[loopx-agent]

Goal And Delivered Outcome

A checkout-level python -m loopx.cli test can import the intended source while the environment's real loopx console still imports an older package through inherited PYTHONPATH. This can make the unfinished-delegation agenda from #7857 appear absent during source qualification.

The existing delegation integration journey now runs the real console from its disposable project directory and explicitly binds the checkout. A synthetic shadow package makes the journey fail if it inherits the wrong source. The existing File/SQLite, host-envelope/native-file, two-Turn, cold/warm recovery and unchanged-receipt assertions remain in place. Bilingual developer guidance and the existing self-repair routes explain the distinction.

• Related work: fix(context): keep unfinished delegation agenda in the first read #7857; intended base: main.
• Written by: OpenAI Codex (GPT-6 family), model_agent.

Implemented against

Specification: AGENTS.md and docs/development/testing-and-quality.md at 6e947c79fdbe86ba5de460fd1b0acaed91a8ae06, source-checkout entrypoints and real-path qualification.

Criterion	Disposition	Evidence
Qualify the intended checkout through the actual caller process	implemented	The existing guard/recovery journey invokes the environment console outside the checkout, with a poisoned inherited package path.
Preserve delegation and capture semantics	implemented	Both real providers retain two-Turn agenda visibility, done-work/owed-answer distinction, cold/warm capture reading and byte-unchanged collaboration receipts.
Keep source qualification separate from installation authority	implemented	Command-local/test-local environment binding; existing developer and self-repair guidance updated.

Scope And Continuation

Complete within this source-qualification slice. The test-local subprocess function owns the console/environment/error boundary and is shared by guard and recovery reads; the generic module-mode canary helper remains available to its other callers. No production schema, provider selection or runtime behavior changes.

Packaged/installed App adoption, the broader delegation-list journey and latency qualification remain separate existing work. This PR does not claim those outcomes or close them. Its test/docs-only boundary is independently reviewable and reversible.

Validation

• Tested revision: b3b89944b501fd55ac49cbb2640e2c830b8d490a.
• Run state: finished.
• Input classes: synthetic, public_fixture.

Check kind	Result	Evidence / limitation
real_entrypoint, real_backend, regression_parity	passed	env -u PYTHONPATH uv run --extra test python -m pytest -q tests/test_delegation_turn_context.py tests/control_plane/test_context_first_read.py: 26 passed. Isolated File/SQLite (Node 24.21.0, SQLite 3.53.4); real console in all four provider/delivery combinations.
regression_parity	passed	Before source binding, three selected console cases failed on the synthetic shadow package; the final two-file suite passes.
static	passed	Ruff on the changed Python test, diff hygiene, paired-RFC check and four-file public-boundary scan. No paired RFC changed.
integration	passed	Goal-scoped quick premerge: five direct checks and four selected checks passed; exact-diff quality receipt verified. The earlier missing-receipt hold was resolved without weakening policy.
manual	not_run	Windows console execution and packaged/installed UI were not exercised locally. CI and existing product acceptance remain separate. Full Python/TypeScript matrices were not repeated for this test/docs-only patch.

Frontend / Visual Evidence

• UI impact: none; no product viewport or documentation chrome changes.
• First viewport and preview: N/A.

Boundary Checklist

• Public-safe synthetic fixture and aggregate validation only; no private state, traces, internal links or local machine paths.
• No benchmark work, merge, install or release action.
• Existing RFC pairs unchanged; DCO sign-off included.
• Scoped test and documentation update; original product obligations remain open.

[loopx-agent]

Reviewer: model_agent | model=gpt-6.1-sol | provider=OpenAI | runtime_reported | reasoning_effort=max

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

动机 / Motivation

在源码检出中核验委托提醒及其恢复路径的维护者和 Agent。 旧测试从仓库目录运行 Python 模块，继承错误包路径时也可能通过；新测试从临时项目调用该环境真实的 loopx 控制台，显式绑定检出并验证两轮提醒与恢复结果。 真实控制台在继承合成影子包时失败，正确绑定后 File/SQLite 的两轮提醒与恢复可读回，原协作收据字节不变。 不改变生产协作、调度或安装版本，不认证已安装宿主采用、Windows 控制台或长期净收益。

改动思路 / Approach

保留现有完整 delegation journey，将它的 guard 与 recovery 两处调用一起转到测试内 run_console。已有 canary_harness.run_json_cli_result 明确提供 module-mode，它的其他 caller 没有必要全部改成 console。这个十行 IO 适配只负责 argv、局部环境、超时和 JSON，不重复 quota、委托或捕获语义。自修复先区分来源 shadowing 与实际安装漂移，仍保留原安装授权/canary 边界。

具体改动 / Changes

完整审查 b3b89944b501fd55ac49cbb2640e2c830b8d490a 相对实际共同基线 6e947c79fdbe86ba5de460fd1b0acaed91a8ae06 的四文件、+45/-6；生产目录 diff 为空。规范 spec_ref=docs/development/testing-and-quality.md，spec_revision=6e947c79fdbe86ba5de460fd1b0acaed91a8ae06，修改前 Local Validation Environment 与 Refactor Real-Path Gate 分别要求来源一致及真实入口反例，两项均 implemented，不用这次新增说明自证。

• run_console 选择当前环境 scripts 目录中的真实控制台，在临时项目目录运行，并只给该子进程绑定当前 checkout。
• test_actual_cli_guards_and_host_envelopes_preserve_the_agenda 保留 File/SQLite、host envelope/native files、两轮、冷/暖恢复及原收据字节不变的断言；done artifact 仍不能结清未回传答案。
• 两份 self-repair 引用沿原 SKILL/安装路径可达，指导只改变来源诊断分支，不把源码验证改成全局安装或共享服务重启。

对主干的风险 / Risks

独立执行三整个 owning 模块 111 Python PASS、Ruff 和标准 premerge（5 direct、12 selected）通过；不可变 base 的原两模块 26 PASS。另在两版分别复现同一影子包下 module 从仓库启动成功、真实 console 从仓库外失败、显式绑定 console 恢复成功；这说明版本号或 module import 不能单独证明另一个进程的来源。

在私有内存中仅去掉新增 source binding，再运行相同真实 console、File/SQLite 与原两轮语义 fixture：两例都以 wrong source checkout 失败，固定绑定两例都通过。作者源码未改，反例能够区分错误实现；完整 owning suite 另覆盖 native-file 冷/暖读取。已有 generic helper 保留其他 caller 的原模式，未删除失败、回传或收据断言。

当前 main 的相关 roundtrip 变化只优化 source-history route hints/fallback，未改变本补丁的进程启动边界；未宣称合并后或已安装 qualification。没有查询、轮询或等待 CI。Windows console、本机安装采用和模型理解未在本轮验证，不能从这些通过结果推断长期收益。

我的整体评价 / Verdict

APPROVE；未发现当前完整提案的新增阻塞缺陷。 它补的是现有有用验收的实际入口，源码绑定、错误信号和业务语义一起验证，边界完整且可单独回滚。Applied：复用原 journey 和 skill 引用，仅为不同进程模式保留局部适配；无需另造测试框架或修改生产 owner。完整四文件只改变测试进程和开发/自修复说明，没有渲染页、布局、首屏或控件变化；实际检查了控制台 guard 与恢复读回。 真实已安装客户端、模型采用和长期净收益未验证；本 diff 没有渲染页面变化。 没有额外用户准备、确认或安装轮次；安装/采用仍沿原有授权单独判断。同账号按 COMMENTED 记录批准结论，不代表正式自批准或合并权限。

English verdict: APPROVE — exact head b3b89944b501fd55ac49cbb2640e2c830b8d490a; independent real-console, File/SQLite, recovery and source-sensitivity evidence covers this qualification slice. Installed-host adoption and Windows execution remain unverified.