[loopx-agent]

Goal And Delivered Outcome

Native context delivery added a small-output boundary that ordinary full-packet output did not have. Without a shell environment declaration, continuations fell back to 8192-byte reads under a 7900-byte output cap; the initial guard always returned at most 3072 content bytes. Combining pre-quota text and that first JSON page also prevented the native hook from recognizing the delivered range.

This repair uses the existing 64000-byte profile for inline and continuation output regardless of that environment declaration. Packets that fit finish inline. Complete newline-delimited JSON records in mixed shell output reach the existing trusted-capture verifier. No additional split-command instruction is needed.

• Outcome basis: maintainer-requested code repair of mixed-output recognition and unnecessarily small pagination.
• Intended base: main; reviewed baseline 787f15e82283ae4cd590dc0ec91d88b09db4e122.
• Observable change: identical synthetic CLI inputs of 46959 bytes require 7 → 1 reads; 468159 bytes require 65 → 8. Mixed stdout now establishes the complete baseline and the next projection is delta. These are transport observations, not model-token or whole-cost savings.

Author Declaration

• Written by: OpenAI Codex model agent (GPT-6.1 Sol).

Implemented against

No separate written specification; the maintainer request described above is the basis. Existing capture integrity and native lifecycle boundaries are retained from docs/heartbeat-automation-prompt.md at the reviewed baseline.

Criterion	Disposition	Symbol / path	Evidence
Repair mixed stdout in code	implemented	_context_pages, existing native hook	Complete/multiple records accepted; embedded, quoted, truncated, corrupt and failed output rejected; original-range recovery tested
Restore reasonably large default reads	implemented	native_context_page_budget, JSON reader, generated recovery commands, inline guard	Missing/invalid/old environment declarations retain 64000-byte profile; small captures complete inline; paired CLI comparison
Preserve original capture and authority	implemented	Existing capture verifier and typed lifecycle owner	Exact bytes/digest/cursors, scope isolation, no extra admission/spend, explicit small-budget recovery, File/SQLite integration

Scope And Continuation

Complete within this transport-repair scope. Python remains the existing native IO adapter; shared lifecycle/state decisions remain with the current typed owner. No new module, command flag, state vocabulary, scheduler, provider or UI is introduced. The obsolete guard-only-small instruction is removed, and the existing English/Chinese required-work-context contract now describes the same defaults; configured-profile instructions otherwise remain byte-identical across six scenario/location combinations.

The explicit JSON recovery reader's default also changes to the large profile. Explicit smaller page/output budgets retain their existing behavior for constrained recovery. Ordinary successful quota output and reuse activation/default-off behavior retain their existing owners. Large stdout does not raise a host's direct-output or retained-history limit.

Validation

• Tested revision: a88619fa02f3269db30e25c044ed7253640477cd.
• Run state: finished.
• Input classes: synthetic, public_fixture.

Check kind	Result	Evidence / limitation
unit	passed	Lifecycle and JSON-reader suites: 196 passed
integration	passed	Seven adjacent context/host/heartbeat modules: 260 passed; 54 existing explicit native-Codex opt-ins skipped
regression_parity	passed	Actual immutable-base/source-after CLI on identical synthetic inputs: 7 → 1 and 65 → 8 reads; mixed-output baseline failure reproduced before and repaired after
real_entrypoint	passed	Isolated final wheel repeats complete inline/large reads and full → delta; 1752 Python/TypeScript modules match source, wheel and isolated installation
static	passed	Required Ruff trees, configured strict mypy (17 sources), semantic advisory, bilingual diff, whitespace and nine-file public-boundary scan
integration	passed	Standard Chat build and wheel build; existing CLI-output-budget regression smoke
manual	passed	Six configured-profile comparisons retain all other instructions/configuration exactly after retiring the obsolete small-guard clause
integration	passed	Standard premerge: 6 direct checks and 19 selected checks, zero failures/manual holds; exact quality receipt cqr_90c4e28e371834b888c2 verified valid

Initial validation failures are preserved: default-budget/legacy handling regressions were repaired, old small-inline assertions were updated to the intentional behavior, and larger saved-page fixtures still exercise continuation/restart with exact bytes. An initial stale-frontend wheel build was repaired by the normal Chat build before the final successful wheel, without a packaging bypass.

Coverage: source and isolated package/CLI transport. The skipped opt-in native-Codex matrix, live App installation, actual HTTP/model-input retention, full hook/IO/billing costs and long-horizon net benefit are not newly qualified by this PR. No model campaign was launched.

The exact-base comparison repeats the same 7 → 1 / 65 → 8 result. A paired disabled-path check preserves the complete off projection and inert hook except for the intentionally enlarged recovery command (including its nested reference). Final documentation refinement changes no Python/TypeScript/test bytes from the tested revision.

Frontend / Visual Evidence

• UI impact: none. Existing native hook/configuration/CLI journey only; no rendered layout changed.
• Before / after / first viewport: N/A.

Type of Change

• Bug fix
• Documentation update
• Test update

LoopX Area

• Control plane
• Host or runtime integration

[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)

动机 / Motivation

没有发现本准确 head 的未解决阻塞问题。使用原生上下文读取的 Agent，尤其是合并前置命令输出与首个上下文页的调用方。

约 47 KB 的完整包原先需要 7 次读取，现在一次返回；约 468 KB 的混合输出原先需要 65 次且无法建立基线，现在 8 次读完并进入下一轮差分。同一不可变输入的字节、摘要和连续游标完整保留，减少分页往返，混合输出无需新增分开执行的提示。本次只修复传输读取；不证明模型保留、完整 token 或费用收益，也不升级运行中的 App。原长程目标中的真实模型输入、全部调用成本和可比任务收益仍未验收。

改动思路 / Approach

复用现有分页器、可信捕获校验和 TypeScript 生命周期 owner；新增代码只识别完整 JSON 行，预算在已有 owner 内统一，未增加状态或授权分支。交付首屏、续页、恢复命令、混合输出识别与相符文档这一完整传输切片。首屏和续页都由现有分页器按完整 JSON stdout 大小收缩正文；基线是否可信仍由原摘要、实际字节、作用域及生命周期判断。它不会因为找到一行 JSON 就相信其中自报的路径或 EOF。

以下是实际 CLI 的相同输入比较，基线为 787f15e82283ae4cd590dc0ec91d88b09db4e122，审阅 head 为 a88619fa02f3269db30e25c044ed7253640477cd，没有 shell 预算声明：

输入及触发	基线	当前 head
46959 字节普通完整包	7 次读取	1 次，首屏 EOF
468159 字节、首页前有其它输出	65 次，下一轮 full	8 次，下一轮 delta

仅加“请分开执行”指令会把框架缺陷留给调用方；无界返回整包又会增加截断风险。现在复用既有 64000 profile，并保留同捕获的显式小预算恢复，是与已授权目标相称的取舍。没有增加手工准备、确认或新配置步骤。

具体改动 / Changes

本次新增需求没有另立已接受的书面规格，依据是维护者当前的代码修复、大页及合并请求。补充核对的既有来源为 docs/reference/required-work-context.md 与 docs/heartbeat-automation-prompt.md，不可变基线 787f15e82283ae4cd590dc0ec91d88b09db4e122；新预算由该请求明确更正，不能用修改后的说明反过来证明实现正确。

9 文件共 +195/-88，生产代码 +42/-22，测试 +93/-10，文档 +60/-56。评审已把仍写旧 3 KiB/7900 默认的中英文协议同步修正；保留历史测试条件、完整源/权限/新鲜度、恢复与不确定性语义。

关键代码讲解

• _context_pages 先保留整个 JSON 解码，再识别完整物理行。嵌入文本、JSON 字符串和截断记录不能当页；最终仍走可信原捕获逐字节校验。
• native_context_capture_view 使用同一首屏/续页预算。完整 stdout 的转义、指令、恢复命令及换行都计入上限，能容纳的包直接结束，不能容纳就保留原游标续读。
• read_context_capture 把省略参数的 JSON 默认改为 64000/64000。Markdown 4096 和显式更小预算保留；过小 metadata 预算报同捕获错误，不产生零进展页。
• native_context_page_budget 删除缺少环境声明就回小页的分支；生成恢复命令也复用原大页配置。

验证为 PASS：196 项 lifecycle/CLI，260 项相邻 resume、retention、saved、host/heartbeat；54 项既有显式原生 Codex opt-in 未运行，单列 SKIPPED。产品源码和测试在最后文档修订后逐字未变；隔离 wheel 已重复 1/8 页、下一轮 delta，1752 个 Python/TS 模块与源码/安装一致。规定 Ruff、mypy17、公开边界和双语/差分检查通过。

对主干的风险 / Risks

64000 字节只是请求和输出呈现预算，不能提高宿主硬容量。最强反例是 CLI 与 hook 都完整，而下游保留历史被截断；当前合成工具测试不能否定它。已观察的缺页、坏字节、错误 scope 或失败退出不建立基线，补读同一原捕获缺失范围可恢复，不重准入、扣额或刷新权限。必要时显式减小 page/output 预算；不会因本地完整 EOF 就声称模型已保留。

复用仍默认关闭。真实 base/head 关闭态比较验证：响应完整字段与指引相同、hook 为 {}，不创建 journal/authority；顶层和嵌套恢复命令均有意从 4096 改为 64000/64000，这个全局 reader 默认变化已经披露。已配置的六组生成指令只去掉过时 small-guard 句，其余配置/指令逐字相同。没有新增 actor、协议、权限或费用门；既有 TS owner 继续负责状态。

语义与 CI 对齐

本次复用既有词表和边界，仅按维护者方向调整输出默认及完整记录框架。首轮验证中的预算/legacy 处理回归已修；旧小页预期改为完整原包断言，续页夹具增大后仍检验精确字节和重启，未删除恢复 oracle。旧构建失败经正常 Chat build 后完成打包，没有豁免；最终准确 head 的标准预合并 PASS：6 项直接检查、19 项选中检查，零失败及人工留置；准确质量回执 cqr_90c4e28e371834b888c2 有效。当前 wait_for_ci=false，未查询、轮询或等待远端 CI。

我的整体评价 / Verdict

这是原长程任务的 justified_increment：减少重复完整读取和工具往返的具体路径已有实证，long_horizon 与 user_experience 在此传输边界 improved，尚不能认证整体任务收益。机制与问题相称；没有新模块、flag 或第二状态 owner，兼容保留对应已有显式小预算/持久捕获消费者。检查了原生 hook、CLI 和配置入口；没有改动渲染布局，因此不需要新的页面截图。正式 App 安装、下游模型输入和长期阅读效果未验证。

准确审阅 head：a88619fa02f3269db30e25c044ed7253640477cd。无阻塞 finding，作者自评采用 COMMENTED，不冒称 GitHub formal approval、第三方独立验收或合并授权。合并仍须最终本地预检与准确 head 的原生 readiness。

English verdict: APPROVE a88619fa02f3269db30e25c044ed7253640477cd for the bounded native transport repair; model retention, live installation and complete cost benefits remain unverified.