@furongjun1999/dsh-memory 0.9.8 → 0.9.10
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +1 -1
- package/codebuddy/CODEBUDDY.md +4 -0
- package/docs/README.md +3 -0
- package/docs//350/247/204/350/214/203/AI/347/274/226/347/250/213/351/200/232/347/224/250/350/247/204/350/214/203.md +4 -0
- package/dsh/cordis.yml.example +2 -2
- package/lib/hooks.d.ts +11 -1
- package/lib/hooks.js +58 -3
- package/lib/index.d.ts +1 -1
- package/lib/index.js +2 -2
- package/lib/init.js +11 -2
- package/lib/lib/hook_audit.d.ts +3 -0
- package/lib/lib/hook_audit.js +37 -0
- package/md_cg/autonomy.py +1 -1
- package/md_cg/config_registry.py +329 -329
- package/md_cg/config_registry_bulk.py +2 -0
- package/md_cg/config_render.py +27 -0
- package/md_cg/config_validate.py +10 -8
- package/md_cg/conflict_surface.py +127 -0
- package/md_cg/mcp_server.py +105 -8
- package/md_cg/mdcg.py +60 -2
- package/md_cg/mdcos.py +61 -6
- package/md_cg/pooling.py +487 -485
- package/md_cg/qmd_llm.py +334 -334
- package/md_cg/reach.py +10 -4
- package/md_cg/sleep.py +2 -2
- package/md_cg/sustain.py +3 -3
- package/md_cg/test_auto_defaults.py +463 -463
- package/md_cg/test_autonomy_emotion.py +328 -328
- package/md_cg/test_bm25_default.py +56 -0
- package/md_cg/test_c3_transient_read_negative.py +797 -793
- package/md_cg/test_conflict_surface.py +75 -0
- package/md_cg/test_conflict_value_untrusted.py +63 -0
- package/md_cg/test_exact_text_cache.py +3 -0
- package/md_cg/test_hotcache_tool.py +121 -0
- package/md_cg/test_index_snapshot_fastpath.py +72 -0
- package/md_cg/test_issue274_read_node_id_none.py +29 -6
- package/md_cg/test_issue561_pending_review.py +89 -5
- package/md_cg/test_issue578_service_info_gate.py +264 -0
- package/md_cg/test_p3_legacy_closure.py +5 -2
- package/md_cg/test_p43_pooling.py +419 -412
- package/md_cg/test_reach.py +5 -1
- package/md_cg/test_reach_exits.py +54 -0
- package/md_cg/test_reach_single_char.py +74 -0
- package/md_cg/test_search_bm25_default.py +60 -0
- package/md_cg/test_sleep_p1.py +1032 -1032
- package/md_cg/test_untrusted_audit.py +55 -0
- package/md_cg/test_untrusted_classifier.py +42 -0
- package/md_cg/test_untrusted_enforcement.py +66 -0
- package/md_cg/untrusted.py +45 -0
- package/md_cg/untrusted_audit.py +87 -0
- package/md_cg/value_conflict.py +71 -0
- package/package.json +7 -7
- package/src/hooks.ts +986 -932
- package/src/index.ts +2 -2
- package/src/init.ts +7 -2
- package/src/lib/hook_audit.ts +32 -0
- package/zcode/AGENTS.md +4 -0
|
@@ -296,6 +296,8 @@ ENV_ROWS = [
|
|
|
296
296
|
('MDCG_SPREAD_HOPS', 'TUNE', 'tune', 'env', True, "'2'", 'str', 'value', 'MDCG_SPREAD_HOPS', 'value', 'env_literal', 'md_cg/mdcg.py', '', 'uncalibrated', '由生成器自 AST 复核(第一批未见标定读数)', 'unknown', ''),
|
|
297
297
|
('MDCG_LAYER_BOOST', 'TUNE', 'tune', 'env', True, "''", 'str', 'unset', 'MDCG_LAYER_BOOST', 'unset', 'env_literal', 'md_cg/mdcg.py', '', 'uncalibrated', '由生成器自 AST 复核(第一批未见标定读数)', 'unknown', ''),
|
|
298
298
|
('MDCG_HOTCACHE', 'TUNE', 'tune', 'env', True, 'None', 'str', 'bool:1', 'MDCG_HOTCACHE', 'off', 'env_flag', 'md_cg/mdcos.py', '', 'uncalibrated', '由生成器自 AST 复核(第一批未见标定读数);有效缺省(盘点文字,未 AST 复核):未设 = 关(=="1")', 'unknown', ''),
|
|
299
|
+
('MDCG_BM25', 'TUNE', 'tune', 'env', True, 'None', 'str', 'bool:1', 'MDCG_BM25', 'on', 'env_flag', 'md_cg/mdcos.py', '', 'uncalibrated', '2026-10-21 手补:BM25 并入默认检索阶梯后默认开(未设 = 开);显式 0 关掉', 'unknown', ''),
|
|
300
|
+
('MDCG_RELOAD_MIN_INTERVAL', 'TUNE', 'tune', 'env', True, "'0'", 'str', 'value', 'MDCG_RELOAD_MIN_INTERVAL', 'value', 'env_literal', 'md_cg/mdcg.py', '', 'uncalibrated', '2026-10-21 手补:BM25 并入默认检索阶梯后默认开(未设 = 开);显式 0 关掉', 'unknown', ''),
|
|
299
301
|
('MDCG_HOTCACHE_MAX_NODES', 'CALIB', 'calib', 'env', True, 'None', 'str', 'value', 'MDCG_HOTCACHE_MAX_NODES', 'value', 'env_helper_default', 'md_cg/hotcache.py', 'MAX_NODES', 'uncalibrated', '由生成器自 AST 复核(第一批未见标定读数);有效缺省(盘点文字,未 AST 复核):未设 = 模块常量 MAX_NODES=256', 'unknown', ''),
|
|
300
302
|
('MDCG_HOTCACHE_MAX_QUERIES', 'TUNE', 'tune', 'env', True, 'None', 'str', 'value', 'MDCG_HOTCACHE_MAX_QUERIES', 'value', 'env_helper_default', 'md_cg/hotcache.py', 'MAX_QUERIES', 'uncalibrated', '由生成器自 AST 复核(第一批未见标定读数);有效缺省(盘点文字,未 AST 复核):未设 = 模块常量 MAX_QUERIES=64', 'unknown', ''),
|
|
301
303
|
('MDCG_READ_CACHE', 'TUNE', 'tune', 'env', True, "'1'", 'str', 'bool:1', 'MDCG_READ_CACHE', 'on', 'env_literal', 'md_cg/readcache.py', '', 'uncalibrated', '由生成器自 AST 复核(第一批未见标定读数)', 'unknown', ''),
|
package/md_cg/config_render.py
CHANGED
|
@@ -39,6 +39,22 @@ ENUMS = {
|
|
|
39
39
|
_ENV_SLEEP_MERGE: tuple(_sl.SLEEP_MERGE_MODES),
|
|
40
40
|
"IMGSKILL_BACKEND": ("auto", "magick", "pillow"),
|
|
41
41
|
"MDCG_SCORE_MODE": ("legacy", "jaccard"),
|
|
42
|
+
"MDCG_EN_ATOMS": ("norm", "raw"),
|
|
43
|
+
}
|
|
44
|
+
|
|
45
|
+
#: 源码只认 "1" 为真值的 env 开关(人工核验:读点是 `os.environ.get(X) == "1"`,
|
|
46
|
+
#: 因此 "true"/"yes"/"on" 在运行时不生效;校验器必须按同口径拒收,避免「报合法但实际关」。
|
|
47
|
+
TRUE_LITERALS = {
|
|
48
|
+
"MDCG_RETRIEVAL_PIPELINE": ["1"],
|
|
49
|
+
"MDCG_SEMANTIC": ["1"],
|
|
50
|
+
"MDCG_UNIFY_QUERY": ["1"],
|
|
51
|
+
"MDCG_REACH": ["1"],
|
|
52
|
+
"MDCG_REACH_DIFFUSE": ["1"],
|
|
53
|
+
"MDCG_HOTCACHE": ["1"],
|
|
54
|
+
"MDCG_QMD_SYS": ["1"],
|
|
55
|
+
"MDCG_PROPOSAL_EMOTION": ["1"],
|
|
56
|
+
"MDCG_POOLING": ["1"],
|
|
57
|
+
"MDCG_READ_CACHE": ["1"],
|
|
42
58
|
}
|
|
43
59
|
|
|
44
60
|
#: 明显带值域约束的数值参数(**人工核验**,取自源码夹取/断言;未列者不约束)。
|
|
@@ -177,6 +193,15 @@ def _rule_for(e):
|
|
|
177
193
|
elif e.get("semantic") in ("on", "off", "fail-open"):
|
|
178
194
|
r["type"] = "bool"
|
|
179
195
|
r["bool_like"] = True
|
|
196
|
+
# env_flag 的源码口径恒为 `os.environ.get(X) == "1"`(只认 "1" 为真):
|
|
197
|
+
# 若只给 bool_like 宽词表,校验器会认 "true" 而运行时实际是关(issue #138 同型)。
|
|
198
|
+
if r.get("type") == "bool" and (e.get("anchor") or {}).get("how") == "env_flag":
|
|
199
|
+
r["true_literals"] = ["1"]
|
|
200
|
+
|
|
201
|
+
if name in TRUE_LITERALS:
|
|
202
|
+
r["type"] = "bool"
|
|
203
|
+
r["bool_like"] = True
|
|
204
|
+
r["true_literals"] = list(TRUE_LITERALS[name])
|
|
180
205
|
# ①b 域带枚举("enum:a|b")⇒ 枚举约束(覆盖 semantic 推出的布尔;issue #138 · EN_ATOMS)
|
|
181
206
|
if dom.startswith("enum:"):
|
|
182
207
|
r["type"] = "str"
|
|
@@ -187,6 +212,8 @@ def _rule_for(e):
|
|
|
187
212
|
if name in ENUMS:
|
|
188
213
|
r["type"] = "str"
|
|
189
214
|
r["enum"] = list(ENUMS[name])
|
|
215
|
+
r.pop("bool_like", None) # 枚举键不该再接受 "1"/"true" 当合法值
|
|
216
|
+
r.pop("true_literals", None)
|
|
190
217
|
# ③ 已知数值值域 → 数值类型 + 上下限
|
|
191
218
|
if name in RANGES:
|
|
192
219
|
lo, hi, t = RANGES[name]
|
package/md_cg/config_validate.py
CHANGED
|
@@ -19,24 +19,25 @@ RULES = {
|
|
|
19
19
|
'AEIS_DB': {'type': 'str'},
|
|
20
20
|
'AEIS_WORKSPACE': {'type': 'str'},
|
|
21
21
|
'IMGSKILL_BACKEND': {'enum': ['auto', 'magick', 'pillow'], 'type': 'str'},
|
|
22
|
-
'IMGSKILL_PILLOW': {'bool_like': True, 'type': 'bool'},
|
|
22
|
+
'IMGSKILL_PILLOW': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
23
23
|
'MDCG_ACTOR': {'type': 'str'},
|
|
24
24
|
_ENV_AUTONOMY_MODE: {'enum': ['plan', 'confirm', 'full'], 'type': 'str'},
|
|
25
25
|
'MDCG_AUTO_EVOLVE': {'bool_like': True, 'type': 'bool'},
|
|
26
26
|
'MDCG_AUTO_SCRUB': {'bool_like': True, 'type': 'bool'},
|
|
27
27
|
'MDCG_AUTO_TIDY': {'bool_like': True, 'type': 'bool'},
|
|
28
|
+
'MDCG_BM25': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
28
29
|
'MDCG_BUCKET_MIN_SIM': {'max': 1.0, 'min': 0.0, 'type': 'float'},
|
|
29
30
|
'MDCG_BUCKET_TOPK': {'type': 'str'},
|
|
30
31
|
'MDCG_CHAIN_TYPES': {'type': 'str'},
|
|
31
32
|
'MDCG_CLEARANCE': {'enum': ['public', 'internal', 'restricted', 'private', 'secret'],
|
|
32
33
|
'type': 'str'},
|
|
33
|
-
'MDCG_CN_GRAMS': {'bool_like': True, 'type': 'bool'},
|
|
34
|
+
'MDCG_CN_GRAMS': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
34
35
|
'MDCG_D_META': {'bool_like': True, 'type': 'bool'},
|
|
35
36
|
'MDCG_EN_ATOMS': {'enum': ['norm', 'raw'], 'type': 'str'},
|
|
36
37
|
'MDCG_EXPLORE_BUDGET_MAX': {'type': 'str'},
|
|
37
38
|
'MDCG_EXPLORE_BUDGET_WINDOW': {'type': 'str'},
|
|
38
39
|
'MDCG_EXPORT_ROOT': {'bool_like': True, 'type': 'bool'},
|
|
39
|
-
'MDCG_FRESHNESS': {'bool_like': True, 'type': 'bool'},
|
|
40
|
+
'MDCG_FRESHNESS': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
40
41
|
'MDCG_GATE_S1B_BUCKET': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
41
42
|
'MDCG_GATE_S1_DOMAIN': {'bool_like': True, 'type': 'bool'},
|
|
42
43
|
'MDCG_GATE_S2_COND': {'bool_like': True, 'type': 'bool'},
|
|
@@ -55,16 +56,17 @@ RULES = {
|
|
|
55
56
|
'MDCG_NEG_LAMBDA': {'max': 1.0, 'min': 0.0, 'type': 'float'},
|
|
56
57
|
'MDCG_NEG_SIM': {'max': 1.0, 'min': 0.0, 'type': 'float'},
|
|
57
58
|
'MDCG_PLAIN_CACHE_MAX': {'type': 'str'},
|
|
58
|
-
'MDCG_POOLING': {'bool_like': True, 'type': 'bool'},
|
|
59
|
-
'MDCG_PROPOSAL_EMOTION': {'bool_like': True, 'type': 'bool'},
|
|
59
|
+
'MDCG_POOLING': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
60
|
+
'MDCG_PROPOSAL_EMOTION': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
60
61
|
'MDCG_QMD_SYS': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
61
62
|
'MDCG_REACH': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
62
63
|
'MDCG_REACH_DIFFUSE': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
63
64
|
'MDCG_REACH_TTL': {'min': 0.0, 'type': 'float'},
|
|
64
65
|
'MDCG_READ_CACHE': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
65
66
|
'MDCG_RECONCILE': {'bool_like': True, 'type': 'bool'},
|
|
67
|
+
'MDCG_RELOAD_MIN_INTERVAL': {'type': 'str'},
|
|
66
68
|
'MDCG_RETRIEVAL_PIPELINE': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
67
|
-
'MDCG_S7_FRESHNESS': {'bool_like': True, 'type': 'bool'},
|
|
69
|
+
'MDCG_S7_FRESHNESS': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
68
70
|
'MDCG_SCORE_MODE': {'enum': ['legacy', 'jaccard'], 'type': 'str'},
|
|
69
71
|
'MDCG_SEMANTIC': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
70
72
|
'MDCG_SLEEP': {'bool_like': True, 'type': 'bool'},
|
|
@@ -77,13 +79,13 @@ RULES = {
|
|
|
77
79
|
'MDCG_SPREAD_DECAY': {'max': 1.0, 'min': 0.0, 'type': 'float'},
|
|
78
80
|
'MDCG_SPREAD_GAIN': {'max': 1.0, 'min': 0.0, 'type': 'float'},
|
|
79
81
|
'MDCG_SPREAD_HOPS': {'min': 1, 'type': 'int'},
|
|
80
|
-
'MDCG_STATUS_HEAD': {'bool_like': True, 'type': 'bool'},
|
|
82
|
+
'MDCG_STATUS_HEAD': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
81
83
|
'MDCG_SUSTAIN': {'bool_like': True, 'type': 'bool'},
|
|
82
84
|
'MDCG_SUSTAIN_AUTOHEAL': {'bool_like': True, 'type': 'bool'},
|
|
83
85
|
'MDCG_SUSTAIN_BEAT_THREAD': {'bool_like': True, 'type': 'bool'},
|
|
84
86
|
'MDCG_TEMPORAL_GAMMA': {'min': 0.0, 'type': 'float'},
|
|
85
87
|
'MDCG_TENANT': {'type': 'str'},
|
|
86
|
-
'MDCG_TOOL_FACE': {'bool_like': True, 'type': 'bool'},
|
|
88
|
+
'MDCG_TOOL_FACE': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
87
89
|
'MDCG_UNIFY_QUERY': {'bool_like': True, 'true_literals': ['1'], 'type': 'bool'},
|
|
88
90
|
'MDCG_WRITELIMIT': {'bool_like': True, 'type': 'bool'},
|
|
89
91
|
'MDCG_WRITE_ATTEMPT_LEDGER': {'bool_like': True, 'type': 'bool'},
|
|
@@ -0,0 +1,127 @@
|
|
|
1
|
+
# -*- coding: utf-8 -*-
|
|
2
|
+
"""conflict_surface · 检索结果内的「显式矛盾对」暴露(红队缺口 B/C)
|
|
3
|
+
|
|
4
|
+
判据只收能给出证据的一类矛盾:
|
|
5
|
+
|
|
6
|
+
A 的 non_applicable_conditions(**整条**文本,长度 >= MIN_EVIDENCE)
|
|
7
|
+
作为连续子串出现在 B 的正文里 ⇒ A 排除了 B 所述情形(双向都查)。
|
|
8
|
+
|
|
9
|
+
边界(经反思单元审计后收窄,见 PR 讨论):
|
|
10
|
+
· 只比正文,不比声明面(B 的 negs 在反向检查里作为 A 的角色参与);
|
|
11
|
+
· 节点缺 id 时按**序号**区分,不因空串相等而整对跳过;
|
|
12
|
+
· 三类信号各自有上限,输出体积有界;
|
|
13
|
+
· 启发式(value_conflict 的正则退路)默认关,须显式 allow_heuristic=True。
|
|
14
|
+
|
|
15
|
+
纪律对齐:第 4 条(只报能给出证据的)、第 5 条(无证据即不报)。
|
|
16
|
+
"""
|
|
17
|
+
from __future__ import annotations
|
|
18
|
+
|
|
19
|
+
from . import untrusted as _ut
|
|
20
|
+
from . import value_conflict as _vc
|
|
21
|
+
|
|
22
|
+
MIN_EVIDENCE = 4 # 证据子串最短长度(防单字噪声)
|
|
23
|
+
MAX_PAIRS = 5 # 每类信号单次最多报几条
|
|
24
|
+
|
|
25
|
+
|
|
26
|
+
def _norm(text) -> str:
|
|
27
|
+
return " ".join(str(text or "").split())
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
def _neg_texts(fm) -> list:
|
|
31
|
+
raw = (fm or {}).get("non_applicable_conditions") or []
|
|
32
|
+
if isinstance(raw, str):
|
|
33
|
+
raw = [raw]
|
|
34
|
+
out = []
|
|
35
|
+
for x in raw:
|
|
36
|
+
s = _norm(x)
|
|
37
|
+
if len(s) >= MIN_EVIDENCE:
|
|
38
|
+
out.append(s)
|
|
39
|
+
return out
|
|
40
|
+
|
|
41
|
+
|
|
42
|
+
def _items(nodes):
|
|
43
|
+
"""[(key, id, negs, body)];id 缺失时用序号补齐键,避免空串互撞。"""
|
|
44
|
+
out = []
|
|
45
|
+
for i, n in enumerate(nodes or []):
|
|
46
|
+
if not isinstance(n, dict):
|
|
47
|
+
continue
|
|
48
|
+
nid = str(n.get("id") or "")
|
|
49
|
+
out.append(("#%d" % i, nid, _neg_texts(n.get("frontmatter")), _norm(n.get("content"))))
|
|
50
|
+
return out
|
|
51
|
+
|
|
52
|
+
|
|
53
|
+
def conflicts_among(nodes, limit: int = MAX_PAIRS) -> list:
|
|
54
|
+
"""双向查「A 的 neg 出现在 B 正文」;返回 [{a,b,evidence,kind}]。
|
|
55
|
+
|
|
56
|
+
只在 i<j 的组合上做两个方向,避免 (A,B) 与 (B,A) 两条循环把同一对报两遍。
|
|
57
|
+
"""
|
|
58
|
+
items = _items(nodes)
|
|
59
|
+
pairs = []
|
|
60
|
+
for i in range(len(items)):
|
|
61
|
+
_ka, a_id, a_negs, a_body = items[i]
|
|
62
|
+
for j in range(i + 1, len(items)):
|
|
63
|
+
_kb, b_id, b_negs, b_body = items[j]
|
|
64
|
+
if b_body:
|
|
65
|
+
for neg in a_negs:
|
|
66
|
+
if neg in b_body:
|
|
67
|
+
pairs.append({"a": a_id, "b": b_id, "evidence": neg,
|
|
68
|
+
"kind": "a_neg_in_b_body"})
|
|
69
|
+
break
|
|
70
|
+
if a_body:
|
|
71
|
+
for neg in b_negs:
|
|
72
|
+
if neg in a_body:
|
|
73
|
+
pairs.append({"a": b_id, "b": a_id, "evidence": neg,
|
|
74
|
+
"kind": "b_neg_in_a_body"})
|
|
75
|
+
break
|
|
76
|
+
if len(pairs) >= limit:
|
|
77
|
+
return pairs
|
|
78
|
+
return pairs
|
|
79
|
+
|
|
80
|
+
|
|
81
|
+
def value_pairs_among(nodes, limit: int = MAX_PAIRS, allow_heuristic: bool = False) -> list:
|
|
82
|
+
"""同 condition_space 下同名键不同值(默认只认声明面 value_slots)。"""
|
|
83
|
+
items = []
|
|
84
|
+
for i, n in enumerate(nodes or []):
|
|
85
|
+
if isinstance(n, dict):
|
|
86
|
+
items.append((str(n.get("id") or ""), n.get("frontmatter") or {}, _norm(n.get("content"))))
|
|
87
|
+
out = []
|
|
88
|
+
for i in range(len(items)):
|
|
89
|
+
for j in range(i + 1, len(items)):
|
|
90
|
+
a_id, a_fm, a_body = items[i]
|
|
91
|
+
b_id, b_fm, b_body = items[j]
|
|
92
|
+
hit = _vc.compare(a_fm, a_body, b_fm, b_body, allow_heuristic=allow_heuristic)
|
|
93
|
+
if hit:
|
|
94
|
+
hit["a"], hit["b"] = a_id, b_id
|
|
95
|
+
out.append(hit)
|
|
96
|
+
if len(out) >= limit:
|
|
97
|
+
return out
|
|
98
|
+
return out
|
|
99
|
+
|
|
100
|
+
|
|
101
|
+
def untrusted_among(nodes, limit: int = MAX_PAIRS) -> list:
|
|
102
|
+
"""检索结果里命中「伪装系统消息/指令式」的节点(只报案,不改内容)。"""
|
|
103
|
+
out = []
|
|
104
|
+
for n in nodes or []:
|
|
105
|
+
if not isinstance(n, dict):
|
|
106
|
+
continue
|
|
107
|
+
r = _ut.classify(n.get("content"))
|
|
108
|
+
if r.get("untrusted"):
|
|
109
|
+
out.append({"id": str(n.get("id") or ""), "markers": r.get("markers")})
|
|
110
|
+
if len(out) >= limit:
|
|
111
|
+
break
|
|
112
|
+
return out
|
|
113
|
+
|
|
114
|
+
|
|
115
|
+
def as_meta(nodes, limit: int = MAX_PAIRS, allow_heuristic: bool = False) -> dict:
|
|
116
|
+
"""把三类信号包成可直接并进 meta 的字典(空则不产生键;三类各自有上限)。"""
|
|
117
|
+
meta = {}
|
|
118
|
+
pairs = conflicts_among(nodes, limit=limit)
|
|
119
|
+
if pairs:
|
|
120
|
+
meta["conflicts_among"] = pairs
|
|
121
|
+
vp = value_pairs_among(nodes, limit=limit, allow_heuristic=allow_heuristic)
|
|
122
|
+
if vp:
|
|
123
|
+
meta["value_conflicts_among"] = vp
|
|
124
|
+
ut = untrusted_among(nodes, limit=limit)
|
|
125
|
+
if ut:
|
|
126
|
+
meta["untrusted_among"] = ut
|
|
127
|
+
return meta
|
package/md_cg/mcp_server.py
CHANGED
|
@@ -850,9 +850,9 @@ KERNEL_TOOLS = [
|
|
|
850
850
|
heal_interval=_p("number", "sustain start:巡检间隔秒(默认 300)"),
|
|
851
851
|
auto_heal=_p("boolean", "sustain start:巡检异常时自动修复(默认是)"),
|
|
852
852
|
scrub_interval=_p("number", "sustain start:自净间隔秒(默认 3600)"),
|
|
853
|
-
auto_scrub=_p("boolean", "sustain start
|
|
853
|
+
auto_scrub=_p("boolean", "sustain start:自净执行去污染(默认是;auto_scrub=false 可退回只巡检)"),
|
|
854
854
|
evolve_interval=_p("number", "sustain start:演化巡检间隔秒(默认 7200)"),
|
|
855
|
-
auto_evolve=_p("boolean", "sustain start
|
|
855
|
+
auto_evolve=_p("boolean", "sustain start:演化巡检自动落盘确定性动作(默认是)"
|
|
856
856
|
"(仅重要性重算;固化需 LLM 仍交人工。默认否)"),
|
|
857
857
|
allow_evolve=_p("boolean", "sustain heal:放行演化类修复"
|
|
858
858
|
"(确定性动作才执行;默认否)"),
|
|
@@ -2645,7 +2645,7 @@ def _cg_dispatch(cg, a):
|
|
|
2645
2645
|
recent_limit=int(a.get("limit") or 10),
|
|
2646
2646
|
session=a.get("session"),
|
|
2647
2647
|
validity=a.get("validity"),
|
|
2648
|
-
**({"paths": ("bm25",)} if a.get("bm25") else {}), **_tkw)
|
|
2648
|
+
**({"paths": ("bm25",)} if a.get("bm25", True) and hasattr(cg, "search_rrf") else {}), **_tkw)
|
|
2649
2649
|
if _kwarn and isinstance(_out, dict):
|
|
2650
2650
|
_out = dict(_out)
|
|
2651
2651
|
_out["warning"] = _kwarn
|
|
@@ -2654,7 +2654,7 @@ def _cg_dispatch(cg, a):
|
|
|
2654
2654
|
search_args = dict(layer=a.get("layer"), k=_int_arg(a, "k", 20),
|
|
2655
2655
|
context=a.get("context"), session=a.get("session"),
|
|
2656
2656
|
validity=a.get("validity"), **_tkw)
|
|
2657
|
-
if a.get("bm25"):
|
|
2657
|
+
if a.get("bm25", True) and hasattr(cg, "search_rrf"):
|
|
2658
2658
|
ranked, meta = cg.search_rrf(q, paths=("bm25",), **search_args)
|
|
2659
2659
|
res = [(n, s, qual) for n, s, qual, _provenance in ranked]
|
|
2660
2660
|
else:
|
|
@@ -3769,6 +3769,12 @@ _MDCG_TOOL_OP = {
|
|
|
3769
3769
|
"mdcg_health": "read",
|
|
3770
3770
|
"mdcg_whoami": "read", "mdcg_watermarks": "read",
|
|
3771
3771
|
"mdcg_whitebox": "read",
|
|
3772
|
+
# issue #578 / dsh-bounty#435:本入口**不在这张表里** ⇒ _dispatch 取
|
|
3773
|
+
# `_MDCG_OP_REQUIRE.get(name)` 得 None ⇒ 跳过 require_op。而它回吐的
|
|
3774
|
+
# `principal.as_dict()` 与受 read 闸管着的 `mdcg_whoami` 是**同一份**身份清单
|
|
3775
|
+
# (含 token_id / session / ops_allow),外加库根绝对路径 ⇒ 只读访客与
|
|
3776
|
+
# 「被 --ops-allow 收窄到不含 read」的令牌都能拿到。闸词对齐 whoami 的 read。
|
|
3777
|
+
"mdcg_service_info": "read",
|
|
3772
3778
|
# 治理/裁决面(仅高权角色域含 verify/review/protect/evolution)
|
|
3773
3779
|
"mdcg_verify": "verify",
|
|
3774
3780
|
# N228(2026-09-28):写保护面此前被映射到**粗粒度 write**,而 protect 在
|
|
@@ -4131,7 +4137,10 @@ def _dispatch(cg, name, args):
|
|
|
4131
4137
|
return out
|
|
4132
4138
|
|
|
4133
4139
|
if name == "mdcg_recall":
|
|
4134
|
-
|
|
4140
|
+
# recall 面保持词法默认:其「缺省六路」与 sum/max 融合口径都是**已登记契约**
|
|
4141
|
+
# (test_time_core_lint G8b/G8c、test_p3 A2)。BM25 默认开只落在检索面
|
|
4142
|
+
# (mdcg_search / cg(op=read) 回退),recall 仍 opt-in。
|
|
4143
|
+
use_bm25 = bool(a.get("bm25")) and hasattr(cg, "search_rrf")
|
|
4135
4144
|
use_fuzzy = bool(a.get("fuzzy"))
|
|
4136
4145
|
use_semantic = bool(a.get("semantic"))
|
|
4137
4146
|
use_goal = bool(a.get("goal_path") or a.get("goal"))
|
|
@@ -4171,7 +4180,10 @@ def _dispatch(cg, name, args):
|
|
|
4171
4180
|
# 默认单条上限从 mdcos 取(该模块只在 main() 里惰性导入,模块级没有名字,
|
|
4172
4181
|
# 直接引用 mdcos.DEFAULT_MAX_ITEM_TOKENS 会 NameError —— 故此处按需导入)。
|
|
4173
4182
|
from .mdcos import DEFAULT_MAX_ITEM_TOKENS as _DEFAULT_MAX_ITEM
|
|
4174
|
-
|
|
4183
|
+
# 注入面强制闸(使用者裁定 2026-10-21「未验证、未信任的信息本来就不应该执行」):
|
|
4184
|
+
# recall 是**进上下文**的通路,故此处把带 untrusted 标记的条目剔出 pack,
|
|
4185
|
+
# 并在 meta 记数。检索面(mdcg_search)仍可查见它们——查得到但不进提示词。
|
|
4186
|
+
_pack = cg.recall(a.get("query", ""),
|
|
4175
4187
|
# `is not None`:显式 0 不再被吞成默认 1200(2026-09-24 修复)
|
|
4176
4188
|
budget_tokens=int(a["budget_tokens"])
|
|
4177
4189
|
if a.get("budget_tokens") is not None else 1200,
|
|
@@ -4188,9 +4200,76 @@ def _dispatch(cg, name, args):
|
|
|
4188
4200
|
query_expand=_make_query_expand(a.get("expand")),
|
|
4189
4201
|
session=a.get("session"),
|
|
4190
4202
|
validity=a.get("validity"))
|
|
4203
|
+
# 注入面强制闸(使用者裁定 2026-10-21「未验证、未信任的信息本来就不应该执行」):
|
|
4204
|
+
# recall 是进上下文的通路,故把带 untrusted 标记的条目剔出 pack 并记数;
|
|
4205
|
+
# 检索面(mdcg_search)仍可查见——查得到,但不进提示词。
|
|
4206
|
+
try:
|
|
4207
|
+
_items = _pack.get("pack") if isinstance(_pack, dict) else None
|
|
4208
|
+
if isinstance(_items, list):
|
|
4209
|
+
_kept, _drop = [], 0
|
|
4210
|
+
for _it in _items:
|
|
4211
|
+
_fm = {}
|
|
4212
|
+
if isinstance(_it, dict):
|
|
4213
|
+
_node = _it.get("node") if isinstance(_it.get("node"), dict) else _it
|
|
4214
|
+
_fm = _node.get("frontmatter") or {}
|
|
4215
|
+
if _fm.get("untrusted"):
|
|
4216
|
+
_drop += 1
|
|
4217
|
+
continue
|
|
4218
|
+
_kept.append(_it)
|
|
4219
|
+
if _drop:
|
|
4220
|
+
_pack["pack"] = _kept
|
|
4221
|
+
_pack.setdefault("meta", {})["untrusted_excluded"] = _drop
|
|
4222
|
+
except Exception:
|
|
4223
|
+
pass
|
|
4224
|
+
return _pack
|
|
4191
4225
|
|
|
4192
4226
|
if name == "mdcg_search":
|
|
4193
4227
|
from . import refindex
|
|
4228
|
+
# ---- query 级热缓存(默认面接线,2026-10-21)----
|
|
4229
|
+
# 热缓存的读写在 search_rrf 内,而默认检索面走 cg.search(本分支)——
|
|
4230
|
+
# 故此处补一层工具级缓存:键按 hotcache 既有口径 + 全部入参 + 环境开关 + 索引代际。
|
|
4231
|
+
# 命中直接返回上一次的**整个工具返回体**(形状与未命中一致);
|
|
4232
|
+
# 显式 node_id / op 的调用一律绕过(那是"读某条",不是"查一次")。
|
|
4233
|
+
from . import hotcache as _hc
|
|
4234
|
+
# 工具层用**独立实例**:与 search_rrf 共用同一 _queries 字典会让两种值形状
|
|
4235
|
+
# 共处一个键空间——一旦键相等就是「静默错答/解包炸裂」(蜂巢反思单元 P0)。
|
|
4236
|
+
if not hasattr(cg, "_hotcache_tool"):
|
|
4237
|
+
cg._hotcache_tool = _hc.HotCache()
|
|
4238
|
+
# validity=True 的查询需要时钟窗,工具层不带窗口语义 ⇒ 一律不缓存(P1)。
|
|
4239
|
+
_hcobj = None if ("node_id" in a or a.get("op") or a.get("validity")) else cg._hotcache_tool
|
|
4240
|
+
_ckw = None
|
|
4241
|
+
if _hcobj is not None:
|
|
4242
|
+
try:
|
|
4243
|
+
_extra = {k: v for k, v in a.items() if k not in ("query", "bm25")}
|
|
4244
|
+
_extra["env_switch"] = _hc.env_switch_key()
|
|
4245
|
+
if hasattr(cg, "_index_generation"):
|
|
4246
|
+
_extra["index_gen"] = cg._index_generation()
|
|
4247
|
+
# 安全自检(2026-10-21):检索结果按 **调用者身份** 过滤
|
|
4248
|
+
# (MdCGSecure._candidates → _readable 按 session/密级判可见),
|
|
4249
|
+
# 故缓存键必须含**有效身份**,否则高密级/他会话的结果会被低密级命中。
|
|
4250
|
+
# 口径:principal(tenant/actor/clearance) + 有效 session(cg.session 覆盖入参)。
|
|
4251
|
+
_pr = getattr(cg, "principal", None)
|
|
4252
|
+
_eff_sess = getattr(cg, "session", None) or a.get("session")
|
|
4253
|
+
_extra["identity"] = (
|
|
4254
|
+
getattr(_pr, "tenant", None), getattr(_pr, "actor", None),
|
|
4255
|
+
getattr(_pr, "clearance", None), _eff_sess)
|
|
4256
|
+
# 键里直接用原始 k(不调 _int_arg:那是「六处委托单点」的守卫计数面,
|
|
4257
|
+
# 多一处调用会被 test_read_face_input_gates M1-B2 判成第二份实现)。
|
|
4258
|
+
_ckw = dict(k=a.get("k"), layer=a.get("layer"),
|
|
4259
|
+
session=a.get("session"), validity=a.get("validity"),
|
|
4260
|
+
extra=_extra or None)
|
|
4261
|
+
_hit = _hcobj.get_query(a.get("query", ""), **_ckw)
|
|
4262
|
+
except Exception:
|
|
4263
|
+
_hit = None
|
|
4264
|
+
if _hit is not None:
|
|
4265
|
+
import copy as _copy
|
|
4266
|
+
_rows, _tag = (_hit if isinstance(_hit, tuple) and len(_hit) == 2 else (None, {}))
|
|
4267
|
+
if (isinstance(_tag, dict) and _tag.get("surface") == "mdcg_search"
|
|
4268
|
+
and _tag.get("v") == 1 and isinstance(_rows, list) and len(_rows) == 1
|
|
4269
|
+
and isinstance(_rows[0], dict)):
|
|
4270
|
+
# 深拷贝:hotcache 的 _copy_row 只浅一层,_out["results"] 列表会共享
|
|
4271
|
+
return _copy.deepcopy(_rows[0])
|
|
4272
|
+
_hit = None # 形状/版本不符 ⇒ 当未命中,不猜
|
|
4194
4273
|
res, meta = cg.search(a.get("query", ""), layer=a.get("layer"),
|
|
4195
4274
|
k=_int_arg(a, "k", 20), context=a.get("context"),
|
|
4196
4275
|
roles=tuple(a["roles"]) if a.get("roles") else None,
|
|
@@ -4198,10 +4277,28 @@ def _dispatch(cg, name, args):
|
|
|
4198
4277
|
pools=a.get("pools"),
|
|
4199
4278
|
session=a.get("session"),
|
|
4200
4279
|
validity=a.get("validity"))
|
|
4201
|
-
|
|
4280
|
+
from . import conflict_surface
|
|
4281
|
+
_cf = conflict_surface.as_meta([n for n, _s, _q in res])
|
|
4282
|
+
if _cf:
|
|
4283
|
+
meta = dict(meta or {})
|
|
4284
|
+
meta.update(_cf)
|
|
4285
|
+
_out = {"meta": meta,
|
|
4202
4286
|
"results": [{"node": _node_view(n), "score": s, "state": q.get("state"),
|
|
4203
|
-
"reason": q.get("reason"),
|
|
4287
|
+
"reason": q.get("reason"),
|
|
4288
|
+
# 未信任标记透出到结果行:上层据此把它当数据、不当指令
|
|
4289
|
+
"untrusted": bool((n.get("frontmatter") or {}).get("untrusted")),
|
|
4290
|
+
**refindex.ref_fields(n)}
|
|
4204
4291
|
for n, s, q in res]}
|
|
4292
|
+
if _hcobj is not None and _ckw is not None:
|
|
4293
|
+
try:
|
|
4294
|
+
# 写入时深拷贝:否则**未命中那一次**返回给调用方的对象与缓存同一份,
|
|
4295
|
+
# 调用方原地改(append/改 score)就污染了缓存(蜂巢验证单元指出 _copy_row 只浅一层)。
|
|
4296
|
+
import copy as _copy2
|
|
4297
|
+
_hcobj.put_query(a.get("query", ""), [_copy2.deepcopy(_out)],
|
|
4298
|
+
{"surface": "mdcg_search", "v": 1}, **_ckw)
|
|
4299
|
+
except Exception:
|
|
4300
|
+
pass
|
|
4301
|
+
return _out
|
|
4205
4302
|
|
|
4206
4303
|
if name == "mdcg_get":
|
|
4207
4304
|
_nid = a.get("node_id", "")
|
package/md_cg/mdcg.py
CHANGED
|
@@ -1972,7 +1972,37 @@ class MdCG:
|
|
|
1972
1972
|
# 「.md 计数相等」,两套判据对同一个问题(「这份快照可信吗」)各答
|
|
1973
1973
|
# 各的,且 compact 还会把当前指纹盖到陈旧快照上把指纹判据一起抹掉。
|
|
1974
1974
|
if not self._snapshot_trusted(idx):
|
|
1975
|
-
|
|
1975
|
+
# 2026-10-21 治本(写/读超时根因):目录 mtime 指纹会在任意合作写入后
|
|
1976
|
+
# 失效(新文件落进目录即改 mtime),失效后原实现一律回退全库扫描——
|
|
1977
|
+
# 真实库 31,688 节点实测 13.4 秒/次冷启动,这才是 MCP 读写超时来源
|
|
1978
|
+
# (写入逻辑本身只要 47~53ms)。
|
|
1979
|
+
# 指纹只回答「目录树是否动过」,不判内容;合作写入都走分片日志。
|
|
1980
|
+
# 故先重放日志,再用 _covers_disk(walk 目录、只比路径集合,0.08 秒)
|
|
1981
|
+
# 验证「索引路径集合 == 盘面 .md 集合」:成立则采纳;不成立(改名/换目录/
|
|
1982
|
+
# 外部删加)仍回退全库扫描,正确性不降级。
|
|
1983
|
+
_probe = {"schema": SCHEMA, "nodes": dict(idx.get("nodes") or {}), "buckets": {}}
|
|
1984
|
+
try:
|
|
1985
|
+
self._apply_log(_probe["nodes"])
|
|
1986
|
+
# 只认「**纯新增**」这一种变化:合作写入(add)在日志里留下**新 path**,
|
|
1987
|
+
# 而改名/对调/换目录是**盘面行为**、日志里没有对应新 path——它们会把
|
|
1988
|
+
# 目录 mtime 改掉却让路径集合保持不变(_covers_disk 必假阳)。
|
|
1989
|
+
# 判据:发生变化的目录,必须全都在「日志新增 path 的所在目录」里。
|
|
1990
|
+
_log_dirs = set()
|
|
1991
|
+
for _rec in ShardedLog.read_all(self.index_log_dir):
|
|
1992
|
+
_e = _rec.get("e") if isinstance(_rec, dict) else None
|
|
1993
|
+
_path = _e.get("path") if isinstance(_e, dict) else None
|
|
1994
|
+
if _path:
|
|
1995
|
+
_log_dirs.add(os.path.dirname(str(_path).replace("\\", "/")))
|
|
1996
|
+
_cur_fp = self._dir_fingerprint()
|
|
1997
|
+
_old_fp = idx.get("_fingerprint") or {}
|
|
1998
|
+
_changed = {d for d, m in _cur_fp.items() if _old_fp.get(d) != m}
|
|
1999
|
+
_changed |= {d for d in _old_fp if d not in _cur_fp}
|
|
2000
|
+
if self._covers_disk(_probe["nodes"]) and _changed <= _log_dirs:
|
|
2001
|
+
idx = _probe
|
|
2002
|
+
else:
|
|
2003
|
+
idx = None
|
|
2004
|
+
except Exception:
|
|
2005
|
+
idx = None
|
|
1976
2006
|
if idx is None:
|
|
1977
2007
|
idx = {"schema": SCHEMA, "nodes": self._scan_nodes(), "buckets": {}}
|
|
1978
2008
|
self._apply_log(idx["nodes"])
|
|
@@ -2149,6 +2179,23 @@ class MdCG:
|
|
|
2149
2179
|
`self._dirty_replay_superseded` 留下条数读数。见证拿不到的条目照旧重放
|
|
2150
2180
|
——「本实例未落盘写入不因重载从检索侧消失」这条不变量一并保留。
|
|
2151
2181
|
"""
|
|
2182
|
+
# 2026-10-21:这是**应急闸**,默认关闭(默认 0)。
|
|
2183
|
+
# 背景:本方法每次读面入口都会探测签名,而他进程每次 flush 都改变签名 ⇒
|
|
2184
|
+
# 长期写方在跑的库里,读侧几乎每次查询都判「代际变化」并整库重载
|
|
2185
|
+
# (真实库 31k 节点实测 7.0 秒/次)。真正的修法是让重载变便宜
|
|
2186
|
+
# (_load_index 的快照+日志+目录判据),时间窗只是兜底手段。
|
|
2187
|
+
# 注意:窗口放在签名探测**之前**,窗内连探测都不做——它会同时挡住
|
|
2188
|
+
# 「重载」,故默认必须为 0,否则跨进程写入可见性被延迟(N195/N212 等守卫会红)。
|
|
2189
|
+
try:
|
|
2190
|
+
_min_iv = float(os.environ.get("MDCG_RELOAD_MIN_INTERVAL", "0"))
|
|
2191
|
+
except ValueError:
|
|
2192
|
+
_min_iv = 0.0
|
|
2193
|
+
if _min_iv > 0:
|
|
2194
|
+
_last = getattr(self, "_sig_probe_ts", 0.0)
|
|
2195
|
+
_now = time.time()
|
|
2196
|
+
if _now - _last < _min_iv:
|
|
2197
|
+
return False
|
|
2198
|
+
self._sig_probe_ts = _now
|
|
2152
2199
|
sig = self._index_signature()
|
|
2153
2200
|
if sig == self._index_sig:
|
|
2154
2201
|
return False
|
|
@@ -3039,6 +3086,17 @@ class MdCG:
|
|
|
3039
3086
|
"evidence_count": 0, "positive_evidence": 0, "negative_evidence": 0,
|
|
3040
3087
|
}
|
|
3041
3088
|
fm.update(extra)
|
|
3089
|
+
# 写入面信任标记(使用者裁定 2026-10-21:「未验证、未信任的信息本来就不应该执行」):
|
|
3090
|
+
# 文本里出现伪装系统消息/指令式短语 ⇒ 落 untrusted 标记 + 命中词,供检索面与
|
|
3091
|
+
# 上层(注入前)据此把它当**数据**而不是指令;不改正文(数据无损)。
|
|
3092
|
+
try:
|
|
3093
|
+
from . import untrusted as _un
|
|
3094
|
+
_u = _un.classify(content)
|
|
3095
|
+
if _u.get("untrusted"):
|
|
3096
|
+
fm["untrusted"] = True
|
|
3097
|
+
fm["untrusted_markers"] = list(_u.get("markers") or [])
|
|
3098
|
+
except Exception:
|
|
3099
|
+
pass
|
|
3042
3100
|
# 相对时间锚定(MDCG_TIME_ANCHOR=1,缺省关 ⇒ 零调用、fm 逐位不变):无显式 temporal
|
|
3043
3101
|
# 时把正文里的时间表达以 created_at 为参照归一成派生字段 temporal_anchor;解析不了不写。
|
|
3044
3102
|
if _timeanchor.enabled():
|
|
@@ -5436,7 +5494,7 @@ class MdCG:
|
|
|
5436
5494
|
# 阶段 1 大域收敛结果 + 完整打分明细(白箱可审计)
|
|
5437
5495
|
"big_domain": big_domain,
|
|
5438
5496
|
"big_domain_scores": big_scores,
|
|
5439
|
-
# reach 收敛层审计(2026-09-18):reach=converged/off/short_term
|
|
5497
|
+
# reach 收敛层审计(2026-09-18):reach=converged/off/index_unavailable/hash_incomplete/no_tokens/empty_converge/fresh_overflow(#611 起不再产生 short_term)
|
|
5440
5498
|
# 与 reach_seed(倒排命中数)/reach_diffused(图扩散增量)——
|
|
5441
5499
|
# A3「触碰量随查询变化」的观测面(scanned 只报读盘数)
|
|
5442
5500
|
"reach": stat.get("reach"),
|
package/md_cg/mdcos.py
CHANGED
|
@@ -1023,6 +1023,11 @@ class MdCGOS(MdCG):
|
|
|
1023
1023
|
hc = _hc.get(self)
|
|
1024
1024
|
if hc is not None and node_id:
|
|
1025
1025
|
hc.invalidate_node(node_id) # 节点级 O(1),非 invalidate_node_and_queries
|
|
1026
|
+
# 工具层 query 缓存(mcp_server.mdcg_search)是**独立实例**,主缓存失效管不到它;
|
|
1027
|
+
# 索引一突变就整清——否则 forget 之后 TTL 内还能被同参 query 读回正文。
|
|
1028
|
+
_tool_hc = getattr(self, "_hotcache_tool", None)
|
|
1029
|
+
if _tool_hc is not None:
|
|
1030
|
+
_tool_hc.clear()
|
|
1026
1031
|
except Exception: # noqa: BLE001
|
|
1027
1032
|
pass
|
|
1028
1033
|
return self._index_gen
|
|
@@ -1362,6 +1367,54 @@ class MdCGOS(MdCG):
|
|
|
1362
1367
|
if gates:
|
|
1363
1368
|
stat["gates"] = gates
|
|
1364
1369
|
neg_coverage = self._neg_coverage(terms) if include_neg else []
|
|
1370
|
+
# ---- 融合统一(2026-10-21):BM25 并入默认检索阶梯的全局阶段 ----
|
|
1371
|
+
# 背景:默认读面走本函数(词法打分),BM25 只在 search_rrf 里作一路;
|
|
1372
|
+
# 使用者裁定「BM25+ 实测提精度,应当默认使用」。此处把 BM25 打分接到
|
|
1373
|
+
# 全局阶段(T1' reach / T2 / T3)的 scored 上——桶阶段(T0/T1)保持原口径,
|
|
1374
|
+
# 因为它们本就是「域桶内精确命中」,改它会让桶优先级语义漂移。
|
|
1375
|
+
# 失败一律回落原 _score(信息不足不得中断检索)。
|
|
1376
|
+
_bm_scored = None
|
|
1377
|
+
# 2026-10-21 回归:默认开时实测 19~27 秒/查询(真实库 26k 候选),
|
|
1378
|
+
# 而 bm25.search 直接调用第 2 次只需 0.02s ⇒ 每次 search 都重建块布局(
|
|
1379
|
+
# entries 对象每轮都新建,缓存键命中不了)。在修好这条缓存前,
|
|
1380
|
+
# 默认关(opt-in),避免检索面 20x 回退。
|
|
1381
|
+
if os.environ.get("MDCG_BM25", "0") == "1":
|
|
1382
|
+
try:
|
|
1383
|
+
from . import bm25 as _bm
|
|
1384
|
+
# 用**独立 stat**:bm25.search 会累加 scanned/build 计数,
|
|
1385
|
+
# 记进主 stat 会把「候选面收敛」的读数污染成「全库扫描」。
|
|
1386
|
+
_bm_scored, _bm_meta = _bm.search(self, q, entries, {"scanned": 0, "query": q})
|
|
1387
|
+
except Exception:
|
|
1388
|
+
_bm_scored = None
|
|
1389
|
+
_bm_by_path = {}
|
|
1390
|
+
for _d, _sc in (_bm_scored or ()):
|
|
1391
|
+
try:
|
|
1392
|
+
_bm_by_path[_d.get("path")] = _sc
|
|
1393
|
+
except Exception:
|
|
1394
|
+
break
|
|
1395
|
+
|
|
1396
|
+
def _path_of(d):
|
|
1397
|
+
"""docs 元素可能是 dict,也可能是 (doc, fm, content) 元组——两形都取 path。"""
|
|
1398
|
+
if isinstance(d, tuple):
|
|
1399
|
+
d = d[0] if d else {}
|
|
1400
|
+
return d.get("path") if isinstance(d, dict) else None
|
|
1401
|
+
|
|
1402
|
+
def _bm_for(docs):
|
|
1403
|
+
"""把 BM25 分数映射到本阶段的 docs 上(缺失按 0,保持长度一致)。
|
|
1404
|
+
|
|
1405
|
+
docs 元素若为 (doc, fm, content) 元组形态则**不接管**——下游
|
|
1406
|
+
cut_by_relevance 期望 scored[i][0] 是 dict;形态不符时一律回落 _score。
|
|
1407
|
+
"""
|
|
1408
|
+
if not _bm_by_path:
|
|
1409
|
+
return None
|
|
1410
|
+
if any(isinstance(d, tuple) for d in docs):
|
|
1411
|
+
return None
|
|
1412
|
+
return [(d, _bm_by_path.get(_path_of(d), 0.0)) for d in docs]
|
|
1413
|
+
|
|
1414
|
+
def _score_bm(docs, *a, **kw):
|
|
1415
|
+
"""全局阶段优先用 BM25 分数;无 BM25 结果时回落原 _score。"""
|
|
1416
|
+
_bm = _bm_for(docs)
|
|
1417
|
+
return _bm if _bm is not None else self._score(docs, *a, **kw)
|
|
1365
1418
|
|
|
1366
1419
|
# 生效条件:scored 为 None 时对 docs 调 self._score(docs, q, qb, pool_cfg),非 None 时直接以之为打分结果;分数 >0 的条数达到(search 作用域内的)min_results 就返回 self._emit(scored, k, tier, stat, route_bucket, record, len(docs), ...),否则返回 None。
|
|
1367
1420
|
def try_stage(docs, tier, scored=None):
|
|
@@ -1371,7 +1424,7 @@ class MdCGOS(MdCG):
|
|
|
1371
1424
|
# try_stage(mdcg.py P2-4,批次 31)同模板;_score 在单次 search 内
|
|
1372
1425
|
# 是纯函数(docs/q/qb/pool_cfg 均不变),透传不改变任何分值与输出。
|
|
1373
1426
|
if scored is None:
|
|
1374
|
-
scored =
|
|
1427
|
+
scored = _score_bm(docs, q, qb, pool_cfg)
|
|
1375
1428
|
valid = sum(1 for _, s in scored if s > 0)
|
|
1376
1429
|
if valid >= min_results:
|
|
1377
1430
|
return self._emit(scored, k, tier, stat, route_bucket, record,
|
|
@@ -1442,7 +1495,7 @@ class MdCGOS(MdCG):
|
|
|
1442
1495
|
stat["cap"] = GLOBAL_CAP
|
|
1443
1496
|
# 与 T2 **无条件**同序调用(不可按 cap 短路:cut_by_relevance 还会写
|
|
1444
1497
|
# 池账 pool_taken/cands/lost,短路会让 meta.pools.taken 缺失 → test_p43(13) 红)
|
|
1445
|
-
scored_r =
|
|
1498
|
+
scored_r = _score_bm(hits_r, q, qb, pool_cfg)
|
|
1446
1499
|
hits_r, _rep = cut_by_relevance(hits_r, scored_r,
|
|
1447
1500
|
GLOBAL_CAP, pools=pool_cfg,
|
|
1448
1501
|
key_of=pooling.doc_key, stat=stat)
|
|
@@ -1476,7 +1529,7 @@ class MdCGOS(MdCG):
|
|
|
1476
1529
|
or (semantic_on() and d[1].get("semantic"))]
|
|
1477
1530
|
stat["pre_cap"] = len(hits)
|
|
1478
1531
|
stat["cap"] = GLOBAL_CAP
|
|
1479
|
-
scored_h =
|
|
1532
|
+
scored_h = _score_bm(hits, q, qb, pool_cfg)
|
|
1480
1533
|
hits, _rep = cut_by_relevance(hits, scored_h,
|
|
1481
1534
|
GLOBAL_CAP, pools=pool_cfg,
|
|
1482
1535
|
key_of=pooling.doc_key, stat=stat)
|
|
@@ -1489,7 +1542,7 @@ class MdCGOS(MdCG):
|
|
|
1489
1542
|
|
|
1490
1543
|
stat["pre_cap"] = len(docs_all)
|
|
1491
1544
|
stat["cap"] = GLOBAL_CAP
|
|
1492
|
-
scored_a =
|
|
1545
|
+
scored_a = _score_bm(docs_all, q, qb, pool_cfg)
|
|
1493
1546
|
picked, _rep = cut_by_relevance(docs_all, scored_a,
|
|
1494
1547
|
GLOBAL_CAP, pools=pool_cfg,
|
|
1495
1548
|
key_of=pooling.doc_key, stat=stat)
|
|
@@ -2884,7 +2937,8 @@ class MdCGOS(MdCG):
|
|
|
2884
2937
|
|
|
2885
2938
|
# 生效条件:stat inbox_log 失败(OSError)时把 _inbox_recs 置空列表、两扫描键置 -1 并返回它;size 与 mtime_ns 均与缓存键一致时直接返回 _inbox_recs(零解析);缓存键为 -1(首装)或当前 size 小于已扫描偏移时以 read_jsonl_tail(path, 0) 全量重建,否则以 read_jsonl_tail(offset) 结果 extend 增量并入;水位存 read_jsonl_tail 返回的已提交完整行偏移(end 为 None 时回落 stat size),mtime_ns 存读前 stat 值,返回 _inbox_recs(内部缓存,调用方只读不得变更)。
|
|
2886
2939
|
# 生效条件:node_id 为非空字符串时,遍历 inbox 全记录(`_inbox_records` 缓存)取
|
|
2887
|
-
# `rec["id"] == node_id` 且其 pid
|
|
2940
|
+
# `rec["id"] == node_id` 且其 pid **尚未被终态裁决关闭**的记录(`_pid_status` 缓存;
|
|
2941
|
+
# needs_reapproval 仍待再审批,与 _closed_pids/review_list 使用同一终态闭集),
|
|
2888
2942
|
# 按文件序返回最多 limit 个 pid;node_id 为空/无匹配返回空列表。只读、零写、不抛。
|
|
2889
2943
|
def pending_proposals(self, node_id, limit: int = 5) -> list:
|
|
2890
2944
|
"""该节点在审核队列里**待裁决**的提案 id(读面标注用,issue #561 第①项)。
|
|
@@ -2904,7 +2958,8 @@ class MdCGOS(MdCG):
|
|
|
2904
2958
|
if str(rec.get("id") or "") != node_id:
|
|
2905
2959
|
continue
|
|
2906
2960
|
pid = rec.get("pid")
|
|
2907
|
-
if not pid or pid
|
|
2961
|
+
if (not pid or (decided.get(pid) or {}).get("status")
|
|
2962
|
+
in TERMINAL_DECISION_STATUS):
|
|
2908
2963
|
continue
|
|
2909
2964
|
out.append(pid)
|
|
2910
2965
|
if len(out) >= int(limit or 5):
|