[BigDataDZ]

On native Windows with a CP936 parent locale, the real peer CLI emits valid UTF-8 JSON but the shared collaboration test helper decodes it using the locale default. A legitimate Chinese/emoji brief makes the reader thread raise UnicodeDecodeError, followed by json.loads(None); the CLI's successful result is lost.

Decode the existing helper's stdout/stderr explicitly as UTF-8, retaining strict decoding. Add a real subprocess round trip covering a Unicode brief, receiver adoption, Unicode reply and original parent-request identity, plus the existing unknown-agent rejection behavior. Run that case in the existing native Windows peer CI step. Production code and authority rules are unchanged.

Validation on main dc9fdbb5186618ec2e797a9d7545e2fe524fee56, native Windows/Python 3.12 with CP936:

• New regression failed before the fix with the actual reader-thread/JSON failure; after the fix it passed.
• Full peer-collaboration module: 41 passed, 1 POSIX FIFO skip (428.68 s).
• Independent read-only review found no blockers and separately ran the new case successfully.
• Changed-file Ruff, whitespace and public/private boundary checks passed. The advisory inspected no production vocabulary paths; it is not semantic-equivalence evidence.

The existing shared test helper is the appropriate owner; no new abstraction or production refactor is needed. This qualifies the multilingual fixture and its CI wiring, not all Windows collaboration routes or a production behavior change. Hosted CI remains pending.

[BigDataDZ]

CI follow-through: the native Windows job stopped in the earlier CLI fixture step, before this PR's new peer round trip ran. Three long-path cases emit brief JSON of 10,522 characters against 10,500. Main dc9fdbb5186618ec2e797a9d7545e2fe524fee56 has the same failures; the relevant fixture/output-owner files are unchanged in this PR.

The cause is native JSON path escaping at equal physical path length, not the peer UTF-8 reader change. #6843 proposes a measured, JSON-only characterization adjustment plus real depth comparisons and an owned short pytest namespace. Its focused native checks and full real base/head CLI differential pass locally. Hosted qualification remains pending; this job is not relabeled passed and has not qualified the new peer case.

This PR's independent native result remains 41 passed / 1 POSIX FIFO skip, including the successful Unicode request/adoption/reply round trip. Maintainer review and hosted execution remain separate.

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max

Exact head: 218fab584b599dd6564f598496c3c7d81455b9a7

动机

未发现阻塞性问题。

在中文 Windows 环境运行协作回归测试的维护者，会遇到成功的 CLI 输出被测试读取器误解码的问题。

例如真实 CLI 返回含中文和 emoji 的请求与回复：旧测试辅助函数沿用父进程 CP936，读取时抛异常而丢失成功结果；修复后按既有 UTF-8 输出契约严格解码，并验证完整往返及原请求关联。

本轮真实 CLI 和存储往返在父进程 CP936 编码模型下由基线失败变为当前通过；默认 UTF-8、ASCII、未知 Agent 拒绝与非法字节严格拒绝均保持。

范围是共享测试读取器和现有 Windows CI 选择器；产品代码、权限与用户工作流均不改变。

改动思路

修复既有共享 CLI 测试辅助函数的解码参数，并用真实请求、采纳和回复流程验证；不复制启动器、不更改生产编码 owner，也不引入新抽象或协作协议。

本 PR 的完整边界是测试端正确消费现有 UTF-8 JSON 和原生 Windows 作业的用例接线；生产行为及更广 Windows 路由的资格保持独立。

不处理会保留可重复失败，分别在消费者打补丁会重复错误处理或启动逻辑。当前共用 owner 已是最小有效边界；正常流程无需增加确认、重填信息或新配置。只读诊断/严格解码与真实效果继续由现有层分别负责，不能从测试或缓存成功推断更高权限。

具体改动

完整 dc9fdbb5186618ec2e797a9d7545e2fe524fee56 → 218fab584b599dd6564f598496c3c7d81455b9a7：2 路径 +23/-0。

关键代码讲解

• cli：已有测试辅助函数运行真实模块 CLI，再检查退出码并解析 JSON。新增 UTF-8 参数同时用于 stdout/stderr，保留 strict 解码、30秒超时和隔离 cwd；它修复消费者而不改变生产输出。

• test_peer_cli_preserves_unicode_request_and_reply：沿用真实场景完成请求、读取、采纳、回复和返回读取，逐字比较中文及 emoji，核对原 parent_request_id，并验证未知 Agent 仍被拒绝。现有 Windows peer 作业新增这个准确测试节点。

未找到单独规定此私有测试辅助函数 codec 的书面规范（no_spec）；独立依据是未改动的 cli 模块转发和 entrypoint.pin_utf8_stdio，以及真实 peer 请求、采纳、结果返还的既有行为。整个 loopx、apps、packages 与依赖声明的 base/head 字节相同。新例沿用已有场景，不增加真实渠道或账户接触。

对主干的风险

完整 peer-collaboration 模块39项通过，3条未改动的原生 Windows 条件用例跳过；五项 direct 和两项 selected 全部执行通过。 exact-scope change-quality receipt_missing 仍是独立合并资格限制。

独立同脚本四类实际子进程/原生存储场景：基线只有 CP936 父解码器的中文/emoji 往返失败，当前四类全过；原默认 UTF-8、CP936 ASCII 与非法字节的严格拒绝保持。只有父进程 stdlib 默认 codec 被控制，子进程仍运行真实 CLI，生产源码没有注入或 overlay。未知 Agent 仍被 Core 拒绝，返回内容和原 parent 逐字比对。该模型验证解码器修复，不能当成原生 Windows 整机或 hosted job 已执行。

已读作者关于既有 long-path brief JSON 预算失败和 #6843 的说明；相关输出/预算 owner 未被本 PR 修改，不在此宣称修复或托管资格。新增 YAML 节点使用既有 Windows peer step，静态指向实际通过的新测试函数，未更改 required check 名称或跳过规则。

语义与 CI 对齐

变更停留在本地测试传输与 CI 选择，没有新共享词汇、Actor 或授权规则；生产树逐字节相等。标准 canary 的实际检查均执行通过，但外层因 receipt_missing 仍非通过，保留独立合并资格门。遵循 wait_for_ci=false，未查询、轮询、等待或触发 CI。

我的整体评价

problem_context=goal_achieved 仅覆盖本次有界缺陷；long_horizon 和 user_experience 均为 preserved。可移植测试恢复消费成功结果，产品的持续任务和用户工作流没有变化。结构和规模 proportionate，现有共享 helper 已足够，未发现需要在本 PR 再加的抽象。CP936 是 POSIX 上的真实父解码器模型，未执行原生 Windows 或托管 CI；作者报告的原生41/1结果保留独立来源。 不以用例数推断长期收益，不授予合并、升级或真实 peer 消息权限。

English verdict: APPROVE at 218fab584b599dd6564f598496c3c7d81455b9a7. The existing test helper now strictly decodes the CLI UTF-8 wire independently of parent locale; real request/adoption/reply lineage and authority negatives pass, with production byte-identical. Local validation is recorded with native-host limits and the separate missing managed quality receipt.