feat(execution): the session continuation surface, and the arms' checks read data

agent-context.yaml

@@ -194,7 +194,16 @@ routes:
       - src/integration/tool-contract.ts
     owners:
       - docs/design/design.md
-    tests: [tests/integration/agent-surface.test.ts, tests/integration/chain-projection.test.ts, tests/integration/config.test.ts, tests/integration/controller-channel.test.ts, tests/integration/evidence.test.ts, tests/integration/lab-capabilities.test.ts, tests/integration/tool-contract.test.ts]
+    tests:
+      [
+        tests/integration/agent-surface.test.ts,
+        tests/integration/chain-projection.test.ts,
+        tests/integration/config.test.ts,
+        tests/integration/controller-channel.test.ts,
+        tests/integration/evidence.test.ts,
+        tests/integration/lab-capabilities.test.ts,
+        tests/integration/tool-contract.test.ts,
+      ]
     verify:
       blocking: [check, test:product, build]
       advisory: []
@@ -203,12 +212,15 @@ routes:
     paths:
       - src/integration/ooo-board.ts
       - src/integration/ooo-candidate.ts
+      - src/integration/ooo-dispatch.ts
       - src/integration/ooo-execution.ts
       - src/integration/ooo-fusion-plan.ts
       - src/integration/ooo-mutation.ts
       - src/integration/ooo-patch.ts
       - src/integration/check-ticket.ts
       - src/integration/check-runner.ts
+      - src/integration/ooo-session-facts.ts
+      - src/integration/ooo-session-mechanism.ts
       - src/integration/task-advisers.ts
       - src/integration/task-coordinator.ts
       - src/integration/task-semantics-interleavings.ts
@@ -235,7 +247,12 @@ routes:
     owners:
       - docs/design/design.md
       - docs/design/tiered-disclosure-design.md
-    tests: [tests/integration/leaf-summarizer.test.ts, tests/integration/summary-drain.test.ts, tests/core/store/node-summaries.test.ts]
+    tests:
+      [
+        tests/integration/leaf-summarizer.test.ts,
+        tests/integration/summary-drain.test.ts,
+        tests/core/store/node-summaries.test.ts,
+      ]
     verify:
       blocking: [check, test:product, build]
       advisory: []

docs/decisions/implemented/2026-08-29-session-active-graph-runtime.md

@@ -0,0 +1,122 @@
+# Session-owned Active Graph runtime
+
+[中文](2026-08-29-session-active-graph-runtime.zh-CN.md)
+
+**Status:** implemented
+**Approved:** unrecorded
+**Relates to:** [Session AG runtime blueprint](../../design/session-active-graph-runtime-design.md)
+
+## Problem
+
+NMG used "Active Graph" for a query-scoped retrieval result while the Pi adapter separately kept a
+flat `SessionRuntimeAg` for recent tool state. The two structures served different parts of working
+memory, duplicated lifecycle logic, and left the task-state term in `AG_t = Project(STG, LTG, q_t,
+task_t)` without a stable runtime owner. `activeGraphId` also named the retrieval trace, so a mutable
+working graph and an immutable exposure record could not be distinguished.
+
+Hierarchical Activation (HA) and the Memory-Graph Reasoner (MGR) already provided candidate
+activation and graph traversal primitives, but they remained beside the runtime rather than operating
+on one bounded working graph.
+
+## Decision
+
+AG is a **session-owned, mutable, memory-resident runtime graph**. It is the only working-memory
+container, but it remains non-authoritative: durable truth and provenance stay in STG/LTG, and AG
+disappears when its owning session is released.
+
+AG holds task frames, STG/LTG references, bounded tool observations, temporary relations, unresolved
+working state, reasoning artifacts, activation metadata, and a disclosure ledger. It may keep one
+active task frame and a small bounded set of cooling frames, so a task switch does not destroy state
+and a return does not require reconstructing everything from the transcript.
+
+Each model-visible retrieval freezes an immutable `ProjectionRevision` from the mutable AG. Four
+identities are kept distinct:
+
+- `agId`: the session working graph;
+- `taskFrameId`: one semantic task partition inside AG;
+- `projectionId`: one immutable selection/disclosure/feedback boundary;
+- `boardChannelId`: a Task Board coordination channel.
+
+The update is:
+
+```text
+candidates_t = Project(STG, LTG, q_t, TaskBelief_t)
+AG_(t+1) = Update_B(AG_t, candidates_t, observations_t, TaskBelief_t)
+Projection_t = Freeze(VisibleSubset(AG_(t+1)))
+```
+
+`B` is a hard total budget over nodes, edges, evidence, tokens, graph depth, temporary observations,
+reasoning steps, task frames, and latency. HA scores activation, cooling, reactivation, and budget
+allocation. MGR may traverse the selected AG subgraph and emit bounded hypothetical nodes or
+reasoning edges. HA can then rescore those artifacts before a projection is frozen.
+
+AG has three typed edge layers which must not silently reinforce each other:
+
+1. semantic edges referenced from STG/LTG;
+2. activation/attention edges produced by HA;
+3. hypothetical reasoning/operator edges produced by MGR.
+
+Activation is not truth, and an MGR result is not a memory write. MGR artifacts start as attributed,
+TTL-bound hypotheses and can reach STG/LTG only through a separate verified or explicit `remember`
+path. Persistent HA/MGR model weights, if later justified, live in versioned controller/Lab state
+rather than AG.
+
+The query-scoped `ActiveGraph` became a projection revision, the Pi adapter's flat `SessionRuntimeAg`
+was removed, and the injection window moved into the AG disclosure ledger. The APIs of the day were
+implementation evidence, not compatibility requirements for the target design.
+
+The criteria this decision set, each one a requirement on the runtime and its host wiring:
+
+- the normative design distinguishes AG, task frame, projection revision, and Task Board channel
+  identities;
+- AG is memory-resident and session-owned; no AG content is persisted as authoritative semantic
+  memory;
+- a projection revision freezes exact model exposure and supports later exact get, attribution,
+  verified outcomes, and replay after AG mutation;
+- tool observations and retrieved semantic references share one total AG budget without becoming
+  durable writes;
+- HA fast state is isolated by session/branch; its activation cannot increase semantic confidence or
+  edge stability by itself;
+- MGR uses only bounded selected AG subgraphs, records derivation provenance, and emits hypothetical
+  TTL-bound artifacts;
+- task-switch tests cover continuation, A-to-B switch, A-to-B-to-A return, shared constraints, false
+  switches, compaction, and session cleanup;
+- the query-scoped AG, the Pi runtime AG, and the continuation map are migrated or removed rather
+  than kept as permanent compatibility layers.
+
+## Alternatives considered
+
+1. **Keep query-scoped AG and add a separate task-state manager.** This is the smallest
+   implementation change but retains two working-memory containers and makes compaction/task-return
+   behavior adapter-specific.
+2. **Persist AG as a third semantic graph.** Rejected because temporary activation, tool state, and
+   hypotheses would become confused with durable memory and shared truth.
+3. **Make MGR or HA own working memory.** Rejected because scoring and reasoning engines should
+   remain replaceable capabilities; neither should own evidence, session lifecycle, or disclosure
+   provenance.
+4. **Treat the entire session as one task.** Rejected because topic drift causes contamination and
+   repeated query hashes provide a poor estimate of independent tasks for stability learning.
+
+## Consequences
+
+- The runtime that carries this decision is `src/core/session-active-graph.ts`: one active frame plus
+  a bounded cooling set, frame-local parent chains, a unified item/character budget across all
+  frames, `ttlMs` on artifacts, and the disclosure ledger a host writes with `markDisclosed`.
+- The three compatibility layers the decision named are gone. The Pi adapter's runtime AG was
+  removed, the continuation map no longer exists, and the query-scoped AG is now the projection
+  revision; the identity collision the problem statement named is resolved by the four identities.
+- Which capability stands where, and how each criterion maps to code, is owned by the
+  [runtime blueprint](../../design/session-active-graph-runtime-design.md) §4; this record does not
+  restate that status table.
+- The disclosure ledger is host-neutral: the Pi extension, the Claude plugin, WorkBuddy and DSH mark
+  a projection through the runtime instead of keeping an injection window of their own, so a host that
+  kept one would be a second home for the same rule.
+
+## Deferred
+
+- The full multidimensional shared account (blueprint 4.3): the runtime bounds items and characters
+  across all frames, and the semantic/tool/reasoning account as one ledger remains open.
+- Automatic MGR admission (blueprint 4.4): the runtime primitive for TTL-bound, attributed artifacts
+  is implemented; admitting MGR output without an explicit act remains deferred.
+- HA admission and rescoring on AG (blueprint 4.6): intentionally deferred until natural-utility
+  evidence exists; explicit Lab invocation and its isolation remain available.

docs/decisions/proposed/2026-08-29-session-active-graph-runtime.zh-CN.md → docs/decisions/implemented/2026-08-29-session-active-graph-runtime.zh-CN.md

@@ -2,52 +2,49 @@
 
 [English](2026-08-29-session-active-graph-runtime.md)
 
-**Status:** proposed
-
-截至 2026-08-29 已部分实现：protocol v9 与 `SessionActiveGraphRuntime` 已提供
-daemon 所有的会话状态、不可变 projection 身份、projection-to-trace 来源映射、Pi
-工具/Task Board 观察接入、确定性释放、按会话隔离的 HA 快状态，以及受 projection
-预算约束的 MGR 调用。自动 task/branch 生命周期、统一总预算、共享披露账本和带 TTL
-推理产物尚未满足验收条件，因此本决策继续保持 proposed。
+**Status:** implemented
+**Approved:** unrecorded
+**Relates to:** [会话 AG 运行时蓝图](../../design/session-active-graph-runtime-design.md)
 
 ## 问题
 
-NMG 当前用 Active Graph 表示一次查询的检索结果，同时 Pi 适配器另有扁平的
+NMG 曾用 Active Graph 表示一次查询的检索结果，同时 Pi 适配器另有扁平的
 `SessionRuntimeAg` 保存近期工具状态。这两套结构分别承担工作记忆的一部分，重复了
 生命周期逻辑，也没有为 `AG_t = Project(STG, LTG, q_t, task_t)` 中的稳定任务状态
 提供明确所有者。`activeGraphId` 同时等于检索轨迹 ID，因而无法区分可变工作图和
 不可变暴露记录。
 
 层次化激活（HA）和 Memory-Graph Reasoner（MGR）已经提供候选激活与图遍历原语，
-但目前位于运行时旁路，没有共同作用于一张受预算约束的工作图。
+但当时位于运行时旁路，没有共同作用于一张受预算约束的工作图。
 
-## 提案
+## 决策
 
-将 AG 重新定义为**会话所有、可变、纯内存的运行时图**。它是唯一的工作记忆容器，
-但仍然不是权威记忆：持久事实和来源留在 STG/LTG，拥有它的会话释放时 AG 消失。
+AG 是**会话所有、可变、纯内存的运行时图**。它是唯一的工作记忆容器，但仍然不是权威
+记忆：持久事实和来源留在 STG/LTG，拥有它的会话释放时 AG 消失。
 
 AG 可包含任务分区、STG/LTG 引用、受限工具观察、临时关系、未解决工作状态、推演
 产物、激活元数据和披露账本。它可保留一个活跃任务分区以及少量有界的 cooling 分区，
 从而在任务切换时不销毁状态，返回旧任务时也不必完全依赖 transcript 重建。
 
-每次向模型披露内容时，从可变 AG 冻结一个不可变的 `ProjectionRevision`。明确区分：
+每次向模型披露内容时，从可变 AG 冻结一个不可变的 `ProjectionRevision`。四个身份
+必须明确区分：
 
 - `agId`：会话工作图；
 - `taskFrameId`：AG 内一个语义任务分区；
 - `projectionId`：一次不可变的选择、披露与反馈边界；
 - `boardChannelId`：Task Board 协作频道。
 
-目标更新过程为：
+更新过程为：
 
 ```text
 candidates_t = Project(STG, LTG, q_t, TaskBelief_t)
 AG_(t+1) = Update_B(AG_t, candidates_t, observations_t, TaskBelief_t)
 Projection_t = Freeze(VisibleSubset(AG_(t+1)))
 ```
 
-`B` 继续作为节点、边、证据、token、图深度、临时观察、推演步数、任务分区和延迟的
-总硬预算。HA 负责激活、降温、重新激活和预算分配；MGR 可遍历选中的 AG 子图并产生
-受限的假设节点或推理边；随后 HA 可在冻结 projection 前重新评分这些产物。
+`B` 是节点、边、证据、token、图深度、临时观察、推演步数、任务分区和延迟的总硬预算。
+HA 负责激活、降温、重新激活和预算分配；MGR 可遍历选中的 AG 子图并产生受限的假设
+节点或推理边；随后 HA 可在冻结 projection 前重新评分这些产物。
 
 AG 内部必须区分三层边，且它们不能静默互相强化：
 
@@ -59,9 +56,22 @@ AG 内部必须区分三层边，且它们不能静默互相强化：
 假设，只能通过独立的验证或显式 `remember` 路径进入 STG/LTG。若以后需要持久化
 HA/MGR 参数，它们属于版本化 controller/Lab 状态，而不属于 AG。
 
-当前 query-scoped `ActiveGraph` 改为 projection revision。Pi 的扁平
-`SessionRuntimeAg` 在工具观察进入共享 session AG 后退化为短暂事件接入缓存或被删除；
-injection window 并入 AG 的披露账本。现有 API 只是实现现状，不构成目标设计的兼容要求。
+query-scoped `ActiveGraph` 改为 projection revision，Pi 适配器的扁平
+`SessionRuntimeAg` 已删除，injection window 并入 AG 的披露账本。当时的 API 是实现
+现状，不构成目标设计的兼容要求。
+
+本决策设定的条件，每条都是对运行时及其宿主接线的要求：
+
+- 规范设计明确区分 AG、task frame、projection revision 和 Task Board channel；
+- AG 只存在内存并归属会话；AG 内容不会作为权威语义记忆持久化；
+- projection revision 冻结模型实际看到的证据，并在 AG 变化后仍支持精确 get、归因、
+  验证结果与回放；
+- 工具观察和语义记忆引用共享一份 AG 总预算，但不会因此成为持久写入；
+- HA 快状态按 session/branch 隔离；激活本身不能提高语义置信度或边稳定度；
+- MGR 只消费受预算约束的 AG 子图，保留派生来源，并输出带 TTL 的假设产物；
+- 任务切换测试覆盖连续任务、A→B、A→B→A、共享约束、误切换、压缩和会话清理；
+- 当前 query AG、Pi runtime AG 和 continuation map 被迁移或删除，而不是成为永久
+  兼容层。
 
 ## 考虑过的替代方案
 
@@ -74,23 +84,25 @@ injection window 并入 AG 的披露账本。现有 API 只是实现现状，不
 4. **整场 session 视作一个任务。** 拒绝；主题漂移会造成污染，而每次 query hash
    也不能可靠表示边稳定度所需的独立任务。
 
-## 验收标准
-
-- 规范设计明确区分 AG、task frame、projection revision 和 Task Board channel。
-- AG 只存在内存并归属会话；AG 内容不会作为权威语义记忆持久化。
-- projection revision 冻结模型实际看到的证据，并在 AG 变化后仍支持精确 get、归因、
-  验证结果与回放。
-- 工具观察和语义记忆引用共享一份 AG 总预算，但不会因此成为持久写入。
-- HA 快状态按 session/branch 隔离；激活本身不能提高语义置信度或边稳定度。
-- MGR 只消费受预算约束的 AG 子图，保留派生来源，并输出带 TTL 的假设产物。
-- 任务切换测试覆盖连续任务、A→B、A→B→A、共享约束、误切换、压缩和会话清理。
-- 当前 query AG、Pi runtime AG 和 continuation map 被迁移或删除，而不是成为永久兼容层。
-
-## 风险
-
-- 过度任务切分会破坏因果连续性；切分不足会保留无关状态。
-- 若激活、推演、语义置信度和稳定度没有类型隔离，HA 与 MGR 会形成自强化回路。
-- 可变会话状态增加并发、分支所有权、清理和确定性回放的复杂度。
-- 多个 cooling task frame 可能消耗内存与 prompt 预算却没有实际收益。
-- 在共享运行时落地前先改适配器，会制造更多重复实现；应先实现 Core session AG 和
-  projection 契约，再做宿主接线。
+## 后果
+
+- 承载本决策的运行时是 `src/core/session-active-graph.ts`：一个活跃分区加有界
+  cooling 集合、分区内的 parent chain、跨分区的统一条目/字符预算、产物上的
+  `ttlMs`，以及宿主用 `markDisclosed` 写入的披露账本。
+- 问题里点名的三层兼容结构均已消失：Pi adapter 的 runtime AG 被删除，continuation
+  map 不再存在，query-scoped AG 已成为 projection revision；四个身份之分也消除了
+  当时 `activeGraphId` 兼作轨迹名的歧义。
+- 每项能力的当前状态以及它与代码的对应关系由[运行时蓝图](../../design/session-active-graph-runtime-design.md)
+  §4 拥有，本记录不重述那张状态表。
+- 披露账本是宿主中立的：Pi 扩展、Claude 插件、WorkBuddy 和 DSH 都通过运行时标记
+  projection，而不再各自保留 injection window；若某宿主仍保留一份，同一规则就有了
+  第二个家。
+
+## 未完成项
+
+- 多维共享总账（蓝图 4.3）：运行时已有跨分区的条目与字符上限，语义/工具/推演合成
+  一本账仍未完成。
+- MGR 自动准入（蓝图 4.4）：带 TTL 与来源的产物原语已实现，未经显式动作即接纳
+  MGR 输出仍然延后。
+- AG 上的 HA 准入与重新评分（蓝图 4.6）：在有自然效用证据前有意延后；显式 Lab
+  调用及其隔离仍然可用。

docs/decisions/implemented/2026-09-17-arms-get-their-own-driver.md

@@ -2,6 +2,7 @@
 
 **Status:** implemented
 **Approved:** explicit
+**Superseded by:** [The dispatch loop is shared](2026-09-19-dispatch-loop-is-shared.md)
 Date: 2026-09-17
 Branch: feat/ooo-run-namespace
 **Relates to:** [task-unit semantics design](../../design/task-unit-semantics.md),
@@ -148,3 +149,10 @@ not "not started".
   refusals.
 - `evals/ooo-execution/cycle.test.ts` — 23 cases, unchanged and green: the driver's behaviour is
   preserved.
+
+## What the later supersession leaves standing
+
+The supersession is partial and the later record states its scope: what it moves is where the loop that
+executes a given plan lives. Everything else here stands - the arms keep their own driver, their spec
+declarations and their measurement record, the product still does not own a planning platform, and
+`src/integration/ooo-cycle.ts` stays the specific external-window experiment it is.

docs/decisions/implemented/2026-09-17-arms-get-their-own-driver.zh-CN.md

@@ -2,6 +2,7 @@
 
 **Status:** implemented
 **Approved:** explicit
+**Superseded by:** [派发循环是共享的](2026-09-19-dispatch-loop-is-shared.zh-CN.md)
 Date: 2026-09-17
 Branch: feat/ooo-run-namespace
 **Relates to:** [task-unit 语义设计](../../design/task-unit-semantics.md)、
@@ -114,3 +115,7 @@ in the plan` 失败；把汇合点从 `C` 改名也同样失败。可读性那
   `["A","B","C"]` 会令其失败）、上述两种拒绝、一个计划值同时到达 store 与轮次、spec 映射及其默认值、
   以及解析器的拒绝。
 - `evals/ooo-execution/cycle.test.ts` —— 23 条，未改动且全绿：驱动器行为被保留。
+
+## 这次部分取代之后仍然成立的
+
+取代是部分的，范围写在后一条记录里：被搬动的是**执行一份给定计划的循环住在哪**。本记录其余内容全部成立——臂保留自己的驱动、spec 声明与测量记录；产品仍不拥有规划平台；`src/integration/ooo-cycle.ts` 仍是它那个特定的外部等待窗口实验。