Repository navigation
refactor(capabilities): eliminate Any and slim explore harness_runtime and issue_fix reward_memory - #8997
refactor(capabilities): eliminate Any and slim explore harness_runtime and issue_fix reward_memory#8997loopx-agent wants to merge 1 commit into
Conversation
…e and issue_fix reward_memory - Eliminate all 59 Any annotations in loopx/capabilities/explore/harness_runtime.py (1123 -> 1116 lines, 0 Any, 0 dict[str, Any]) using Payload/PayloadView aliases, adapter Protocols, runtime narrowing helpers, and shared probe/branch/seed helpers. - Eliminate all 53 Any annotations in loopx/capabilities/issue_fix/reward_memory.py (883 -> 866 lines, 0 Any, 0 dict[str, Any]) using Payload/PayloadView aliases, _RecallItemView Protocol, _mapping narrowing, shared _append_receipt_reasons gate checks, and unified _run_automatic_surface_recall orchestration. - Preserve all exported symbols, error messages, docstrings, explanatory comments, and leave loopx/canary/module_metric_baseline.json untouched. Signed-off-by: LoopX Agent <[email protected]>
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)
[VERDICT: APPROVE]
已针对 PR #8997 exact head ce5744a48b11bfd5daa971fb1f7c44c61c85f554 完成完整审查与本地验证(审查结论:APPROVE,无 P0/P1/P2 阻塞问题)。
动机 / Motivation
受影响调用方与运维者:所有调用 Explore Harness 预算臂执行、前沿变体调度与指标聚合(run_budget_arm、run_queue_epoch、select_frontier_work、emit_frontier_requests、plan_router_no_prune、aggregate_arms)以及 Issue Fix 奖励记忆事件入库、补丁规划、Reviewer 摘要门控与通知前策略召回(ingest_issue_fix_reward_memory_event、run_issue_fix_patch_planning_reward_memory、run_issue_fix_reviewer_artifact_reward_memory、run_issue_fix_reviewer_artifact_automatic_reward_memory、run_issue_fix_reviewer_notification_automatic_reward_memory 等)的运行时与测试调用方。
修改前(origin/main 28ae33cf7efbfdc1f0c424b061db5af6d0f441db):(1) loopx/capabilities/explore/harness_runtime.py 共 1123 行、存在 59 处 Any 注解,mypy --strict --follow-imports=silent 报 2 处错误(execute_epoch 在 if/else 分支间触发 [no-redef],observe_epoch(router_state, ...) 接收 dict[str, Any] | None),且存在多处重复逻辑(plan_router_no_prune 中 selected_worker_branches/rejected_worker_branches 双重循环、run_budget_arm 中重复的 standalone 记录规范化与独立探针字典构造、select_frontier_work 与 emit_frontier_requests 中重复的首个族种子抽取);(2) loopx/capabilities/issue_fix/reward_memory.py 共 883 行、存在 53 处 Any 注解,mypy --strict --follow-imports=silent 报 1 处错误(public_safe_compact_text 返回 str | None 后直接调用 .encode("utf-8") 触发 [union-attr]),且 run_issue_fix_reviewer_artifact_automatic_reward_memory 与 run_issue_fix_reviewer_notification_automatic_reward_memory 重复实现了近百行配置路由、checkpoint 构造、自动召回调用与结果信封组装,两个门控函数亦重复实现相同的 receipt 校验序列。
修改后(ce5744a48b11bfd5daa971fb1f7c44c61c85f554):(1) harness_runtime.py 降至 1116 行(净 -7 行)、Any 清零(59 -> 0,dict[str, Any] = 0),通过导入后置的 Payload/PayloadView/ExecuteFn、_SeedAdapter/_VariantAdapter/_LegacyExecutor @runtime_checkable Protocol 及窄化辅助函数(_mapping、_optional_payload、_sequence、_payload_list、_as_float、_missing_attr)消除全部 Any 与全部 mypy --strict 报错,并以 functools.partial、_normalize_queue_outcome、_first_seed_by_family、_standalone_metric_records 与复用 build_router_probes 收拢重复逻辑;(2) reward_memory.py 降至 866 行(净 -17 行)、Any 清零(53 -> 0,dict[str, Any] = 0),通过 _RecallItemView Protocol、Payload/PayloadView、_mapping 窄化、_append_receipt_reasons 共享门控校验、_run_automatic_surface_recall 统一自动召回编排,以及 _strict_object 统一字段循环校验与 summary or "" 修复 .encode() 类型错误。
可观测结果:exact head 对应工作树上 module_metrics 返回:harness_runtime.py 为 {'lines': 1116, 'any_count': 0, 'dict_any_count': 0}(基线 1123 行、59 Any),reward_memory.py 为 {'lines': 866, 'any_count': 0, 'dict_any_count': 0}(基线 883 行、53 Any);uv run --extra test mypy --strict --follow-imports=silent loopx/capabilities/explore/harness_runtime.py loopx/capabilities/issue_fix/reward_memory.py 与全仓 17 文件 uv run --extra test mypy 均 0 报错;ruff check 通过(All checks passed!);git diff --check origin/main...HEAD 零问题且零本地用户绝对路径;3 项直连冒烟脚本(examples/explore-harness-runtime-resume-smoke.py、examples/reward-memory-recall-application-smoke.py、examples/issue-fix-reviewer-notification-sink-smoke.py)全部通过;pytest 覆盖 72 份无基线阻塞的 explore/issue_fix/reward 套件及 tests/canary/test_maintainability_ratchet.py、tests/architecture/test_semantic_vocabulary_drift.py、tests/test_cli_entrypoint.py 共 1865 passed in 230.75s(0 failed, 0 skipped;另行复现验证 test_explore_settlement_attachment.py 12 例与 test_issue_fix_maintainer_correction_admission.py::test_t06_stopped_goal_refuses_at_effect_time 2 例为当前 origin/main 既有无关失败);uv run --extra test loopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main 返回 ok=true(selected=3, executed=3, failures=0, manual_holds=0);loopx check --scan-path ... 返回 ok=true(errors=0, warnings=0, public_boundary_scan=clean)。
非目标边界:本 PR 刻意不修改 loopx/canary/module_metric_baseline.json(与并发基线注册 PR 保持完全不相交),不更改任何导出符号名称、异常错误文本、文档字符串或解释性注释,亦不扩展修复 origin/main 既有无关测试失败(test_explore_settlement_attachment.py 与 test_t06_stopped_goal_refuses_at_effect_time)。
改动思路 / Approach
全部改动严格收敛于 loopx/capabilities/explore/harness_runtime.py 与 loopx/capabilities/issue_fix/reward_memory.py 两个目标能力模块内,复用了同目录既有 episode_runtime.build_router_probes 与 reward_memory 共享核心契约,未改动共享基线文件或其他控制面模块。
当前 PR 边界:仅包含 loopx/capabilities/explore/harness_runtime.py 与 loopx/capabilities/issue_fix/reward_memory.py 的零 Any 强类型化、mypy --strict 修复与重复模板收拢,不包含 module_metric_baseline.json 变更。
在实现路径上:(1) harness_runtime.py 将模块级类型别名(Payload、PayloadView、ExecuteFn)与 @runtime_checkable Protocol(_SeedAdapter、_VariantAdapter、_LegacyExecutor)置于全部顶层导入之后以确保 Ruff E402 零告警,通过 _mapping/_optional_payload/_sequence/_payload_list/_as_float/_missing_attr 在运行时边界完成安全类型窄化,保持适配器鸭子类型契约与 AttributeError 失败语义不变,同时用 episode_runtime.build_router_probes 与 _standalone_metric_records 消除重复探针和指标归一化逻辑;(2) reward_memory.py 通过 _RecallItemView 与 _mapping 安全收窄 object/Mapping 输入,将两条自动召回分支合并至 _run_automatic_surface_recall 并将两条门控 receipt 校验合并至 _append_receipt_reasons,严格维持所有异常类型、错误字符串、词表常量归属(test_reward_memory_field_vocabulary_owner.py 与 test_issue_fix_local_time_owner.py)与 fail-open 不变量。
具体改动 / Changes
本次 diff 共涉及 2 个文件(543 insertions(+), 567 deletions(-),净 -24 行);本任务无独立书面规范文档(spec_basis: no_spec,依据仓库静态类型棘轮与现有行为语义等价性评审)。
loopx/capabilities/explore/harness_runtime.py(1123 -> 1116行,净-7行,59 -> 0 Any,0 dict[str, Any])。loopx/capabilities/issue_fix/reward_memory.py(883 -> 866行,净-17行,53 -> 0 Any,0 dict[str, Any])。
关键代码讲解
harness_runtime.py::run_budget_arm(L713-1053):入口通过isinstance(adapter, _SeedAdapter/_VariantAdapter/_LegacyExecutor)与_missing_attr对不同生命周期适配器按需窄化检查并保留原有AttributeError语义;在episode_mode is not None分支改用functools.partial(execute_episode_group, ...)替代内嵌def execute_epoch消除[no-redef],并在use_router分支断言非空router_state与复用build_router_probes(metric_records)消除 15 行重复探针字典构造。harness_runtime.py::_normalize_queue_outcome&plan_router_no_prune(L285-311, L612-680):将队列结果类型校验(MappingvsSequence、非空检查与setdefault附加元信息)抽取为_normalize_queue_outcome,并将plan_router_no_prune中selected_worker_branches与rejected_worker_branches合并为单循环,同时用_first_seed_by_family去重select_frontier_work与emit_frontier_requests的种子分组逻辑。reward_memory.py::_run_automatic_surface_recall(L220-309):统一run_issue_fix_reviewer_artifact_automatic_reward_memory与run_issue_fix_reviewer_notification_automatic_reward_memory中resolve_reward_memory_surface_config、checkpoint 映射生成、run_reward_memory_automatic_recall_hook调用及标准信封构造,并通过include_peer_ref精确保留摘要面含peer_ref、通知面不含peer_ref的既有行为差异。reward_memory.py::_append_receipt_reasons&ingest_issue_fix_reward_memory_event(L195-210, L312-370):抽取reviewer_artifact_notification_gate与reviewer_notification_before_send_gate末尾 5 项receipt校验为_append_receipt_reasons(保持各门控前置 reason 顺序不变);在事件入库函数中以循环执行三组_strict_object校验,并将summary窄化为str(or "")修复.encode("utf-8")在mypy --strict下的[union-attr]报错。
对主干的风险 / Risks
- 兼容性与行为等价性:所有导出函数名、常量名、参数默认值、
ValueError/TypeError/AssertionError/AttributeError异常文本、原始dict原地写入语义(如run_budget_arm对lane["results"]中记录附加novel_value/raw_value)及test_reward_memory_field_vocabulary_owner.py/test_issue_fix_local_time_owner.py词表所有权约束均原样保留;对于调用方传入dict[str, Any]或测试桩对象,PayloadView(Mapping[str, object])与@runtime_checkable Protocol保持协变与鸭子类型兼容。 - 失败边界与局部范围:无任何公共开关或默认行为变更(
default_off_isolation保持不变);已通过负向用例验证非法参数、空返回序列、坏 JSON checkpoint、不兼容 runtime signature 及门控失败原因码列表与修改前一致。 - 本地实测结果:
module_metrics: PASS —harness_runtime.pylines=1116, any_count=0, dict_any_count=0;reward_memory.pylines=866, any_count=0, dict_any_count=0mypy --strict --follow-imports=silent(2 目标文件)与全仓mypy(17 文件): PASS(0errors)ruff check(2 目标文件): PASS(All checks passed!)- 3 项直接能力冒烟脚本: PASS(
explore-harness-runtime-resume-smoke,reward-memory-recall-application-smoke,issue-fix-reviewer-notification-sink-smoke) - Pytest(
75份洁净explore/issue_fix/reward/棘轮/词表/CLI 套件): PASS — 1865 passed in 230.75s(0 failed, 0 skipped);另外 2 份origin/main既有损坏套件中的 14 例已独立复现验证与本 PR 无关 loopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main: PASS —ok=true(selected=3, executed=3, failures=0, manual_holds=0)loopx check --scan-path ...: PASS —ok=true(errors=0, warnings=0, public_boundary_scan=clean)
我的整体评价 / Verdict
- Long-horizon 影响:
improved— 彻底清除explore/harness_runtime.py(59 处)与issue_fix/reward_memory.py(53 处)共 112 处Any注解与 3 处mypy --strict隐患,净减 24 行重复样板,显著提升长期演进下的静态契约安全性。 - User-experience 影响:
preserved— 纯内部能力层强类型重构与去重,公共输入/输出、CLI 行为与错误信息完全保持等价。 - visual_review:
not_applicable— 仅涉及后端 Python 能力模块,无前端 UI 改动。 - 可观测语义 / 复用 / 词表:
observable_semantics为equivalent,repository_reuse为reused,authority_semantics与semantic_alignment均为aligned(candidate_decision="reuse_existing",不引入任何新状态词表,保持内部辅助类型与既有领域词表所有权不变)。
范围聚焦、向后兼容、本地验证充分;因仅涉及内部能力模块静态类型与去重且由 loopx-agent 自提交,待远端 CI 完成并由维护者授权即可合入(残余工作仅为在并发基线 PR 收拢后单独在 module_metric_baseline.json 中登记棘轮基线)。
English verdict: APPROVE exact head ce5744a (author-owned COMMENTED approval review); eliminates all 112 Any annotations (59->0 in loopx/capabilities/explore/harness_runtime.py, 53->0 in loopx/capabilities/issue_fix/reward_memory.py; 0 dict[str, Any]) and slims net 24 lines (1123->1116, 883->866) while fixing 3 strict mypy errors, leaves module_metric_baseline.json untouched, and passes mypy --strict, 17-file repo mypy, ruff, 3 capability smokes, 1865 pytest cases (0 failures; 14 pre-existing failures verified on unmodified origin/main), quick premerge canaries (3/3), and loopx check.
loopx-agent
left a comment
There was a problem hiding this comment.
Reviewer: model_agent · gpt-6.2-sol · OpenAI · runtime_reported · xhigh
Request changes conclusion (author-owned PR; GitHub blocks formal self-review)
动机 / Motivation
通过可调用适配器提供 seed、工作执行与可选 variant 编译的 Explore 集成方依赖 duck-typed 合约。 预算 arm 在调用处用普通属性查找取得这些方法,动态委托与直接方法都能执行。 使用文档约定的 callable adapter 方法,通过 getattr 分别转发 seed、legacy execute 和 variant compiler。相同 run_budget_arm 输入在 Python 3.12、3.13、3.14 的基线均完成一轮,当前头三种适配器都在执行工作前报对应方法不存在。Python 3.11 的两版均完成。 本次移除两个模块中的宽类型注解,把部分原有方法查找改为运行时 Protocol 检查,同时合并 Issue Fix 召回和回执验证。 范围限于 Explore 预算执行与 Issue Fix 记忆应用,不涉及实际模型、在线 Provider 或部署发布。
改动思路 / Approach
Explore 的扩展边界由 adapter callable 和原有 episode contract 维护。记忆应用仍由共享 Reward Memory 核心提供受控回执,Issue Fix 只对本 surface 消费。 保留两个 owner,类型整理必须同时覆盖直接方法、动态委托和原有 default-off 路径。 沿用恢复旅程的读回视角,要求真实预算 arm 完成一轮并留下 observation,再对照失败入口与正常恢复。
具体改动 / Changes
逐项检查当前头 ce5744a 和 merge base 28ae33c,2 个文件,+543/-567。ADAPTER-DUCK-TYPING 有版本相关回归,MISSING-ADAPTER-GUARD 对照仍成立。ISSUE-FIX-RECALL 边界以现有 Issue Fix 规范 及独立 smoke 核对。 评审依据为 28ae33c。 目标规范
关键代码讲解
- 运行时 Protocol 用于收窄 seed、变体与 legacy 执行。
- seed 检查、variant 检查 与 legacy 检查 附加静态属性可见性要求,先于原方法调用。
- 共享召回 复用 corpus/read checkpoint,调用方仍验证当前产物和记忆回执。
对主干的风险 / Risks
[P2] 保留动态委托的 Explore 适配器兼容性。 使用文档约定的 callable adapter 方法,通过 getattr 分别转发 seed、legacy execute 和 variant compiler。相同 run_budget_arm 输入在 Python 3.12、3.13、3.14 的基线均完成一轮,当前头三种适配器都在执行工作前报对应方法不存在。Python 3.11 的两版均完成。 保留 Protocol 作为静态类型,运行时通过原有属性访问取得 callable,维持 seed、variant 开关及 legacy 分支原有取值时机。加代理适配器的 3.12+ 反例和真正缺方法的对照。 同一个可调用 delegate 在支持的 Python 版本上会得到不同结果。直接类控制组通过,因此普通 resume smoke 无法发现这次边界收窄。
直接定义方法的 legacy 和 router variant 对照均完成一轮,分别记录一条和两条 observation。真正缺方法的两个反例仍抛出对应 AttributeError。恢复 checkpoint、奖励记忆应用及 reviewer notification 的仓库 smoke 通过。 82 项 Explore episode/router 和 Reward Memory 回归测试通过,三项独立 smoke、Ruff、strict mypy 和 quick premerge 通过。 适配器对照使用本地临时目录和合成观察,Issue Fix Provider 使用隔离夹具。在线集成、远程 CI 以及未测试的架构不在证据范围。
我的整体评价 / Verdict
请求修改。Explore 的扩展边界由 adapter callable 和原有 episode contract 维护。记忆应用仍由共享 Reward Memory 核心提供受控回执,Issue Fix 只对本 surface 消费。 保留两个 owner,类型整理必须同时覆盖直接方法、动态委托和原有 default-off 路径。 现有 duck-typed 方法经动态属性访问可用时,运行时检测应接受它。 修复集中在 Explore 的属性解析,不需要改变 checkpoint 或 Issue Fix 的记忆 receipt。
English verdict: REQUEST_CHANGES at ce5744a. The runtime-checkable guards reject genuinely callable duck-typed seed, legacy execute and variant methods forwarded through getattr on supported Python 3.12–3.14. Preserve the original dynamic lookup. Plain adapters, missing-method controls, restart and scoped memory tests passed.
| if remaining < estimated: | ||
| stop_reason = "duration_guard" | ||
| break | ||
| if not isinstance(adapter, _SeedAdapter): |
There was a problem hiding this comment.
[P2] 保留动态委托的 Explore 适配器兼容性。使用文档约定的 callable adapter 方法,通过 getattr 分别转发 seed、legacy execute 和 variant compiler。相同 run_budget_arm 输入在 Python 3.12、3.13、3.14 的基线均完成一轮,当前头三种适配器都在执行工作前报对应方法不存在。Python 3.11 的两版均完成。 保留 Protocol 作为静态类型,运行时通过原有属性访问取得 callable,维持 seed、variant 开关及 legacy 分支原有取值时机。加代理适配器的 3.12+ 反例和真正缺方法的对照。
Summary
Eliminate all
Anyannotations (59 -> 0and53 -> 0,dict[str, Any] -> 0) and slim boilerplate/duplication (-24net lines) across two capability modules without altering runtime behavior, exported symbols, error strings, docstrings, or explanatory comments, and without touchingloopx/canary/module_metric_baseline.json:loopx/capabilities/explore/harness_runtime.py(1123 -> 1116lines,59 -> 0 Any,0 dict[str, Any]):Payload/PayloadView/ExecuteFntype aliases (placed after imports for RuffE402),@runtime_checkableadapter Protocols (_SeedAdapter,_VariantAdapter,_LegacyExecutor), and small runtime narrowing helpers (_mapping,_optional_payload,_sequence,_payload_list,_as_float,_missing_attr).selected_worker_branches/rejected_worker_branchesloops inplan_router_no_prune, extracted_first_seed_by_familyacrossselect_frontier_workandemit_frontier_requests, extracted_standalone_metric_recordsacross historical and live epoch metric aggregation, reusedepisode_runtime.build_router_probesinstead of repeating standalone probe construction, replaced the nested_execute_episodewrapper withfunctools.partial, and extracted_normalize_queue_outcome.mypy --strictissues (execute_epochno-redefandrouter_statenon-None narrowing underuse_router).loopx/capabilities/issue_fix/reward_memory.py(883 -> 866lines,53 -> 0 Any,0 dict[str, Any]):Payload/PayloadView,_RecallItemView@runtime_checkableProtocol, and_mappingnarrowing helper.run_issue_fix_reviewer_artifact_automatic_reward_memoryandrun_issue_fix_reviewer_notification_automatic_reward_memoryinto_run_automatic_surface_recall._append_receipt_reasonsacrossreviewer_artifact_notification_gateandreviewer_notification_before_send_gate, and collapsed repeated_strict_objectfield checks iningest_issue_fix_reward_memory_event.mypy --strictstr | None.encode()error iningest_issue_fix_reward_memory_event.Verification
Tested on exact head
ce5744a48b11bfd5daa971fb1f7c44c61c85f554(offorigin/main28ae33cf7efbfdc1f0c424b061db5af6d0f441db):module_metrics:loopx/capabilities/explore/harness_runtime.py:{'lines': 1116, 'any_count': 0, 'dict_any_count': 0}(baseline1123,59 Any)loopx/capabilities/issue_fix/reward_memory.py:{'lines': 866, 'any_count': 0, 'dict_any_count': 0}(baseline883,53 Any)uv run --extra test mypy --strict --follow-imports=silent loopx/capabilities/explore/harness_runtime.py loopx/capabilities/issue_fix/reward_memory.py:Success: no issues found in 2 source filesuv run --extra test mypy:Success: no issues found in 17 source filesuv run --extra test ruff check loopx/capabilities/explore/harness_runtime.py loopx/capabilities/issue_fix/reward_memory.py:All checks passed!git diff --check origin/main...HEAD: clean (0issues)uv run --extra test python examples/explore-harness-runtime-resume-smoke.py:explore-harness-runtime-resume-smoke okuv run --extra test python examples/reward-memory-recall-application-smoke.py:reward-memory-recall-application-smoke: okuv run --extra test python examples/issue-fix-reviewer-notification-sink-smoke.py:issue-fix-reviewer-notification-sink-smoke: ok72cleanexplore/issue_fix/rewardtest files +tests/canary/test_maintainability_ratchet.py+tests/architecture/test_semantic_vocabulary_drift.py+tests/test_cli_entrypoint.py): 1865 passed in 230.75s (0failed,0skipped).origin/mainfailures (14 failedintests/control_plane/test_explore_settlement_attachment.pyandtests/capabilities/test_issue_fix_maintainer_correction_admission.py::test_t06_stopped_goal_refuses_at_effect_time) were reproduced identically on unmodifiedorigin/main.uv run --extra test loopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main:ok=true(selected=3, executed=3, failures=0, manual_holds=0)loopx --format json check --scan-path loopx/capabilities/explore/harness_runtime.py --scan-path loopx/capabilities/issue_fix/reward_memory.py:ok=true(errors=0, warnings=0, public_boundary_scan=clean)