[loopx-agent]

Problem and resulting behavior

An enabled native guard always emitted its first frame with the 64,000-byte default, even when a caller needed a smaller host output budget. The existing --max-page-output-bytes was accepted only by the continuation reader, so truncation could hide the first continuation command before recovery was discoverable.

The same option now applies to enabled JSON native should-run and context-project. It bounds the complete successful frame, including metadata and newline, and preserves the cap in returned continuation and full-recovery commands. Omitting it retains the existing 64,000-byte default and ambient-environment invariance. Disabled or invalid uses reject before quota admission; presentation never creates another decision, renews a lease or spends quota.

Native regression fixtures now size each batch and its UTF-8 offsets consistently, apply the generated 20,000-token profile's declared direct-output budget to the first frame too, and retain the explicit 10,000-token history/eight-read case. Real model-visible bytes, original hashes/cursors, full recovery and exactly two original guards remain independent assertions. An unknown dependency accepts the typed owner's null status while retaining must-attempt and the bounded-three poll limit; one trailing blank line is removed.

The bilingual 2.2.0 notes document this per-command option, merged ordinary self-claim behavior and verified community attribution. No optional feature is enabled by this change, and no host history limit or frozen qualification criterion increases.

Validation

At exact head 0c3c68b6f15df87d54fc5203ee0e264bd974fe97: the complete actual-native context module and related CLI/lifecycle/dependency/machine-guard modules passed 280 tests, 0 skips, using the real Codex binary and a deterministic isolated local provider. Required Ruff, mypy, syntax, nine-path public-boundary, all 12 bilingual release-usage surfaces and risk-based premerge passed with no failures or manual holds. CLI help/manpage preflight passed. Source-before/source-after identity matched and isolated cleanup succeeded. Failed and interrupted earlier attempts remain failed and are not relabeled as qualification.

Final integrated workspace, installed package, paid model and publication-channel qualification remain separate.

Existing contract

docs/reference/required-work-context.md, docs/development/testing-and-quality.md budget decisions, and the existing digest-bound reader/native activation owners. The implementation extends an existing explicit transport limit rather than making a shell variable authoritative or adding another state owner. No frontend layout or business capability changes.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=max
Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

APPROVE，精确 head 0c3c68b6f15df87d54fc5203ee0e264bd974fe97，没有未解决的阻塞发现。

动机 / Motivation

在原生 Codex 中显式启用差分上下文、且工具输出或保留历史预算较小的用户与验证维护者。 首个 guard 固定返回64KB，而现有输出上限只允许用于后续读取；首帧截断后，模型可能连续页命令也拿不到。 同一显式参数现可限制已启用原生 guard／context-project 的完整首帧，并在返回的续页与完整恢复命令中保留；完整字节、digest、游标及原结算仍独立核验。 不启用差分功能、不增加宿主历史容量、不改原决策或配额权限，也不把本地脚本化 provider 当作真实模型或正式发布资格。

改动思路 / Approach

复用既有 reader 的完整stdout预算和native activation owner，只把原显式上限传至首帧；不以环境变量控制输出、不新增协议、状态或第二次准入。 本PR交付已有原生上下文入口的显式小输出限制、真实fixture预算匹配和对应双语说明；同时修正两个当前发布回归的null/EOF断言与已合并内容的发布说明。 首帧和后续分页使用同一个完整stdout预算，先验证现有原生启用范围，再执行原准入／读取。省略参数仍是64KB；显式较小值只改变本次表示，不改变源内容、宿主历史配置、lease、配额或原结算。原生模型必须从实际收到的完整首帧发现续页，不能读磁盘上的正确副本来掩盖模型输入截断。

具体改动 / Changes

按 base f40d2930dea9426c2e86daef37d31bce7d4d4ab5 的 docs/reference/required-work-context.md 核对 COMPLETE-STDOUT、ORIGINAL-RECOVERY、EXPLICIT-ONLY。三个生产文件复用已有CLI参数、activation和reader：启用的JSON should-run/context-project接受正上限，完整metadata／换行计入上限，返回的续页和完整恢复命令保留同一cap。复制恢复命令而不改sealed capture。显式off、未启用或非正值在quota authority之前拒绝；原默认64KB和环境声明无关。

关键代码讲解

• native_context_capture_view：Clamp whole-frame cap through existing reader;copy returned recovery command.
• handle_native_context_command：Validate allowed JSON enabled scope and positive limit before quota authority.
• project_saved_context_capture：Pass caller cap for original saved projection without collecting another decision.
• test_native_known_history_budget_keeps_larger_batches_and_recovery：Keep10k history/eight-read batch;cap every serialized page and match UTF8 offsets.
• test_native_context_cli_without_retention_stays_full_and_preserves_receipt：File/SQLite complete original readback,explicit7900/default64k,one receipt/zero spend.

测试没有把预算上调来跳过失败。显式10,000 token历史和八次批量读取保留，每份serialized输出4KB，UTF8 offsets按相同payload上限规划；所有batch元素都携带cap。生成的20,000 token宿主配置仍使用64KB输出，首帧和续页都使用该profile声明的直接工具预算。未知历史继续一次一份、有界读取，原2,000 token工具限制和native deadline保留；原256字节短读和大页场景仍执行。包装shell的pipefail与tee保持原结构，cap放在producer命令上，避免把整条pipeline重新shlex序列化成另一条命令。

File/SQLite正例分别重建完整UTF8 bytes和digest，重复guard仍只有一条heartbeat receipt、零spend且source不变；saved context-project读同一原capture，不产生新guard。真实native独立byte map要求所有原字节连续到齐，包含最后的停止条件，恢复后仍只有原来两个guard。未知依赖断言接受typed owner的null，must_attempt=true和全部轮询上限3保持；EOF只删空白。发布说明补入实际已合并的7510自claim默认及12位社区贡献者，未把这些文档当发布凭证。

PASS：Exact head:complete native/CLI/lifecycle/dependency cohort 280passed/0skipped;required Ruff,mypy,syntax,9path public boundary,12surface bilingual usage and risk-based premerge passed;source unchanged/cleanup successful. CLI help/manpage预检也通过。CI依现有Goal策略未查询。旧的低预算失败、错误pipeline尝试及已中断的旧head整模块运行均保留，不计为通过。

对主干的风险 / Risks

主要风险是cap仅限制内容而metadata仍溢出、后续命令丢cap、分页缩小却保留旧offset，或把磁盘存在／EOF当成实际保留。现有完整stdout reader计入所有framing，真实File/SQLite与实际Codex下一次输入逐字验证，并保留digest、原source、unique receipt、零重复guard/settlement反例。没有新协议版本或兼容分支；旧capture/receipt格式与默认路径保留，回退可撤销本PR。每条命令的cap不改变宿主总历史或整批预算，也不设置长期宿主偏好；不支持的环境与外部真实provider仍需独立资格。

我的整体评价 / Verdict

复用既有 reader 的完整stdout预算和native activation owner，只把原显式上限传至首帧；不以环境变量控制输出、不新增协议、状态或第二次准入。 本PR交付已有原生上下文入口的显式小输出限制、真实fixture预算匹配和对应双语说明；同时修正两个当前发布回归的null/EOF断言与已合并内容的发布说明。 架构上这是现有native IO／presentation adapter的参数传递，控制面语义owner继续拥有准入、lease与结算；fixture本地工具预算选择复用一个helper，避免多个分支再漂移。相较增大历史预算、压缩任务指导或另起准入，该方案直接解决可复现的完整首帧缺口，并保持原实际输入的独立反例。没有UI布局或新业务能力，现有原生交互入口完成验证；不能由此推断App／Bot或长期研究收益已经资格。统一发布commit的全仓、安装、付费模型与公开渠道仍待资格；本PR没有新科学结论、实际长期收益或完整跨平台资格声明。

English verdict: APPROVE 0c3c68b6f15df87d54fc5203ee0e264bd974fe97. Explicit native first-frame limits, complete original recovery and default/off parity are verified through real CLI and native-input tests. Required local gates passed; integrated release, installed artifacts and paid-model qualification remain separate.