[cocolord]

Delivery brief

• Goal/source: make each short LoopX CLI process in the heartbeat chain pay only for the code it calls.
• Current gap: the PR-wait and turn-outcome hooks each import one owner submodule (issue_fix.pr_wait_hook, reward_memory.turn_hooks). Python first initializes the package. Both package __init__ facades eagerly re-exported every workflow: report sources, PR descriptions, issue-fix reward memory, evaluation, dogfood and Codex App staging. Those modules were imported on quota should-run and heartbeat-prompt without being called.
• Observable result: importing a hook no longer imports unrelated capability workflows. Every facade export is still available and is the identical object from its owner module.
• Owning boundary: loopx/capabilities/issue_fix/__init__.py and loopx/capabilities/reward_memory/__init__.py. This uses the existing _EXPORTS + module __getattr__ idiom already used by loopx/capabilities/periodic_report/__init__.py. It does not change TS authority, scheduling, settlement, defaults or output contracts.

Change

• Each facade maps its unchanged __all__ to owner submodules in _EXPORTS.
• __getattr__ imports the owner on first access, then caches the value in module globals. Lazy from . import name in turn_hooks and monkeypatch.setattr on the facade keep working.
• __dir__ lists all exports.
• Name→owner pairs were generated from the original import statements; all 53 + 5 names match __all__.
• New fresh-process test in tests/test_cli_entrypoint.py:
  • importing both hooks must not load issue_fix.{periodic_report,pr_description,reward_memory} or reward_memory.{codex_app_outcome,dogfood,evaluation};
  • every export must be identical to its owner's object and listed in dir();
  • this test fails on base.

Evidence

All measurements use a synthetic SQLite fixture on macOS. Base is 04c5799b8; head swaps only the two facade files inside the same checkout.

Output parity: for quota should-run, heartbeat-prompt --thin and status, base and head JSON outputs are equal once timestamps are normalized. The compressed scheduler host-facts chunks also match after decoding and normalizing timestamps.

Modules imported per command:

Command	Base	Head	Change
quota should-run	568	551	−17
heartbeat-prompt --thin	353	342	−11
status	414	414	0

• Modules no longer loaded on should-run:
  • issue_fix: periodic_report, pr_description, reward_memory, repository_snapshot, repository_commit_evidence, metrics_supplement, outcome_projection;
  • domain_packs, domain_state;
  • reward_memory: decision, dogfood, evaluation, evaluation_fixtures, health, memory_utility, utility_reducer.
• Modules no longer loaded on heartbeat-prompt: reward_memory codex_app_outcome, decision, dogfood, evaluation, evaluation_fixtures, health, memory_utility, outcome_lifecycle, read_authority, runtime_hooks, utility_reducer.

Isolated hook import cost (n=20, alternating order, after importing effect_runtime): 348 ms → 179 ms median. Head was faster in 20/20 pairs, and 227 → 162 loopx modules were loaded.

End-to-end CLI time is not distinguishable from noise at this sample size. Most of the deferred modules are still imported by other paths in the full command.

Command	Wall-clock, n=10 (paired median, head faster)	CPU time, n=16 (paired median, head faster)
should-run	−19 ms (5/10)	−65 ms (8/16)
heartbeat-prompt	−9 ms (5/10)	−19 ms (8/16)
status	−1 ms (5/10)	—

The expected per-command gain is tens of milliseconds, not the ~110 ms this Todo originally estimated.

Validation at head:

• pytest on the reward-memory, issue-fix, PR-wait, outbound-guidance and agent-turn-recall capability suites, plus tests/test_cli_entrypoint.py, tests/test_loopx_turn_driver.py, the import-boundary/owner architecture tests and the compact-identifier owner test: 1022 passed.
• The facade example smokes all passed: examples/reward-memory-*-smoke.py (8), issue-fix-pr-description-builder-smoke.py and openviking-project-peer-provider-smoke.py.
• loopx canary premerge --from-git-diff --git-diff-base origin/main (standard tier): 6 direct and 9 selected checks, 0 failures, 0 manual holds.

Not covered

• Linux and Windows were not measured.
• The end-to-end latency win is small and was not established as statistically distinguishable. The deterministic result is fewer imported modules and lower hook import cost.
• This changes runtime behavior in loopx/**, so it needs maintainer review and merge authorization. It is not self-merged.

[songoow]

Reviewer: model_agent; GPT-6; OpenAI; self_reported.

动机 / Motivation

受影响的是每次唤醒都加载 issue-fix 或 reward-memory 子模块的 CLI 进程。以前导入一个子模块会展开整个包的公开facade；本项把未选择的导出推迟到首次访问。已验证所有公开导出仍解析到原所有者对象，轻量hook导入不再展开未选中的工作流。本项不改变召回、写入、奖励评估策略，也不证明完整任务延迟或token费用下降。

这次判断绑定PR #7749的完整head 52fa115。评价对象是已经存在的实际调用任务与局部可观察结果；重复调用带来的成本或错误恢复路线是本项动机，不能把较小响应或更少导入直接等同于长期任务质量。

改动思路 / Approach

使用有限显式名称映射和模块级 getattr；首次访问导入原模块并缓存原对象，__all__和dir保留发现性。本 PR 的合理边界是两个包facade的导入时机；业务函数和状态所有者保持原位。

沿入口跟踪到原所有者、返回值和失败恢复后，这个实现仍在现有边界上处理问题。无改动会保留原成本或错误指引；另起一套业务规则、持久化状态或通用缓存会增加同步义务，此处没有这类必要性。新增说明和投影应服务原调用方完成任务，不能改变原来的准入、权限、结算或来源完整性要求。

具体改动 / Changes

关键代码讲解

• getattr：将原公开函数名映射到原模块，未知名称继续抛出AttributeError。 输入与拒绝分支沿用现有调用路径；这一处的返回内容或依赖时机改变，不把局部投影提升为新的业务判定。
• getattr：按需返回原class/function对象并缓存，不包装或复制业务实现。 输入与拒绝分支沿用现有调用路径；这一处的返回内容或依赖时机改变，不把局部投影提升为新的业务判定。

调用者公开名称、对象身份和业务实现保持；导入时的异常时机有意变为首次真正访问相应导出。正向场景是：分别访问每个公开导出并比较原所有者对象身份。负向场景是：只导入轻量hook时禁止未选择的heavy模块；未知属性保持拒绝。这些是依据实际源码与已有调用测试核对的行为边界；失败和恢复继续由原owner负责，没有以返回字段存在代替真实状态读取。

规范依据为原公开调用行为、既有所有者和相应回归；已检索改动路径与周边文档，没有为这次局部性能/适配修正找到另一个已接受专项RFC。作者的量测叙述不充当独立功能合同。

实际验证 / Validation

3 项 facade/导出身份/导入隔离测试 PASS；所有改变Python源码编译和 diff hygiene PASS。

执行命令：python -m pytest -q tests/test_cli_entrypoint.py -k 'facade or reward_memory or issue_fix'。本轮统计只包含上述实际执行；作者PR内微基准、旧head完整回归和旧review的历史量测都已阅读，但不冒称本轮重新执行的结果。

对主干的风险 / Risks

调用者公开名称、对象身份和业务实现保持；导入时的异常时机有意变为首次真正访问相应导出。最需要防止的情况是：只导入轻量hook时禁止未选择的heavy模块；未知属性保持拒绝。已核对相关调用方和已有review，未重复发布旧head意见；本轮对新增head的源代码与局部回归单独作判断。原持久化状态和业务所有者保持，仍然不能由小范围通过推断整仓或所有部署入口已经合格。

当前精确head失败证据：test-shard (6)、test-shard (3)、test-shard (2)、chat-bundle-browser (1)、dashboard-browser (2)。当前多个test-shard失败日志未成功取得，且chat与dashboard存在实际失败；源码与局部导入测试不足以归因全部当前失败。 未取得的分片日志明确留作未知；同名错误或改动文件不相交都不足以建立pre_existing归因。

最小补齐是针对当前实际失败取得日志，并提供不可变base/head同命令同条件归因，或能独立覆盖该失败不变量的当前真实入口证据。已完成的聚焦回归不应无条件重做，也不要求无关功能重写；如果归因确认是回归，再修相应最小代码路径。未完成的检查不是单凭pending就要求改代码的理由。

我的整体评价 / Verdict

REQUEST_CHANGES：当前阻断是具体的验证证据缺口。聚焦审查尚未发现可复现的阻断性生产代码缺陷，不能把未归因红灯写成代码finding。 审查仅适用于exact head 52fa115。

本 PR 的合理边界是两个包facade的导入时机；业务函数和状态所有者保持原位。 已验证所有公开导出仍解析到原所有者对象，轻量hook导入不再展开未选中的工作流。 已检查 loopx/capabilities/issue_fix/init.py:getattr；loopx/capabilities/reward_memory/init.py:getattr；变更是CLI/内部读取或文本协议，没有修改浏览器页面布局，视觉截图不适用。

当前多个test-shard失败日志未成功取得，且chat与dashboard存在实际失败；源码与局部导入测试不足以归因全部当前失败。 本项不改变召回、写入、奖励评估策略，也不证明完整任务延迟或token费用下降。所有本轮探针使用合成fixture，没有运行真实模型任务或发起业务作业；本review陈述审查与验证事实，不声称已经发布部署或完成合并。

English verdict: REQUEST_CHANGES — exact head 52fa115. Focused code review and the listed local regressions found no confirmed blocking implementation defect. Approval is held for the named current validation gaps; resolve their source-bound evidence rather than changing unrelated code.