha_str('m7-fake-activation-pre-v1\u0000' + obs))
+            activation_id = 'act_' + first32(sha_str('m7-fake-activation-v1\u0000' + obs))
             cand = {
                 'candidateId': 'cand_' + first32(sha_str(activation_id + '\u0000' + mid + '\u0000' + str(i))),
                 'memoryId': mid,
@@ -262,7 +273,7 @@ class Worker:
             candidates.append(cand)
         if not candidates:
             return None
-        activation_id = 'act_' + first32(sha_str('m7-fake-activation-pre-v1\u0000' + obs))
+        activation_id = 'act_' + first32(sha_str('m7-fake-activation-v1\u0000' + obs))
         created = req.get('sentAt', 0)
         ttl_steps = 2
         activation = {
@@ -340,11 +351,33 @@ class Worker:
         require((record_count == 0) == (page_count == 0), 'invalid-payload', 'count-consistency')
         require(p.get('indexPolicyVersion') == INDEX_POLICY, 'invalid-payload', 'indexPolicyVersion')
         if self.active_sync is not None:
-            return [self.index_ack(req, p, 'begin', False, reason='sync-in-progress')]
+            prev = self.active_sync
+            prev_id = str(prev.get('syncId', ''))
+            try:
+                last_seen = float(prev.get('lastActivityAt') or prev.get('startedAt') or 0)
+            except Exception:
+                last_seen = 0.0
+            idle = (time.time() - last_seen) if last_seen else 0.0
+            same_key = (str(prev.get('workspaceRef', '')) == ws_ref
+                        and prev.get('scope') == scope)
+            superseded = same_key and prev_id != sync_id
+            idle_dead = idle > SYNC_TAKEOVER_IDLE_SECONDS
+            if superseded or idle_dead:
+                # 只在「同 key 被新 syncId 取代」或「超过阈值无任何活动」时接管；
+                # 跨工作区并发、以及正在推进的长同步，仍按原样回 sync-in-progress。
+                self.counts['sync_takeovers'] = self.counts.get('sync_takeovers', 0) + 1
+                diag('index_sync takeover: %s -> %s (%s)' % (
+                    prev_id or '(none)', sync_id,
+                    'superseded' if superseded else 'idle:%.0fs' % idle))
+                self.active_sync = None
+            else:
+                return [self.index_ack(req, p, 'begin', False, reason='sync-in-progress')]
         self.active_sync = {
             'syncId': sync_id, 'workspaceRef': ws_ref, 'scope': scope, 'memoryIndexVersion': miv,
             'recordCount': record_count, 'pageCount': page_count, 'tuples': tuples,
             'pages': {}, 'next_page': 0, 'received': 0, 'page_digests': [],
+            'startedAt': time.time(),
+            'lastActivityAt': time.time(),
         }
         return [self.index_ack(req, p, 'begin', True)]
 
@@ -412,6 +445,7 @@ class Worker:
         st['next_page'] = page_no + 1
         st['received'] += len(records)
         st['page_digests'].append(recomputed)
+        st['lastActivityAt'] = time.time()
         return [self.index_ack(req, base, 'page', True, extra={
             'pageNo': page_no, 'receivedPages': len(st['pages']), 'receivedRecords': st['received']})]