[loopx-agent]

Goal And Delivered Outcome

#7687 已合入实际 descriptor.command / next_command 分页。本 PR 保留其后的一个独立缺口：完整内容的 byte-map oracle 允许相同字节重叠，因此成功路径中重复读取一页仍能通过。

在既有原生 saved-context 夹具中记录每次捕获的预期 UTF-8 byte offset，要求下一页接续上一页，并核对实际内容长度与返回游标。普通与 progressive 分支统一复用现有 last_context_page，删除重复 JSON 提取和导入。通用 byte-map oracle 保留，其他恢复测试仍可验证合法的重复交付。

• Basis: existing context recovery contract and controller/executor C2.
• Observable result: 在同一生产代码与原时限下，故意重复第一页时，主干夹具通过；本提案准确拒绝 offset 0 与预期 59760 的冲突。正常完整模块全部通过。
• Scope: latest-main integration leaves one test file, +8/-9. No production reader, page default, prompt, schema, timeout or authority change.

Author Declaration

• Written by: model_agent — OpenAI Codex (GPT-6 family).
• Implemented against the existing native reader and saved-context contract linked above.

Criterion	Disposition	Path	Evidence
成功路径完整读取不重复或跳过字节	implemented within fixture	tests/control_plane/test_saved_context_project.py	Actual native repeated-page counterexample; complete source module
完整内容、digest、Stop、生命周期、恢复身份和恰好一次语义保留	retained	same file	All 145 current-main assertions retained; 148 final assertions

Bounded refactor: use the existing parser in both branches; keep only local per-capture cursor state with an independent failure oracle. No new framework or parallel decision owner.

Validation

Tested revision: 66b17f9fc621226c1ed9c3e4f2c10af80c244039, integrating main d3d9ad209d0a8f57a7582e0165f2fad19df26dcf. Run state: finished. Inputs: synthetic, public_fixture.

Kind	Result	Evidence and limits
real_entrypoint	passed	Complete saved-context source module: 43 passed, zero skips, 242.11 s. All six native profiles, including App server/machine/code mode/progressive and both worker modes. Actual Codex 0.163.0-alpha.11.1, deterministic local loopback inference and disposable File/SQLite authority; App 30 s and worker 60 s limits unchanged.
regression_parity	passed	Disposable exact-main/candidate native App-server fixtures inject the same redundant first-page read against identical current production code. Main accepts it; candidate rejects at the expected cursor assertion. The experiment verifies that injection actually occurred. It measures oracle sensitivity, not latency or complete cost benefit.
regression_parity	passed	AST comparison retains every one of main's 145 assertions; three added assertions cover page success and byte continuity. Original complete content/digest, lifecycle, Stop and no-replay checks remain.
static	passed	Declared Python CI Ruff scope, configured mypy 17 source files, staged/committed bilingual RFC check, diff check, DCO and public-boundary scan. The semantic advisory finds no supported new vocabulary carrier; its documented test/dynamic-analysis limitations remain.

The first disposable counterexample copies failed before execution because the existing fixture module was absent from their import path. Those preparation failures remain recorded; correcting only that experiment's import path produced the behavioral results above. They are not runtime failures or sensitivity evidence.

Exact-scope author quality receipt: cqr_a802a4a631eece2b4a5f, valid/pass for fingerprint a802a4a631eece2b4a5f0c458dfa37ed2848092bbf234615910344dad0c33c2b; safe-fix allowed=true/applied=false, blocker/warning/advisory=0. Author qualification is separate from independent review and merge readiness.

Standard source-checkout premerge: 5 direct checks passed, zero selected catalog/risk checks for this test-only diff, zero failures, skips, quality failures or manual holds. The exact current quality receipt is valid. Remote CI, independent review and merge readiness remain separate; no self-merge authority is granted.

Coverage boundary: this is source/native fixture validation. No new wheel qualification, installed App adoption, model understanding, sustained latency, complete outcome cost, remote CI or independent approval is claimed. #7687's final wheel evidence remains tied to its own revision. No live installation, model call or benchmark campaign was performed. C1/#7430 and C2's actual consumer adoption/full-cost acceptance continue under their existing owners; this test slice does not close them.

Frontend / Visual Evidence

UI impact: none; first viewport, before/after and preview: N/A (test-only diff). No frontend, Lark or CLI product operation is extended.

Type Of Change

• Test update
• Narrow refactor of an existing validation consumer
• Host/runtime integration coverage
• Public diff excludes private state, paths, raw captures, logs and credentials
• All proposal commits have DCO sign-off

[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)

Exact head: 66b17f9fc621226c1ed9c3e4f2c10af80c244039；不可变 base：d3d9ad209d0a8f57a7582e0165f2fad19df26dcf。

动机 / Motivation

维护原生上下文恢复验证的贡献者，需要知道成功路径是否真的按游标逐页读取。

旧 byte-map 允许相同字节重叠，重复读第一页仍会通过；新夹具会在成功路径的游标不连续时准确失败。

实际原生 reader 的重复第一页反例由旧夹具接受、新夹具拒绝，正常完整模块和原有断言保留。

不改变生产 reader、读取预算、合法重放、权限、模型或已安装 App 的行为，也不完成 C1/C2 整体验收。

这补的是现有验证的敏感性：完整内容的摘要一致，并不能证明成功路径没有反复读同一页。判断不是来自作者的通过数量，而是让真实 reader 重复一页后，对照两版夹具的实际行为。

改动思路 / Approach

复用既有 last_context_page，将连续游标断言留在当前成功路径夹具；通用 byte-map 继续允许合法恢复中的幂等重复。

当前切片只强化现有 source/native 测试的证据敏感性；没有生产改动、新夹具框架或新的产品义务。

同一 production reader 的重试或并发交付仍可合法重复；本 PR 没有把成功路径的局部游标要求提升为全局产品规则。也没有新建第二个 fixture 或完整 reader：每次新 capture 初始化一个本地期望 offset，读取后按实际 UTF-8 字节长度推进，再核对下一游标或 EOF 总长度。

具体改动 / Changes

完整 diff 只有 tests/control_plane/test_saved_context_project.py，+8/-9。普通和 progressive 分支复用既有 last_context_page，删除手工 JSON 提取和重复 import。AST 对照确认主干原有 145 条断言全部保留，当前共 148 条；这项机械检查仅证明断言保留，语义敏感性由下述真实反例证明。

关键代码讲解

• 原生测试 callback 先要求当前工具结果是成功 page。普通完整读取随后核对 offset == read_offsets[name]，用 len(content.encode()) 累积字节，检查 next_offset 或末页 total_bytes。它仍跟随实际生成的 descriptor.command 和 next_command。
• last_context_page 读取最新工具结果，支持现有普通及 code-mode 包装；最新结果不合法时不会用一个更早的成功 page 掩盖。generic byte-map、完整原始内容和生命周期断言保持不变。

spec_ref：docs/reference/required-work-context.md；spec_revision：d3d9ad209d0a8f57a7582e0165f2fad19df26dcf。

• Saved context cache / 已保存上下文缓存：implemented，选定完整读取时遵循真实游标、完整 UTF-8 内容和原摘要，恢复不重新准入或扣额。
• Selected capture details / 按需读取捕获详情：implemented，progressive profile 继续只读所需索引/字段；局部 EOF 不成为全量 retention 证明。

对主干的风险 / Risks

未发现当前阻塞性问题。 我在隔离复制的两版测试里注入相同的一次冗余第一页命令，生产代码保持完全相同：旧夹具接受了重复的实际 reader 输出，当前夹具在 offset 0 与已经累积的期望游标处拒绝。两个实验的注入标记均独立读回，避免把没有真正执行反例的设置失败当成证据。

正常当前模块 43 passed、0 skipped，覆盖六个实际 native profile，包括普通 App、machine、code mode、progressive 与两种 worker 路径；使用实际 Codex 0.163.0-alpha.11.1 和确定性的本地 loopback inference，没有真实模型调用。App 30s、worker 60s 限制未变。原内容、digest、Stop、生命周期、原 Turn 与一次结算/无重放断言仍执行。

最初两份树外 native 反例漏载了仓库的隔离 fixtures，在读页前因隐式默认路由冲突而失败；它们不是 mutation 证据。加载原有 fixtures 后，在同一生产代码上取得上述旧通过、新拒绝结果，实际模型输入中也分别读回两次 offset0。原准备失败、正常通过和预期断言失败均分别保存。

标准 premerge 5 direct + 0 selected 全部通过，零 failed/skip/manual hold，Ruff 通过；测试-only 风险选择没有额外 catalog/risk 项。已对照现有 checkpoint helper、同文件相关历史与完整当前 PR 队列：这里没有复制现有 smoke，而是强化已采用 caller 的不同 oracle。共享账号本身不证明同一 Agent，也没有据此作贡献限制判断。未查询或等待 CI。

我的整体评价 / Verdict

APPROVE 当前有界测试修复，delivery 判断为 goal_achieved，只指这一项 oracle 缺口。long_horizon 与 user_experience 为 preserved：生产行为没有改变，验证更能识别冗余工具工作；不能由此宣称模型更理解上下文或实际任务成本已经下降。

相邻简化已采用既有 parser，并将 clean-success cursor 留在本夹具；通用幂等恢复 oracle 和既有原生 profile 都有独立语义，应继续保留。仅改测试夹具，生产 CLI、网页、App 和视口均未改变。 已安装宿主、真实模型采用和长期完整成本收益未验证；没有需要补图的渲染表面。

English verdict: APPROVE — 66b17f9fc621226c1ed9c3e4f2c10af80c244039. The real repeated-page counterexample passes the old fixture and fails the strengthened cursor oracle; all 43 current native tests pass and all 145 prior assertions remain. Production behavior and broader C1/C2 qualification are unchanged.