feat(pr-review): reuse repository-owned review experience with Reward Memory

docs/architecture/rfcs/loopx-overall-roadmap-v0.md

@@ -96,7 +96,7 @@ These directly determine whether a long-running team is usable. A directory or R
 | [Goal Vision/replan](../../reference/protocols/goal-vision-replan-contract-v0.md), [work graph](../../reference/protocols/task-graph-projection-v0.md), [peer runtime](../../reference/protocols/peer-agent-runtime-v1.md), [supervisor](../../reference/protocols/peer-supervisor-v0.md) | S2/S3 | Exercise dependencies, replanning, acceptance and handoff in one real case; aggregate closeout consumes acceptance facts |
 | [Quota](../../quota-allocation.md), [cadence](../../operations/long-task-cadence-policy.md), [attention](../../operations/attention-queue.md) | S5/S7 | Budget exhaustion, deferral and blocking expose next triggers/readback; scale without frequent full-state polling |
 | [Material lifecycle](../../reference/protocols/material-lifecycle-architecture-v0.md), [material frontier](../../reference/protocols/agent-material-frontier-v0.md), [authority registration](../../operations/authority-source-registration.md) | S6 | Agents discover roadmap/RFC revisions and record reads; reading grants neither agreement nor authority; archival preserves raw-source ownership |
-| [Decision Context](../../../loopx/capabilities/decision_context/README.md), [Reward Memory](../../../loopx/capabilities/reward_memory/README.md), [Semantic Preference](../../../loopx/capabilities/semantic_preference/README.md), [Turn Recall](../../../loopx/capabilities/agent_turn_recall/README.md) | S6/S11 | Distinguish facts/preferences/advice/attribution/authority; scoped recall, expiry and outcome-feedback counterexamples first |
+| [Decision Context](../../../loopx/capabilities/decision_context/README.md), [Reward Memory](../../../loopx/capabilities/reward_memory/README.md), [Semantic Preference](../../../loopx/capabilities/semantic_preference/README.md), [Turn Recall](../../../loopx/capabilities/agent_turn_recall/README.md) | S6/S11 | Distinguish facts/preferences/advice/attribution/authority; scoped recall, expiry and outcome-feedback counterexamples first. Start review learning with the [#5944 Git-owned experience](../../../loopx/capabilities/pr_review_queue/experiences/README.md): preserve machine and maintainer-directed frames, deliver qualified advice to the explicitly opted-in review Agent, then verify current-head adoption and useful review outcomes. Context delivery closes no quality/utility gate; held-out false-blocker, cost and attention controls follow [the existing utility RFC](post-outcome-memory-utility-attribution-v0.md#91-review-learning-pilot-repository-experience-before-utility) |
 | [Issue Fix](../../../loopx/capabilities/issue_fix/README.md), [PR Review](../../../loopx/capabilities/pr_review_queue/README.md), [Change Quality](../../../loopx/capabilities/change_quality/README.md), [Integration Branch](../../../loopx/capabilities/integration_branch/README.md), [Change Window](../../../loopx/capabilities/repository_change_window/README.md) | S8/S12 | Engineering caller: issue→implementation→independent review→validation→authorized delivery; source-head drift invalidates old evidence |
 | [Explore](../../../loopx/capabilities/explore/README.md), [Benchmark Toolkit](../../../loopx/capabilities/benchmark_toolkit/README.md), [auto-research](../../product/use-cases/auto-research/README.md) | S8/S11 | Second collaboration journey: question/hypothesis→parallel experiments→independent interpretation→successor; no uplift/failure remain valid outcomes |
 | [Content Operations](../../../loopx/capabilities/content_ops/README.md), [Periodic Report](../../../loopx/capabilities/periodic_report/README.md), [office operations](../../product/use-cases/office-operations/README.md), [domain packs](../../product/domain-capability-packs.md) | S5/S8 | Public-safe projections feed sinks; separate drafts/review/publish authority/outcomes; generic sinks do not depend on project-private documents |

docs/architecture/rfcs/loopx-overall-roadmap-v0.zh-CN.md

@@ -82,7 +82,7 @@ managed 与 attached 的工作对话都应能持续在 LoopX 中进行：沿用
 | [Goal Vision/replan](../../reference/protocols/goal-vision-replan-contract-v0.md)、[work graph](../../reference/protocols/task-graph-projection-v0.md)、[peer runtime](../../reference/protocols/peer-agent-runtime-v1.md)、[监督](../../reference/protocols/peer-supervisor-v0.md) | S2/S3 | 将跨工作依赖、重规划、验收与 handoff 放进同一个真实案例；aggregate closeout 必须消费验收事实 |
 | [quota](../../quota-allocation.md)、[cadence](../../operations/long-task-cadence-policy.md)、[attention](../../operations/attention-queue.md) | S5/S7 | 预算耗尽/延期/被阻塞时有明确下一次触发及用户回读；百 Agent 不靠高频全文轮询 |
 | [材料生命周期](../../reference/protocols/material-lifecycle-architecture-v0.zh-CN.md)、[材料 frontier](../../reference/protocols/agent-material-frontier-v0.md)、[authority 注册](../../operations/authority-source-registration.md) | S6 | 路线/RFC 更新能由 Agent 按 revision 发现并登记阅读；read receipt 不表示同意或获得权限；归档不丢原始来源 |
-| [Decision Context](../../../loopx/capabilities/decision_context/README.md)、[Reward Memory](../../../loopx/capabilities/reward_memory/README.md)、[Semantic Preference](../../../loopx/capabilities/semantic_preference/README.md)、[Turn Recall](../../../loopx/capabilities/agent_turn_recall/README.md) | S6/S11 | 区分事实、偏好、建议、归因和权威；同一 scope 的回忆/失效/结果反馈负例先行 |
+| [Decision Context](../../../loopx/capabilities/decision_context/README.md)、[Reward Memory](../../../loopx/capabilities/reward_memory/README.md)、[Semantic Preference](../../../loopx/capabilities/semantic_preference/README.md)、[Turn Recall](../../../loopx/capabilities/agent_turn_recall/README.md) | S6/S11 | 区分事实、偏好、建议、归因和权威；同一 scope 的回忆/失效/结果反馈负例先行。以 [#5944 的 Git 经验](../../../loopx/capabilities/pr_review_queue/experiences/README.zh-CN.md)起步：保留机器与维护者指引下的口径，将合格建议投递到明确 opt-in 的 review Agent，再验证当前 head 的采用与有用评审结果。投递不关闭质量/utility gate；held-out 误阻断、成本和注意力对照遵循[现有 utility RFC](post-outcome-memory-utility-attribution-v0.zh-CN.md#91-评审学习试点先沉淀仓库经验再证明效用) |
 | [Issue Fix](../../../loopx/capabilities/issue_fix/README.md)、[PR Review](../../../loopx/capabilities/pr_review_queue/README.md)、[Change Quality](../../../loopx/capabilities/change_quality/README.md)、[Integration Branch](../../../loopx/capabilities/integration_branch/README.md)、[Change Window](../../../loopx/capabilities/repository_change_window/README.md) | S8/S12 | 作为工程团队的端到端 caller：问题→实现→独立 review→验证→批准范围内交付；head 漂移不复用旧证据 |
 | [Explore](../../../loopx/capabilities/explore/README.md)、[Benchmark Toolkit](../../../loopx/capabilities/benchmark_toolkit/README.md)、[auto-research](../../product/use-cases/auto-research/README.md) | S8/S11 | 作为第二类协作旅程：问题/假设→并行实验→独立解释→后继；无提升/失败也形成有效结果 |
 | [Content Operations](../../../loopx/capabilities/content_ops/README.md)、[Periodic Report](../../../loopx/capabilities/periodic_report/README.md)、[office operations](../../product/use-cases/office-operations/README.md)、[domain packs](../../product/domain-capability-packs.md) | S5/S8 | 公共投影驱动显示；草稿、review、发布权限与结果分离；不将 generic sink 绑定项目私有文档 |

docs/architecture/rfcs/post-outcome-memory-utility-attribution-v0.md

@@ -1,6 +1,6 @@
 # RFC: Post-Outcome Memory Utility Attribution v0
 
-- Status: Accepted; Stage 1 shipped, Stage 2 delivered by this PR, Stage 3+ remain proposed
+- Status: Accepted; Stage 1 and Stage 2 implemented, Stage 3+ remain proposed
 - Supersedes / closes: none
 - Date: 2026-08-15
 - Tracking issues: [#3214](https://github.com/huangruiteng/loopx/issues/3214), [#3824](https://github.com/huangruiteng/loopx/issues/3824)
@@ -368,6 +368,48 @@ bounded rank-prior effect plus readback receipt.
 Stage 4 runs a bounded pilot with held-out or counterfactual evaluation before
 any default-on ranking influence.
 
+### 9.1 Review-learning pilot: repository experience before utility
+
+[PR #5944](https://github.com/loopx-project/loopx/pull/5944) supplies the first
+[repository-owned comparison and experience](../../../loopx/capabilities/pr_review_queue/experiences/README.md).
+Earlier machine approvals covered technical recovery design; a later
+maintainer-directed model review requested a first useful recovered Goal,
+earlier packaged interaction and useful post-recovery results. Preserve each
+frame, actor provenance, API target commit, body-declared head, evidence and
+unresolved outcome. Different frames and source heads prevent treating this
+as a controlled experiment or automatic gold label.
+
+The first pilot phase stores public distilled evidence in Git using the existing
+`procedural_experience_contract_v0`. `pull-request-review` owns the source and
+read-only file adapter; `reward_memory` retains qualification, recall and the
+TypeScript context-delivery decision. Self-repair points to the same case and
+repairs its owning workflow. No second store, shared decision owner, utility
+score or provider write is introduced.
+
+A registered Agent needs qualified Reward Memory enablement, automatic recall
+and an explicit `pull_request_review.review` surface. Its actionable PR packet
+can deliver at most three repository-matched, exact-readback experiences.
+Historical verdict labels stay outside retrieval inputs. Disabled, other-surface,
+unqualified and unrelated-repository calls preserve their base behavior;
+malformed/changed sources fail open. Native public advice extends only this
+opt-in caller boundary; provider ranking, configuration, action authority and
+exact-head work selection are unchanged.
+
+This phase accepts a versioned case, qualified experience, real source readback
+and JSON/Markdown delivery at the production CLI, plus off-parity and failure
+controls. Context delivery does not establish semantic adoption or better
+reviews. Explain adoption, rejection or irrelevance in the existing
+`problem_context` judgment with current-head evidence.
+
+The next phase pins one common review frame, artifact revision, model and budget
+across baseline/experience arms on held-out cases, including justified internal
+prerequisites and false-blocker controls. Retain disagreement; measure actionable
+findings, incorrect blockers, review cost and human attention per accepted review.
+A maintainer correction is evidence, not an unconditional reward. Only verified
+application and work-outcome lineage can feed the existing utility observation
+and reducer. The original case has no such lineage, so its utility is unproven.
+This pilot closes no Stage 3/4 ranking or general review-quality qualification.
+
 ## 10. Validation criteria
 
 Focused fixtures must prove:

docs/architecture/rfcs/post-outcome-memory-utility-attribution-v0.zh-CN.md

@@ -1,6 +1,6 @@
 # RFC：结果后记忆效用归因 v0
 
-- 状态：已接受；Stage 1 已交付，Stage 2 由本 PR 交付，Stage 3+ 仍为提案
+- 状态：已接受；Stage 1 与 Stage 2 已实现，Stage 3+ 仍为提案
 - 替代 / 关闭：无
 - 日期：2026-08-15
 - 跟踪 issue：[#3214](https://github.com/huangruiteng/loopx/issues/3214)、[#3824](https://github.com/huangruiteng/loopx/issues/3824)
@@ -282,6 +282,38 @@ Stage 3 增加 OpenViking readback；只有 provider 协议真实存在时，才
 
 Stage 4 先做有界 pilot、held-out 或 counterfactual evaluation，再考虑让排序影响 default-on。
 
+### 9.1 评审学习试点：先沉淀仓库经验，再证明效用
+
+以 [PR #5944](https://github.com/loopx-project/loopx/pull/5944)作为第一条
+[仓库持有的判断对照与经验](../../../loopx/capabilities/pr_review_queue/experiences/README.zh-CN.md)。
+先前机器评审覆盖恢复的技术设计，后续维护者指引下的模型评审要求明确第一个有用的
+恢复 Goal、更早的打包交互和恢复后有用结果。分别保留口径、actor 来源、API target
+commit、正文声明的 head、证据和未解决的结果。口径与 source head 差异意味着
+这不是受控实验，也不能自动生成金标准标签。
+
+试点第一阶段把公共蒸馏证据存入 Git，复用现有
+`procedural_experience_contract_v0`。`pull-request-review` 持有来源与只读文件
+adapter；`reward_memory` 保留资格化、recall 和 TypeScript 上下文投递判断。
+self-repair 指向同一个案例并修复原有工作流，不新建 store、共享 decision owner、
+utility 分数或 provider 写入。
+
+已注册 Agent 须具备资格化的 Reward Memory enablement、自动 recall 和明确的
+`pull_request_review.review` surface。可执行 PR packet 才可投递最多三条仓库匹配、
+精确回读的经验；历史 verdict 标签不进入检索输入。关闭、其他 surface、未资格化
+和无关仓库调用保持原行为；损坏或变化中的来源 fail-open。内置公共建议只扩展此
+opt-in caller 边界，不改变 provider ranking、配置、行动权限或 exact-head 选工。
+
+本阶段验收是版本化案例、合格经验、真实来源回读与生产 CLI 的 JSON/Markdown 投递，
+加关闭等价和失败对照。上下文投递不证明语义采用或评审质量变好。
+在现有 `problem_context` 判断中，用当前 head 证据说明采用、拒绝或不适用。
+
+下一阶段在 held-out 案例上固定共同的评审口径、产物 revision、模型和预算，比较
+baseline/experience 两组，包含合理内部前置阶段与误阻断对照。保留分歧，衡量每份
+被接受评审的可行动发现、错误 blocker、评审成本与人工注意力。
+维护者纠正是证据，不是无条件 reward。只有验证后的采用与工作结果 lineage 才能进入
+现有 utility observation/reducer；原案例没有该 lineage，效用仍未证明。
+此试点不关闭 Stage 3/4 排序资格或通用评审质量资格。
+
 ## 10. 验证标准
 
 Focused fixture 必须证明：

loopx/capabilities/pr_review_queue/README.md

@@ -1275,3 +1275,21 @@ its existing values, and a new namespace uses capability defaults.
 审阅优先级），部分修改保留既有目标值，新覆盖使用 capability 默认值。关闭时不
 查询、轮询或等待 CI；本地必需验证、当前提交评审、评论及权限检查仍然适用。
 GitHub `BLOCKED` 只提示另需管理员授权，不授予合并权限。
+
+## Repository experience for opted-in review Agents
+
+The existing Reward Memory experiment can deliver Git-versioned procedural
+advice in actionable review packets. Enable the registered Agent, automatic
+recall and the explicit `pull_request_review.review` surface in its existing
+configuration. See [operation, readback and disable](experiences/README.md).
+The [#5944 comparison](experiences/pr-5944-review-frame.md) is the first stored
+case. Historical verdicts are outside retrieval; packet delivery proves neither
+adoption nor utility. Base/off paths, queue selection and review authority retain
+their existing contracts; the native file reader writes no provider.
+
+已开启的 review Agent 可通过现有 Reward Memory 实验，在可执行 review packet 中
+收到 Git 版本化的过程经验。须启用已注册 Agent、自动 recall，并在原配置中明确指定
+`pull_request_review.review` surface，参见[操作、回读与关闭](experiences/README.zh-CN.md)。
+[#5944 判断对照](experiences/pr-5944-review-frame.md)是第一条案例。
+历史 verdict 不进入检索，packet 投递不证明采用或效用。关闭路径、选工和评审权限
+仍遵循现有合同；内置文件 reader 不写 provider。