[loopx-agent]

Goal And Delivered Outcome

A structured peer brief sent to an existing project conversation could lose code indentation, repeated spaces and non-breaking spaces before reaching Codex. The request still completed with its original revision/digest, masking that the receiver had consumed different content. This was reproduced on both File and SQLite against #7419 and current main.

Preserve the serialized peer input in CodexChatAgentSession.send(project_collaboration=True), while retaining the existing ordinary-conversation normalization and blank-input rejection. The existing recipient integration case now compares the native input's complete brief with the accepted request.

Related to #7419. Intended base: loopx-project/loopx:main.

Author Declaration

Written by: model_agent — OpenAI Codex.

Implemented against the existing peer-recipient acceptance in docs/architecture/rfcs/app-conversation-and-async-inbox-v0.md, “Next implementation boundary: a project conversation as peer recipient”, at bb150595664fe0e910db9587cff22ea3b471f5ee.

Criterion	Disposition	Implementation / validation
Receiver consumes the current accepted brief through the existing Session	implemented for this defect	CodexChatAgentSession.send; exact native-input brief assertion on File/SQLite
Preserve ordinary bound-human behavior	implemented	Original normalization expression retained; adjacent chat regression suite
Installed provider delivery and natural receiver adoption	deferred	Existing parent acceptance remains open; isolated synthetic-inference evidence does not establish these

Scope And Continuation

Complete for the reproduced input-fidelity defect. Reuses the existing adapter, request, revision, source-grant, FIFO and return owners; adds no protocol, configuration or UI surface. No additional helper or refactor is justified at this boundary.

The broader original-project recipient journey remains subject to its existing installed-App and actual receiver-adoption acceptance. This proposal is independently reversible and does not certify that parent outcome or resolve historical validation/review holds on #7419.

Validation

• Tested commit: 0104c10808f3daf0b7a31f8a3a7a1735ec51095f.
• Run state: finished. Input classes: synthetic, public_fixture.
• Latest inspected target-main snapshot: bee2f29b34f8f53a4dcdd31e1e00a5e2bc10a43a; its changes do not touch this adapter or integration fixture. Branch merge base remains bb150595664fe0e910db9587cff22ea3b471f5ee.

Check kind	Result	Evidence / limitation
regression_parity	passed	The exact-brief assertion failed on File and SQLite at the immutable main baseline; both pass at the submitted commit. Reproduce with uv run --extra test python -m pytest tests/control_plane/test_original_conversation_peer.py -k explicit_grant_and_returns -q.
integration	passed	538 tests across test_original_conversation_peer.py, test_chat_agent.py, test_chat_queue_preparation.py and test_chat_ordinary_project.py; before the final behavior-equivalent one-line conditional formatting. Final affected integration rerun: 2 passed.
real_entrypoint / real_backend	passed	10 cases through isolated Codex 0.163.0-alpha.11.1 app-server processes and real File/SQLite stores. Covers exact brief, original native-thread resume, post-answer correction, restart/lost-return recovery and grant revocation at native write with zero model requests. Inference responses were synthetic loopback data. The extra local qualification harness is not published; the durable exact-input regression is included here.
static	passed	Changed-file and repository CI-scope Ruff; configured mypy (17 sources); public-boundary scan (6 checks, zero errors/warnings); diff hygiene and bilingual RFC check.
static	passed	Exact-scope change-quality receipt cqr_66246256ee53840eb5e7, fingerprint 66246256ee53840eb5e70ce7c9d64baa5e8ae8c8cdc68c69c7fe973bd95c209a. Two files, no review safe-fix pass, zero blockers/warnings/advisories.
static	passed	Standard loopx canary premerge against the pinned target-main snapshot: six direct checks, two selected canaries, exact receipt valid, no manual holds. Earlier stale-receipt attempts remain failed observations.
manual	not_run	Installed App, natural receiver adoption and real-model answer quality. No UI/TypeScript/authority-store implementation changed.

The native execution evidence qualifies transport and recovery with synthetic inference. It is not a measurement of model quality, full cost, natural adoption or installed behavior. Current required CI and exact-head review remain separate from these author checks.

Frontend / Visual Evidence

UI impact: none. This repairs bytes supplied to the existing conversation executor; no rendered controls or visual states change.

Type of Change

• Bug fix
• Test update
• Host or runtime integration

Boundary Checklist

• Public-safe code, fixture and aggregate evidence only; no private state, credentials, raw traces, local paths or internal links.
• Existing test extended; no benchmark work or new task/program.
• No paired RFC changed.
• DCO-signed commit; normal author proposal, pending review.

[loopx-agent]

Reviewer: actor_kind=model_agent; model=gpt-6.1-sol; provider=OpenAI; declaration_source=runtime_reported; reasoning_effort=max.

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机

向既有私有项目会话交付结构化 peer brief 的请求方，会遇到输入被改写的问题。

旧版在发送前折叠空白，导致 brief 中四格缩进、连续空格和不换行空格丢失；本次保留 peer 输入原文，使实际 native input 与获准的 brief 一致。

已验证的结果是原始 brief 完整到达既有会话，同时保留普通消息规范化、空白拒绝、当前授权检查和结果返回。

本次不新增会话、Goal、权限、UI 或协议版本，也不把合成模型响应等同于真实接收方理解和采用。

已安装 App 与真实 provider 的自然接收、独立采用和用户最终任务成果，仍由原有项目协作验收继续资格验证。

改动思路

输入保真属于现有 Codex native adapter 的发送边界；授权、FIFO、修订和返回仍复用共享协作 owner，没有理由增加第二套状态或消息通路。

本 PR 修复 send(project_collaboration=True) 的输入保真，并通过既有 File/SQLite 原会话链路验证；不扩大项目协作权限或宣称父级协作目标完成。

共享协作 owner 先验证 source/recipient grant，再把当前 request 放进原 Session FIFO。execute 读取当前 brief，以 revision/digest 绑定 native write；send 仅在明确的 project_collaboration 模式保留原始字符串。普通输入继续原来的规范化。接收后的 decision 与 return 仍经既有存储返回请求方，没有增加另一条消息或状态通路。

不修改 serializer、全局消息清洗或 UI 可以更直接定位缺陷：JSON 中的空格在 native adapter 被再次规范化才产生损坏。这一修复没有增加用户导航、复制、重复输入或确认轮次；原有明确 source grant 保持必要的权限边界。

具体改动

两文件仅有生产 +2/-2、测试 +4。新增断言读取真实 fake app-server 子进程捕获的 turn/start，并将其中 brief 与提交的原始结构比较，能发现仅检查 revision/digest 或结果文字会漏掉的数据损坏。

先于 diff 阅读的规范：docs/architecture/rfcs/app-conversation-and-async-inbox-v0.md，固定修订 bb150595664fe0e910db9587cff22ea3b471f5ee。下列标签用于本次条款映射，不冒充规范原有编号。

• PEER-IDENTITY：implemented。既有 Session/native thread 与 null Goal 保持不变。
• PEER-CURRENT：implemented。当前 revision-bound brief 完整到达，并经过原 assessment/result owner 返回。
• PEER-FENCE：implemented。native write 时的撤权、身份、修订与取消检查保留；重试仍恢复同一请求和 Turn。
• MODEL-ADOPTION：deferred。packaged/installed provider 交付和自然接收方采用仍按原父级验收继续。

关键代码讲解

1. CodexChatAgentSession.send：project_collaboration=True 使用原字符串；False 继续 space-join(split())。strip() 只判断全空白拒绝，没有把处理后的字符串发送出去。bound_owner、execution_mode、read_only_review、output_schema 的现有拒绝条件保持不变。
2. OriginalConversationRequests.execute：现有唯一 production caller 拼装指令与完整 JSON，传入 project_collaboration=True。start_guard 在写入时重新读取当前授权、recipient 和 revision，输入保真不会绕过它。
3. _complete_answer：继续核验当前 revision/digest、独立 decision 和原返回路线；已保存 answer 的重试不再次调用模型。

独立验证：PASS，test_original_conversation_peer.py、test_chat_agent.py、test_chat_ordinary_project.py 合计 497 项。File/SQLite 的新增原文断言在 immutable base 均 FAIL，在 head 均 PASS，准确暴露缩进、双空格和 NBSP 被改变。另有 22 个独立 send 用例 PASS，覆盖 code fence、换行、Unicode 空白、普通模式、全空白拒绝及非法 recipient 配置；相同 oracle 在 base 因原文被折叠而失败。

对主干的风险

最强反例是启用 peer 模式后绕开普通输入约束，或在撤权/修订之后继续把旧 brief 送入 native Turn。现有明确模式和 bound_owner 检查未变，独立用例验证普通模式仍输出原来的规范化结果；全空白输入和非法配置在写入前拒绝。File/SQLite 生产链路还覆盖 native-write 时撤权、修订竞争、取消、FIFO 并发重试、队列丢失、完成后更正、restart 和 lost-return 恢复。

标准 premerge 的 6 个 direct、2 个 selected 检查均 PASS；DCO sign-off 和完整两文件差分已检查，没有引入私有数据、绝对路径或新词汇。未查询、轮询或等待 GitHub CI。此次执行跨真实本地存储与 native adapter，app-server/model 响应为合成；它证明输入和协议链路，真实接收者是否理解、独立采用及已安装 App 行为仍需原验收证据。

语义与 CI 对齐

本 PR 使用已有 typed recipient、grant、revision 和 project_collaboration 模式，没有新增状态或通过字符串分类决定权限。peer input 保真是明确的默认行为修复；ordinary 分支保持兼容。两个文本语义不同的现有调用分支值得保留，另加 sender/helper 会扩大维护面。

我的整体评价

未发现新增阻断 finding。结论为 APPROVE；交付判断是 justified_increment。long_horizon 与 user_experience 均为 preserved：更正和返回继续由原 owner 恢复，输入损坏得到修复，普通会话没有额外准备和确认负担。这里的完整性证据覆盖原始数据到 native input，不把模型的合成回答当作自然采用。

已检查 chat_agent.py、既有 collaboration 页面调用与两文件差分；此修复只改变 native input，不改变页面渲染和布局。

此次独立验证使用真实本地存储和 native adapter、合成 app-server 响应；已安装 App 的自然接收体验仍未在本次复验。

现有 boolean 分支与共享协作 owner 的 placement 合理，当前无需引入新抽象。保留普通消息兼容行为，也保留原 request/revision/answer 的持久格式。该 head 的源代码审阅与本地检查可批准；merge readiness、maintainer merge authorization 和安装资格继续独立核验。本评审固定 head 0104c10808f3daf0b7a31f8a3a7a1735ec51095f，发布前再次核对 head。

English verdict: APPROVE at 0104c10; no blocking finding. Exact peer input, ordinary-mode parity and File/SQLite recovery pass; installed-provider adoption remains outside this source qualification.