dsh-memu-memory 1.1.3 → 1.2.0

package/tests/fixtures/active-memory-ab/cases/13-clean-scale-5000.json

@@ -0,0 +1,137 @@
+{
+  "schemaVersion": 1,
+  "kind": "active-memory-ab-case",
+  "id": "clean-scale-5000",
+  "draft": true,
+  "notGold": true,
+  "gold": false,
+  "review": "pending",
+  "annotationSource": "synthetic-proxy",
+  "interference": "off",
+  "title": "规模梯度 5000 事件：越过 largeTranscriptEvents，薄回合不再吃饥饿豁免",
+  "annotationNotes": "合成代理标注。与 clean-large-thin 同证据、同标签、同 expect 形状，唯一变量是 padding.count，用于标定 largeTranscriptEvents 阈值。padding 是确定性工具轮询，文本完全相同，去重后只算 1 个工作单元。",
+  "snapshot": {
+    "domain": "coding",
+    "idleAtSnapshotEndMinutes": 20,
+    "lastInjectedAt": null,
+    "events": [
+      {
+        "type": "turn/start",
+        "at": "2026-10-30T02:00:00.000Z"
+      },
+      {
+        "type": "user/message",
+        "at": "2026-10-30T02:00:00.000Z",
+        "data": {
+          "source": {
+            "kind": "user"
+          },
+          "content": [
+            {
+              "type": "text",
+              "text": "先按计划把迁移脚本跑通，日志我看产出就行。"
+            }
+          ]
+        }
+      },
+      {
+        "type": "assistant/message",
+        "at": "2026-10-30T02:00:00.000Z",
+        "data": {
+          "message": {
+            "content": [
+              {
+                "type": "text",
+                "text": "开始跑。"
+              }
+            ]
+          }
+        }
+      },
+      {
+        "type": "step/end",
+        "at": "2026-10-30T02:00:00.000Z",
+        "step": 1
+      },
+      {
+        "type": "turn/end",
+        "at": "2026-10-30T02:00:30.000Z"
+      },
+      {
+        "type": "padding",
+        "at": "2026-10-30T02:01:00.000Z",
+        "startAt": "2026-10-30T02:01:00.000Z",
+        "count": 1247,
+        "intervalMs": 1000,
+        "tool": "read",
+        "text": "polling: no change"
+      },
+      {
+        "type": "turn/start",
+        "at": "2026-10-30T02:10:00.000Z"
+      },
+      {
+        "type": "user/message",
+        "at": "2026-10-30T02:10:00.000Z",
+        "data": {
+          "source": {
+            "kind": "user"
+          },
+          "content": [
+            {
+              "type": "text",
+              "text": "继续"
+            }
+          ]
+        }
+      },
+      {
+        "type": "assistant/message",
+        "at": "2026-10-30T02:10:00.000Z",
+        "data": {
+          "message": {
+            "content": [
+              {
+                "type": "text",
+                "text": "还在跑。"
+              }
+            ]
+          }
+        }
+      },
+      {
+        "type": "step/end",
+        "at": "2026-10-30T02:10:00.000Z",
+        "step": 1247
+      },
+      {
+        "type": "turn/end",
+        "at": "2026-10-30T02:10:00.000Z"
+      }
+    ]
+  },
+  "labels": {
+    "expected": [
+      {
+        "id": "mem-ab-scale-plan-scope",
+        "class": "decision",
+        "sourceSeq": 1,
+        "note": "迁移脚本先跑通，日志以产出为准"
+      },
+      {
+        "id": "mem-ab-scale-thin-tail",
+        "class": "promise",
+        "sourceSeq": 4994,
+        "note": "末尾薄回合（“继续”）本身没有新事实，用于量漏记"
+      }
+    ],
+    "mainProactive": [],
+    "userReminders": []
+  },
+  "expect": {
+    "maxBackgroundChecks": 0,
+    "blockedReasons": [
+      "large-transcript-thin"
+    ]
+  }
+}

package/tests/fixtures/active-memory-ab/cases/14-clean-scale-20000.json

@@ -0,0 +1,137 @@
+{
+  "schemaVersion": 1,
+  "kind": "active-memory-ab-case",
+  "id": "clean-scale-20000",
+  "draft": true,
+  "notGold": true,
+  "gold": false,
+  "review": "pending",
+  "annotationSource": "synthetic-proxy",
+  "interference": "off",
+  "title": "规模梯度 20000 事件：上限区，薄回合被拦且完整前缀种子成本最高",
+  "annotationNotes": "合成代理标注。与 clean-large-thin 同证据、同标签、同 expect 形状，唯一变量是 padding.count，用于标定 largeTranscriptEvents 阈值。padding 是确定性工具轮询，文本完全相同，去重后只算 1 个工作单元。",
+  "snapshot": {
+    "domain": "coding",
+    "idleAtSnapshotEndMinutes": 20,
+    "lastInjectedAt": null,
+    "events": [
+      {
+        "type": "turn/start",
+        "at": "2026-10-30T02:00:00.000Z"
+      },
+      {
+        "type": "user/message",
+        "at": "2026-10-30T02:00:00.000Z",
+        "data": {
+          "source": {
+            "kind": "user"
+          },
+          "content": [
+            {
+              "type": "text",
+              "text": "先按计划把迁移脚本跑通，日志我看产出就行。"
+            }
+          ]
+        }
+      },
+      {
+        "type": "assistant/message",
+        "at": "2026-10-30T02:00:00.000Z",
+        "data": {
+          "message": {
+            "content": [
+              {
+                "type": "text",
+                "text": "开始跑。"
+              }
+            ]
+          }
+        }
+      },
+      {
+        "type": "step/end",
+        "at": "2026-10-30T02:00:00.000Z",
+        "step": 1
+      },
+      {
+        "type": "turn/end",
+        "at": "2026-10-30T02:00:30.000Z"
+      },
+      {
+        "type": "padding",
+        "at": "2026-10-30T02:01:00.000Z",
+        "startAt": "2026-10-30T02:01:00.000Z",
+        "count": 4997,
+        "intervalMs": 1000,
+        "tool": "read",
+        "text": "polling: no change"
+      },
+      {
+        "type": "turn/start",
+        "at": "2026-10-30T02:10:00.000Z"
+      },
+      {
+        "type": "user/message",
+        "at": "2026-10-30T02:10:00.000Z",
+        "data": {
+          "source": {
+            "kind": "user"
+          },
+          "content": [
+            {
+              "type": "text",
+              "text": "继续"
+            }
+          ]
+        }
+      },
+      {
+        "type": "assistant/message",
+        "at": "2026-10-30T02:10:00.000Z",
+        "data": {
+          "message": {
+            "content": [
+              {
+                "type": "text",
+                "text": "还在跑。"
+              }
+            ]
+          }
+        }
+      },
+      {
+        "type": "step/end",
+        "at": "2026-10-30T02:10:00.000Z",
+        "step": 4997
+      },
+      {
+        "type": "turn/end",
+        "at": "2026-10-30T02:10:00.000Z"
+      }
+    ]
+  },
+  "labels": {
+    "expected": [
+      {
+        "id": "mem-ab-scale-plan-scope",
+        "class": "decision",
+        "sourceSeq": 1,
+        "note": "迁移脚本先跑通，日志以产出为准"
+      },
+      {
+        "id": "mem-ab-scale-thin-tail",
+        "class": "promise",
+        "sourceSeq": 19994,
+        "note": "末尾薄回合（“继续”）本身没有新事实，用于量漏记"
+      }
+    ],
+    "mainProactive": [],
+    "userReminders": []
+  },
+  "expect": {
+    "maxBackgroundChecks": 0,
+    "blockedReasons": [
+      "large-transcript-thin"
+    ]
+  }
+}

package/tests/fixtures/active-memory-ab/lead-ai-reviewed-reference.json

@@ -0,0 +1,191 @@
+{
+  "schemaVersion": 1,
+  "kind": "active-memory-ab-lead-reviewed-reference",
+  "status": "lead-ai-reviewed",
+  "gold": false,
+  "annotationSource": "lead-ai-reviewed",
+  "reviewer": "lead-ai",
+  "reviewedAt": "2026-10-30",
+  "sourceDataset": {
+    "status": "synthetic-proxy-not-gold",
+    "gold": false,
+    "source": "local fixture cases/ (unchanged; no synthetic label was promoted)"
+  },
+  "policy": {
+    "seedMode": "complete-contiguous-prefix",
+    "windowedSeed": false,
+    "heldCountsAsCovered": false,
+    "halfStep": "deferred without model scoring",
+    "noHumanTurn": "skipped",
+    "thinNoFact": "no_value only when normal completed check returns no candidates"
+  },
+  "cases": [
+    {
+      "id": "clean-dense-turn",
+      "expectedStatus": "completed",
+      "labels": [
+        {
+          "class": "decision",
+          "sourceSeq": 1,
+          "canonical": "snapshot-worker-does-not-block-main-session"
+        },
+        {
+          "class": "root-cause",
+          "sourceSeq": 1,
+          "canonical": "repeated-full-transcript-forking-drives-input-cost"
+        },
+        {
+          "class": "preference",
+          "sourceSeq": 5,
+          "canonical": "python-installation-is-confined-to-engine-venv"
+        },
+        {
+          "class": "promise",
+          "sourceSeq": 9,
+          "canonical": "synthetic-labels-never-become-gold-without-review"
+        }
+      ],
+      "notes": [
+        "The worker/root-cause decision is at seq 1; the venv constraint is at seq 5; the gold-gating promise is at seq 9."
+      ]
+    },
+    {
+      "id": "clean-thin-turn",
+      "expectedStatus": "completed",
+      "labels": [],
+      "outcome": "no_value",
+      "notes": [
+        "The two human turns contain no durable decision or preference; the fixture label marks an uncovered interval, not a memory."
+      ]
+    },
+    {
+      "id": "clean-starved-small",
+      "expectedStatus": "completed",
+      "labels": [
+        {
+          "class": "decision",
+          "sourceSeq": 1,
+          "canonical": "no_value-requires-normal-completed-empty-result"
+        }
+      ],
+      "notes": [
+        "The no_value rule is an explicit human decision at seq 1; idle/starvation does not itself create a memory."
+      ]
+    },
+    {
+      "id": "clean-large-thin",
+      "expectedStatus": "completed",
+      "labels": [
+        {
+          "class": "decision",
+          "sourceSeq": 1,
+          "canonical": "run-migration-plan-and-use-produced-logs"
+        }
+      ],
+      "notes": [
+        "The later “继续” at seq 7 adds no fact; the synthetic tail label is not a gold memory."
+      ]
+    },
+    {
+      "id": "clean-half-step",
+      "expectedStatus": "deferred",
+      "guardReason": "half-step",
+      "labels": [],
+      "notes": [
+        "The prefix contains an unpaired tool call and snapshot.pendingCheck=true; hard guard prevents scoring, writes, and cursor advance."
+      ]
+    },
+    {
+      "id": "clean-duplicate-reread",
+      "expectedStatus": "completed",
+      "labels": [
+        {
+          "class": "decision",
+          "sourceSeq": 1,
+          "canonical": "do-not-use-windowed-seed"
+        },
+        {
+          "class": "decision",
+          "sourceSeq": 8,
+          "canonical": "receipts-bind-snapshot-digest-and-scope-generation"
+        }
+      ],
+      "notes": [
+        "The credential requirement is stated by the human at seq 8; seq 7 is turn/start and cannot be the source."
+      ]
+    },
+    {
+      "id": "clean-recent-recall",
+      "expectedStatus": "completed",
+      "labels": [
+        {
+          "class": "decision",
+          "sourceSeq": 1,
+          "canonical": "suppress-checks-within-recall-cooldown"
+        }
+      ],
+      "notes": [
+        "The suppression rule is an explicit human instruction at seq 1."
+      ]
+    },
+    {
+      "id": "global-relevant",
+      "expectedStatus": "completed",
+      "labels": [
+        {
+          "class": "preference",
+          "sourceSeq": 1,
+          "canonical": "worker-writes-private-only-and-global-is-read-only"
+        }
+      ],
+      "notes": [
+        "Relevant global entries are read-only context and must not be copied into private writes."
+      ]
+    },
+    {
+      "id": "global-irrelevant-conflict",
+      "expectedStatus": "completed",
+      "labels": [
+        {
+          "class": "decision",
+          "sourceSeq": 1,
+          "canonical": "retain-only-one-backup-after-successful-switch",
+          "disposition": "held-conflict"
+        },
+        {
+          "class": "preference",
+          "sourceSeq": 6,
+          "canonical": "web-gui-port-is-9032",
+          "disposition": "covered"
+        }
+      ],
+      "notes": [
+        "The backup decision conflicts with the global three-backup entry, so it is held and does not count as covered; the port preference is independent."
+      ]
+    },
+    {
+      "id": "no-human-turn",
+      "expectedStatus": "skipped",
+      "guardReason": "no-human-turn",
+      "labels": [],
+      "notes": [
+        "The user-shaped message is plugin-generated (source.kind=plugin:dsh-memu-memory), so this case is not eligible for human-memory scoring."
+      ]
+    }
+  ],
+  "metrics": {
+    "cases": 10,
+    "completed": 7,
+    "deferred": 1,
+    "skipped": 1,
+    "normalNoValue": 1,
+    "labels": 13,
+    "heldLabels": 1,
+    "coveredLabels": 12
+  },
+  "caveats": [
+    "This lead-AI-reviewed reference is an independent review of local fixture evidence, not a gold set and not a claim that the live model produced these labels.",
+    "The 13 labels are non-exhaustive reference annotations; the live run produced no parseable candidates, so model effect and A/B accuracy remain unmeasured.",
+    "Raw prompts, responses, transcripts, and credentials are not stored."
+  ]
+}

package/tests/fixtures/active-memory-ab/manifest.json

@@ -0,0 +1,199 @@
+{
+  "schemaVersion": 1,
+  "kind": "active-memory-ab-manifest",
+  "name": "active-memory 金标 A/B 候选集（合成代理草稿，未人审）",
+  "status": "synthetic-proxy-not-gold",
+  "draft": true,
+  "notGold": true,
+  "gold": false,
+  "review": "pending",
+  "annotationPolicy": "所有标签都是合成代理（annotationSource=synthetic-proxy），按 seq 指向“若由人来标也应记”的事件；没有人工逐条确认，因此不是 gold set。人审候选与 gold gating 分开：只有 review 改为 human-reviewed 之后才可能进入金标。",
+  "axes": [
+    "off",
+    "relevant",
+    "irrelevant-conflict"
+  ],
+  "conditions": [
+    "legacy",
+    "guide",
+    "guide-weighted"
+  ],
+  "rows": [
+    "legacy",
+    "guide",
+    "guide-weighted:oracle",
+    "guide-weighted:no-value"
+  ],
+  "paddingGenerator": "ab-polling-padding-v1",
+  "realModelCalls": 0,
+  "cases": [
+    {
+      "id": "clean-dense-turn",
+      "file": "cases/01-clean-dense-turn.json",
+      "throughSeq": 22,
+      "eventCount": 23,
+      "snapshotSha256": "01b63a470444c849c93239c9b717f3b6049497398c05daaab193d6781093838c",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "密集回合：分数达标就检查，不等空闲"
+    },
+    {
+      "id": "clean-thin-turn",
+      "file": "cases/02-clean-thin-turn.json",
+      "throughSeq": 9,
+      "eventCount": 10,
+      "snapshotSha256": "73cfdd46eb46233ac5129a1a73e5a2a77be34a4bdf3e6d27dc5e182cc39495b7",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "薄回合：没有空闲窗口，也不该唤醒"
+    },
+    {
+      "id": "clean-starved-small",
+      "file": "cases/03-clean-starved-small.json",
+      "throughSeq": 9,
+      "eventCount": 10,
+      "snapshotSha256": "49145876868c032e291dd7964d63ceea6afd8481158bcaf81617561aded8cd82",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "小会话饥饿：空闲到阈值就无条件检查一次"
+    },
+    {
+      "id": "clean-large-thin",
+      "file": "cases/04-clean-large-thin.json",
+      "throughSeq": 2089,
+      "eventCount": 2090,
+      "snapshotSha256": "e828cde9d7ee4242a507c9af294dd0acd0a38df0fd087d42f65351dabb9c5cf1",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "大 transcript 薄回合：不吃饥饿豁免"
+    },
+    {
+      "id": "clean-half-step",
+      "file": "cases/05-clean-half-step.json",
+      "throughSeq": 3,
+      "eventCount": 4,
+      "snapshotSha256": "c6eb87ef676716ddad2e4f5fa32d985bd1c151f70516f72e91e38bd698358b96",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "半步边界：tool/call 没有结果就 deferred，不推进游标"
+    },
+    {
+      "id": "clean-duplicate-reread",
+      "file": "cases/06-clean-duplicate-reread.json",
+      "throughSeq": 13,
+      "eventCount": 14,
+      "snapshotSha256": "d48163e0314151f7ed6e57e267e3f2c4bdf2c7f9879b35e9f5d14d0deab401d0",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "重复读与已写条目：整段前缀重放会再次提议同一条"
+    },
+    {
+      "id": "clean-recent-recall",
+      "file": "cases/07-clean-recent-recall.json",
+      "throughSeq": 9,
+      "eventCount": 10,
+      "snapshotSha256": "2ea689efcda46021415affa143fe55e167b11ed6784dabadce18c27e1cf26e50",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "刚注入过记忆：抑制而不是检查"
+    },
+    {
+      "id": "global-relevant",
+      "file": "cases/08-global-relevant.json",
+      "throughSeq": 4,
+      "eventCount": 5,
+      "snapshotSha256": "0f4d97adf48488f69f3228a1fc7a746472fed18221f3502186e5e086f9934fb9",
+      "interference": "relevant",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "全局干扰（相关）：相关全局条目可以进上下文，但不许被复制进私有写"
+    },
+    {
+      "id": "global-irrelevant-conflict",
+      "file": "cases/09-global-irrelevant-conflict.json",
+      "throughSeq": 9,
+      "eventCount": 10,
+      "snapshotSha256": "0a1d5e3a73965d5efb523e967995036d5693be9afaf22bef9c03721bd6a73963",
+      "interference": "irrelevant-conflict",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "全局干扰（无关且冲突）：冲突条目只进 held，不进可召回覆盖"
+    },
+    {
+      "id": "no-human-turn",
+      "file": "cases/10-no-human-turn.json",
+      "throughSeq": 5,
+      "eventCount": 6,
+      "snapshotSha256": "851fa7965cbe2bb477c2e1c062cf5c13ba198c911cd1f85cb0885cee3d8e0899",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "没有真人回合：纯工具轮询不构成检查理由"
+    },
+    {
+      "id": "clean-scale-100",
+      "file": "cases/11-clean-scale-100.json",
+      "throughSeq": 97,
+      "eventCount": 98,
+      "snapshotSha256": "ec6180fcd444ab6f71345ff571da9ca79c8168f0a036d2e8902b6673e4ee34ce",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "规模梯度 100 事件：低于 largeTranscriptEvents，饥饿豁免仍然生效"
+    },
+    {
+      "id": "clean-scale-1000",
+      "file": "cases/12-clean-scale-1000.json",
+      "throughSeq": 997,
+      "eventCount": 998,
+      "snapshotSha256": "04510b40d08eb03e3b1d04715cff1554324b5f1769bb97c95f78b22e82496f1a",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "规模梯度 1000 事件：仍在阈值下，饥饿豁免生效"
+    },
+    {
+      "id": "clean-scale-5000",
+      "file": "cases/13-clean-scale-5000.json",
+      "throughSeq": 4997,
+      "eventCount": 4998,
+      "snapshotSha256": "dc9095796c6cc122d7b314f59db2efa1963d86226837b7b5234c57ebb671e68f",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "规模梯度 5000 事件：越过 largeTranscriptEvents，薄回合不再吃饥饿豁免"
+    },
+    {
+      "id": "clean-scale-20000",
+      "file": "cases/14-clean-scale-20000.json",
+      "throughSeq": 19997,
+      "eventCount": 19998,
+      "snapshotSha256": "400e916841df1318ea6819bd73b840039b91157b955c9968d08e660c32fc85e4",
+      "interference": "off",
+      "annotationSource": "synthetic-proxy",
+      "review": "pending",
+      "gold": false,
+      "title": "规模梯度 20000 事件：上限区，薄回合被拦且完整前缀种子成本最高"
+    }
+  ]
+}