[loopx-agent]

Goal And Delivered Outcome

The checkpoint handoff regression fails in both native-context modes before it can check history: its fake guard requires --decision-output-root, while the shipped preflight now prepares a capture with --decision-output-dir. It also expects an immediate projected view where the native host now presents the retained original later.

Update that synthetic transport to use the existing prepared-reference formatter and verify the current contract. All four context/participant combinations retain their history assertions; two corrupt-digest cases prove that a substituted reference is rejected before continuation. No runtime, scoring, execution authority or benchmark workload changes.

Author Declaration

• Written by: model_agent · OpenAI · GPT-6.
• Specification: the existing checkpoint preflight and native presentation owner.

Criterion	Disposition	Evidence
Preserve the original guard bytes, digest and private file mode	Implemented	Initial and resumed captures remain distinct and immutable; corrupt references are refused
Defer native presentation without replaying the guard	Implemented	Prepared flags, original capture reference and one guard per explicit preflight are checked
Defer full history only for a native participant	Retained	All four existing context/participant combinations remain covered

Scope And Continuation

Complete within this test-fixture scope. The broader 2.3.0 release continues through its existing qualification and publication gates; this PR does not certify that release.

Validation

• Tested revision: 6469c07491cf6260c6cd94f7ccc9b34dd3df550b.
• Run state: finished. Inputs: synthetic fixtures.
• python -m pytest -q benchmark/tests/test_checkpoint_handoff.py benchmark/tests/test_context_delivery.py: 76 passed, 4 skipped. The skips cover Linux writer process groups on macOS.
• The unchanged test failed in both native-context modes during the release sweep. The updated six-case preflight matrix passes and rejects both injected digest substitutions.
• CI-scoped Ruff, public-boundary scan and diff hygiene passed. Exact-scope change-quality receipt verified.
• Standard premerge: 5 direct checks and 10 selected checks completed, zero blocking check failures. Its benchmark-sensitive manual hold remains for independent review. The maintainability ratchet reports inherited metric debt in loopx/extensions/lark/private_conversations.py; premerge classifies it as an advisory outside this one-file diff. The first premerge attempt lacked repository npm dev dependencies; after installing the lockfile dependencies, the vocabulary check passed. No paid model or benchmark experiment was launched.

Frontend / Visual Evidence

UI impact: none; only a synthetic Python regression fixture changes.

Boundary

One test file, signed-off commit, no paired RFC change. No credentials, private state, local paths, raw trajectories or verifier output are included.

[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)

未发现阻塞问题；结论绑定 6469c07491cf6260c6cd94f7ccc9b34dd3df550b。

动机 / Motivation

维护者在运行 checkpoint handoff 回归时会被过时的假 guard 阻断。同一原生上下文 preflight，旧夹具查找不存在的 --decision-output-root 而失败；新夹具接受已交付的 prepared reference，继续检查原文件和历史边界。四种原组合继续到历史/来源断言，两种被替换的摘要在恢复前拒绝。不认证完整 2.3.0 发布、平台资格、真实模型或长期收益。

改动思路 / Approach

生产 preflight 已经把原始 guard 保存下来，prepared reference 只携带身份和摘要，真正呈现由原生工具调用完成。旧夹具还把输出当作即时 projection，因此只改参数名并不足够。复用已交付的 preparation formatter，独立保留来源与拒绝断言；测试修复不另立运行规则。本 PR 到现有回归夹具为止；完整发布与真实宿主资格由原 owner 继续。没有改动运行时来迎合旧期望，也没有另建一个形状相同的测试入口。

具体改动 / Changes

关键代码讲解

• test_preflight_defers_history_only_when_capture_will_verify_it 保留 native context 与 native participant 的四种原组合，另外覆盖两种摘要被替换的情况。合成 guard 在原生模式检查 prepared/deferred 标志，并使用产品已有 formatter；非原生模式继续使用原输出路径。后续断言独立检查真实源文件、SHA-256、私有文件权限和原正文，避免仅让 formatter 自证。
• capture_before_replan 的生产代码在 base/head 完全相同：读回原文件后，先把返回摘要与原字节比较，再调用历史恢复 gate。有效 native participant 才推迟 full-history 检查；其他组合仍要求立即核完整历史。
• preflight_context_instruction 同样未改，继续让 context-project --initial-read 在实际呈现时运行。两次显式 preflight 各执行一次 guard，保留不同的原始 capture；第二次不能改写第一次的字节。本测试不声称第二次 preflight 是原 guard 的重放。

规范使用改动前 benchmark/README.md，修订 a02be126957953b7a61796e3b99fcfe578eb5174。该文无编号，定位标签 Original quota guard 对应一次原 guard 与原身份恢复，Native retained history 对应完整历史验证边界，Immutable source recovery 对应原字节及替换拒绝；这三项在本夹具范围 implemented。完整 diff 为现有一份测试 +33/-11，产品 preparation、checkpoint 与呈现模块的 Git blob 均与基线一致。

对主干的风险 / Risks

独立 PASS：产品 decision-capture/research-checkpoint 两份测试 106 通过、11 项因未配置显式真实 Host 而跳过；两份目标测试 76 通过；四项 Linux writer process-group 测试在本次 Darwin 环境跳过。旧基线同一原测试为 2 失败 / 2 通过，两个 native case 都因缺少旧参数失败。独立反例把生产摘要比较单独停用后，恰好两种篡改 case 报 DID NOT RAISE ValueError，另外四种通过；原 head 六种均通过。第一次 mutation harness 错误复制了 globals，导致六项一起失败，已保留为无资格结果，修正后的对照才用于结论。

合成 guard、selector 和历史 gate 限定了证明范围：这里验证真实 preflight 的传输、原文件和分支，不证明真实 Codex 历史已经重建。test_context_delivery.py 另经实际 CLI 验证原捕获的只读恢复。Ruff、差异及公开边界扫描通过；标准 premerge 完成 5 项 direct 与 10 项 selected，零阻塞检查失败；benchmark_sensitive 人工 hold 仍在。其 ratchet 的 Lark 模块规模债务被原生门判为 inherited advisory，未改记为修复。移出我创建的工作区日志后，clean plan 的十条命令完全相同，复用原检查。未查询或等待 CI，未执行模型或正式 benchmark 工作负载。

我的整体评价 / Verdict

这是一项完成的回归夹具维修，复用了现有 owner，并用可失败的反例保留了原约束，范围与成本相称。long_horizon 的生产行为 preserved；本 PR 只修复现有 checkpoint 夹具，未改生产路径或呈现。真实宿主、安装版与模型采用未验证；此测试改动不改变页面。没有新增共享 vocabulary、权限或默认策略，local test 参数不授予执行权；完整发布及平台门继续独立判断。

English verdict: APPROVE for the test-fixture repair at 6469c07491cf6260c6cd94f7ccc9b34dd3df550b; release and native-platform qualification remain separate.