Repository navigation
refactor(reward-memory): slim utility_reducer and eliminate all 56 Any - #8976
Conversation
loopx-agent
left a comment
There was a problem hiding this comment.
Reviewer: model_agent · Gemini · Google (self_reported)
Approval conclusion (author-owned PR; GitHub blocks formal self-approval)
已针对 PR #8976 exact head 2ab9e049af57e7fd27ef19e49274434547c9d231 完成完整审查与本地验证(审查结论:APPROVE,无 P0/P1/P2 问题)。
动机 / Motivation
受影响调用方与运维者:所有调用 loopx/capabilities/reward_memory/utility_reducer.py 的 Reward Memory 读取与归约入口(包括 loopx/capabilities/reward_memory/__init__.py、loopx/capabilities/reward_memory/cli.py 与 loopx/capabilities/reward_memory/utility_projection_validation.py),以及负责静态上限控制的 loopx/canary/ 可维护性棘轮维护者。
修改前(513b6f168a):loopx/capabilities/reward_memory/utility_reducer.py 共 1,431 行并存在 56 处 Any(含多处 cast(...) 与 locals().get("context") 动态取用),validate_reward_memory_utility_projection 单函数长达 510 行且含重复的计数字段/边界字段列表与校验样板,同时未纳入 loopx/canary/module_metric_baseline.json 的零 Any 棘轮门禁。修改后(2ab9e049af):utility_reducer.py 降至 1,182 行(净减 249 行),any_count 与 dict_any_count 均降至 0,并以 any_count: 0, dict_any_count: 0, lines: 1197 纳入棘轮基线。
可观测结果:module_metrics 返回 {'lines': 1182, 'any_count': 0, 'dict_any_count': 0};mypy --strict --follow-imports=silent 对 utility_reducer.py 与 utility_projection_validation.py 及全仓 mypy(17 个文件)均报 0 错误;ruff 全过;5 个核心单元测试套件共 213 passed in 24.86s;reward-memory-architecture-smoke(隔离 HOME)与 reward-memory-utility-reducer-smoke 均输出 ok;canary premerge --tier quick ok: true;loopx check --scan-path ... 报 0 错误;所有公开导出符号、词表复用(_SCOPE_FIELDS)、返回字段结构及 ValueError 报错文本逐字等价。
非目标边界:本 PR 不修改 Stage-1 观测契约、Stage-2 归约算法、错误消息文本、utility_projection_validation.py 或任何 CLI/前端接口,亦不授予自动合入权限。
改动思路 / Approach
全部强类型收口、校验拆解与样板收敛均在既有的 loopx/capabilities/reward_memory/utility_reducer.py 边界内完成,通过标准的 object、dict[str, object]、Mapping[str, object]、Sequence[object] 与 _ReducerContext(TypedDict) 替代 Any 与 cast,避免修改下游 utility_projection_validation.py 与 cli.py。
当前 PR 边界:仅变更 loopx/capabilities/reward_memory/utility_reducer.py 与 loopx/canary/module_metric_baseline.json 两个文件(494 insertions(+), 738 deletions(-),净减 244 行),并严格保留全部现有文档字符串、解释性注释与私有辅助入口兼容签名。
- 消除
Any与cast:移除Any与cast导入,以_ReducerContext(TypedDict)与运行时窄化辅助函数_non_negative_int、_str_list、_assert_canonical_list精确收窄映射、字符串列表及非负整数字段。 - 校验职责拆分与去重:将原
510行的validate_reward_memory_utility_projection拆分为_validate_projection_header、_validate_projection_status、_validate_projection_history、_validate_projection_rejections与_validate_projection_review_proposals五个单职责辅助函数,复用_attribution_counts、_CONFLICT_REJECTION_FIELDS与_PROJECTION_*_FIELDS模块常量,并用_SCOPE_FIELDS(来自.memory_utility)替换_rejected_projection中的内联作用域键元组以严格遵守单词表所有权。 - 棘轮基线注册:在
loopx/canary/module_metric_baseline.json字母序位置注册lines: 1197, any_count: 0, dict_any_count: 0。
具体改动 / Changes
本次 diff 仅触及 2 个文件(+494 / -738),无书面规范文档变更(基于既有 Stage-2 reducer 契约与可维护性棘轮执行):
loopx/capabilities/reward_memory/utility_reducer.py(1,431 -> 1,182行,56 Any -> 0):清除全部Any与cast,收敛重复字段与上下文校验,拆分投影回读校验辅助函数,完整保留 docstring、注释与异常信息。loopx/canary/module_metric_baseline.json(+5 / -0,L35-39):在external_conversations.py与chat_actions.py之间注册utility_reducer.py棘轮基线。
关键代码讲解
_ReducerContext/_normalized_context/reward_memory_utility_reducer_identity(utility_reducer.py:101-106,246-276):用_ReducerContext(TypedDict)明确scope: dict[str, str]与三个引用字段类型,由_normalized_context直接对object入参执行_scope与_ref校验,消除了旧版回读验证中对raw.get(...)的 4 处cast(...);reward_memory_utility_reducer_identity直接复用_normalized_context与_digest。_non_negative_int/_str_list/_assert_canonical_list(utility_reducer.py:129-150):为严格mypy提供运行时窄化,保证observation_count、memory_ref_digests、reason_codes与evidence_refs在零Any约束下保持精确的int与list[str]类型,并复用有序唯一规范列表校验。_rejected_projection与reduce_reward_memory_utility_observations(utility_reducer.py:359-380,658-781):前者通过sorted(_SCOPE_FIELDS)复用.memory_utility权威作用域字段集合避免内联重复;后者先显式声明context: _ReducerContext | None = None,移除locals().get("context")反射取用并在review_proposals组装中展开review并拷贝reason_codes。validate_reward_memory_utility_projection及其五个内部校验段(utility_reducer.py:815-1155):保留对.utility_projection_validation的延迟导入以防循环依赖,按头字段、状态机约束、历史一致性、冲突拒绝与审查提案分段校验,同时保持所有错误检查顺序与报错文本逐字一致。
对主干的风险 / Risks
- 调用方兼容与运行时等价:
utility_projection_validation.py仍导入_object、_exact_fields、_bounded等 20 个符号及常量,且cli.py调用reduce_reward_memory_utility_observations。实测mypy --strict --follow-imports=silent同时通过utility_reducer.py与utility_projection_validation.py,全仓mypy(17 个文件)0 错误;test_reward_memory_field_vocabulary_owner.py验证_SCOPE_FIELDS单一所有权与_scope错误文本完全一致。 - 真实路径与边界验证(exact head
2ab9e049af57e7fd27ef19e49274434547c9d231):module_metrics: PASS —{'lines': 1182, 'any_count': 0, 'dict_any_count': 0}mypy --strict --follow-imports=silent loopx/capabilities/reward_memory/utility_reducer.py& 全仓mypy: PASS — 0 errors(含 17 个全仓文件)ruff check loopx/capabilities/reward_memory/utility_reducer.py tests/canary/test_maintainability_ratchet.py: PASS — All checks passedpytest -q -n 4(test_reward_memory_utility_reducer.py,test_reward_memory_field_vocabulary_owner.py,test_maintainability_ratchet.py,test_semantic_vocabulary_drift.py,test_reward_memory_pipeline.py): PASS — 213 passed in 24.86sexamples/reward-memory-architecture-smoke.py(隔离HOME)与examples/reward-memory-utility-reducer-smoke.py: PASS — 均输出okloopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main: PASS —ok: true, selected: 3, failures: 0loopx check --scan-path ...: PASS —ok: True, errors=0, warnings=0
- 残留说明:
tests/control_plane/test_reward_memory_host_isolation.py中stale/both_off8 项断言在未改动的origin/main上已既有一致失败,本 PR 仅触及上述 2 个文件且未引入新回归。
我的整体评价 / Verdict
- Long-horizon 影响:
improved— 彻底清除utility_reducer.py全部56处Any并净减249行,以零Any与1197行上限锁定棘轮,降低后续长程归约规则演进时的类型漂移与回归风险。 - User-experience 影响:
not_applicable— 纯内部 Python 类型收口、校验拆解与静态棘轮基线注册,无终端输出、配置界面或视觉交互变化;visual_review为not_applicable。 - 可观测语义 / 复用 / 词表:
observable_semantics为equivalent,repository_reuse为reused(复用_SCOPE_FIELDS与既有窄化契约,不新建平行词表),authority_semantics与semantic_alignment均为aligned(candidate_decision="reuse_existing")。
改动聚焦、静态检查与真实测试充分通过、无 P0/P1/P2 阻断项。本 PR 为作者自有且涉及 loopx/capabilities/ 运行时,正式合入仍需维护者授权及远端 CI 确认。
English verdict: APPROVE exact head 2ab9e04 (author-owned COMMENTED approval review); eliminates all 56 Any usages and trims loopx/capabilities/reward_memory/utility_reducer.py from 1,431 to 1,182 lines (-249 net) while registering a zero-Any baseline ceiling (lines<=1197) in loopx/canary/module_metric_baseline.json, verified by strict/repo mypy (0 errors), ruff, 213 passing pytest tests across 5 suites, reward-memory smokes, quick premerge canary, and boundary check.
|
This pull request has merge conflicts with Choose the remote for the base repository, not an out-of-date fork. git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEADFor a same-repository clone whose Keep the DCO |
Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>
2ab9e04 to
dab4ff3
Compare
huangruiteng
left a comment
There was a problem hiding this comment.
Reviewer: model_agent · GPT-5 · OpenAI (self_reported)
APPROVE。 精确 head dab4ff3e5613bab957ecbb8995ffc8927385acf9 未发现阻断项;utility_reducer.py 已从 1,431 行、56 个 Any 收敛到 1,182 行、0 个 Any,额外 reward-memory 真实路径覆盖也全部通过。
动机 / Motivation
受影响的是 Reward Memory reducer 的维护者,以及通过 reduce_reward_memory_utility_observations 和 validate_reward_memory_utility_projection 读取公共 utility projection 的 CLI、pipeline 与测试调用方。旧实现使用 56 个 Any/cast,并把约 500 行的 projection readback 校验放在同一函数中;新增字段或修改错误顺序时,类型检查无法保护大部分动态映射,维护者也要在长函数里追踪多组相似检查。一个具体的 before/after 场景是:一个带冲突 delivery、截断 history 和 review proposal 的 projection 通过 readback;修改前动态值依赖 cast 和相同函数内的重复样板,修改后由 _ReducerContext、显式 Mapping[str, object] 与分段 validator 表达,原有拒绝条件、错误文本和返回结构保持不变。
当前基线上的模块指标确实是 1,431 行和 56 个 Any,精确 head 为 1,182 行和 0 个 Any,并注册为零 Any ratchet。非目标是改变 Stage-1 observation schema、Stage-2 utility 算法、排序、CLI、持久化格式或 authority;本 PR 只收敛现有 reducer 边界内的类型和重复校验代码。
改动思路 / Approach
入口仍是现有公开 reducer 与 validation 函数;权威 observation、context 与 projection 继续由 memory_utility、utility_projection_validation 和本模块的既有常量拥有。改动只把 object/Mapping narrow、计数读取、规范列表校验与投影校验分段放入同一模块的 local helper,最后由原 validate_reward_memory_utility_projection 按原有顺序组合。正路径返回同样的 projection 字典;负路径继续抛出原有 ValueError,不新增 retry、持久化或外部 effect。
flowchart LR
A[Observation sequence + context] --> B[reduce_reward_memory_utility_observations]
B --> C[typed observation/context narrowing]
C --> D[Existing subject and history grouping]
D --> E[validate_reward_memory_utility_projection]
E --> F[header/status/history/rejection/review validators]
F --> G[Same public projection]
C -. malformed input .-> H[Existing rejected projection or ValueError]
上图是 source-derived illustrative flow,不是执行证据。当前 PR 边界是 retain:既有 reducer 和 validation entrypoint 已拥有这些规则,拆到新模块只会增加调用层和循环依赖风险,因此选择在文件内做可逆的 local-helper 整理。目标输出与失败语义保持不变,证据见下方矩阵。
具体改动 / Changes
相对 origin/main afba0841b49980310544ad492b885162e4f6c71a,精确 head 修改 2 个文件,+494/-738。utility_reducer.py 清除 56 个 Any 和 cast,module_metric_baseline.json 同时保留主干新增的其他模块并登记本模块真实上限 1182。PR 原先的 1197 ceiling 与本 head 实际指标不符,rebase 时已校正,没有保留无依据 headroom。
没有可引用的新 RFC;此改动属于 no_spec 的内部重构,验收依据是现有 Stage-2 reducer 契约、字段词表所有权测试和真实调用路径。
关键代码讲解
_normalized_context(utility_reducer.py:246)把 scope 和三个 snapshot/version 字段统一收窄为_ReducerContext,替代旧 validation 中的多处cast。_rejected_projection(:359)用同一 context helper 构造可回读的 rejected projection;作用域字段由_SCOPE_FIELDS唯一拥有。reduce_reward_memory_utility_observations(:658)显式保存可选 context,移除了locals().get("context")的动态取用,同时保持 duplicate/conflict/rejected 统计和返回字段。_validate_projection_header(:815)集中校验 exact fields、状态、布尔边界、identity 与 count 上限,生成的错误文本与旧校验一致。validate_reward_memory_utility_projection(:1118)继续作为兼容入口,按原顺序调用 subject、aggregate、header、status、history、rejection 与 review-proposal 检查。
对主干的风险 / Risks
最强回归是 validator 拆分后检查顺序改变,导致同一 malformed projection 从一条诊断变成另一条,或 reducer 在 conflict/truncated history 下返回不同的计数。源码对照确认相关拒绝分支在旧实现中已经存在,新增 helper 只是把原 body 分段;213 条 reducer、词表、ratchet、semantic drift 和 pipeline 测试通过,另有 105 条 Reward Memory host-isolation 测试通过,守护了 context mismatch、conflict、history 和 projection readback 路径。
两个 production smoke 均通过,quick premerge 的 6 个 direct checks 与 3 个 catalog canaries 为 0 failure,strict/full mypy 与 ruff 也通过。远端 CI 在审查时仍在排队,因此作为 diagnostic 而非本地语义证据。回滚只需 revert squash commit,不涉及迁移或持久数据。
我的整体评价 / Verdict
Long-horizon 影响为 improved:reducer 的类型和校验职责更清晰,56 处动态类型与一处反射式上下文读取被移除,并由真实 metric ratchet 防回退。User-experience 影响为 not_applicable:没有 CLI、配置或 renderer 变更,visual_review 不适用。可观测语义为 equivalent,repository reuse 为 reused,authority 与 semantic alignment 都保持既有边界;没有 blocking finding。
English verdict: APPROVE exact head dab4ff3e5613bab957ecbb8995ffc8927385acf9; the rebased reward-memory reducer removes all 56 Any usages and 249 net lines while preserving the existing reduction and validation contract, verified by strict/repo mypy, ruff, 213 targeted tests, 105 host-isolation tests, two production smokes, and quick premerge with zero failures.
Summary
Eliminates all 56
Anyannotations and slims duplicated validation boilerplate inloopx/capabilities/reward_memory/utility_reducer.py, then registers a zero-Anyratchet ceiling inloopx/canary/module_metric_baseline.json.utility_reducer.pymoves from1,431 -> 1,182lines and56 -> 0Anyusages._ReducerContext(TypedDict)replaceAnyandcastwithout adding a second reducer or validation owner.validate_reward_memory_utility_projectionis split into bounded header, status, history, rejection and review-proposal validators while preserving the public entry point, field vocabulary, returned shape and existingValueErrortext.1182lines, not an inflated ceiling.Verification
Rebased onto
origin/mainatafba0841b49980310544ad492b885162e4f6c71a; reviewed headdab4ff3e5613bab957ecbb8995ffc8927385acf9.module_metrics:{'lines': 1182, 'any_count': 0, 'dict_any_count': 0}.utility_reducer.pyandutility_projection_validation.py: no issues.213 passed.105 passed.reward-memory-architecture-smoke.pyandreward-memory-utility-reducer-smoke.py: passed on the production entrypoints.loopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main: passed with 6 direct checks and 3 catalog canaries,failures: 0,manual_hold_count: 0.Boundary & Rollback
No Stage-1 observation schema, Stage-2 utility algorithm, CLI contract, persisted format or authority model changes. The diff is limited to the existing reducer module and its module-metric baseline and is reversible by reverting the squash commit.