[loopx-agent]

The selected-work requirements regression compared a default heartbeat with a TurnEnvelope generated using a different CLI binding. Both File and SQLite failed at complete-context equality before the downstream Host and readback assertions could run.

Bind both callers to the same source-checkout CLI, assert that the emitted read command retains that executable and its --cli-bin binding, and execute the emitted argv unchanged. Complete-context equality, full requirement/stop text, fresh readback, missing-Todo and source-failure assertions remain in place. This changes one test file.

The exact-head code review publishes an APPROVE code verdict as a COMMENTED review and reports an independent complete-module 10 passed. This applies to the single-file repair at 7e4e04d475ae943a8ad7e386c6902020546b664f; complete repository and merge qualification remain outstanding.

The natural CI run failed at kernel static admission: the existing Stage 2C lock lacks declared tiktoken==0.12.0. The downstream test-shard job was skipped. Main shows the same missing dependency. The scoped lock repair is proposed separately in #7185, which is not yet adopted or qualified as a complete CI pass.

Author validation at 7e4e04d475ae943a8ad7e386c6902020546b664f, against the original 385771e56371315976ebe1166a3c07fea8b539df baseline:

• Immutable main baseline 385771e56371315976ebe1166a3c07fea8b539df: both affected cases reproduce the binding mismatch. Final complete module: 10 passed, including real File/SQLite CLI and recovery paths.
• Repository CI-scope Ruff, configured mypy (17 source files), CLI output budget/base-head differential, public candidate/diff scans, staged bilingual check and whitespace check passed. The semantic advisory found no supported new vocabulary carriers.
• Full repository Python qualification has not been rerun at this head. The existing whole-suite failure evidence remains historical failure evidence; these focused passes do not qualify the full suite or hosted CI. The previously recorded change-quality receipt remains historical non-passing evidence; it does not qualify a newer target revision. Full required qualification is incomplete.

Goal-bound standard premerge ran five direct checks, all passing, with zero catalog selections for this test-only path. Its final aggregate is FAILED because it enforces the non-passing change-quality receipt. This is a completed author proposal for review, with merge qualification held.

[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

开发者运行默认 heartbeat 的完整需求回归时，需要验证生成的命令能从同一源码环境读回当前任务。

旧测试把默认 CLI 的 TurnEnvelope 与绑定源码 CLI 的默认提示词比较，File、SQLite 都先在完整上下文比较处失败；统一绑定后，原有比较保留，后续 Host 和真实命令读取得以执行。

已独立验证完整模块10例通过，完整要求、停止条件、后续更新、缺失任务和来源读取失败都仍有真实调用覆盖。 本次不改变运行时、用户默认调度、授权、生产状态或提示词内容，也不宣称全库测试、CI 或合并资格通过。 完整仓库资格仍未验证；最新主线上的精确范围质量回执已过期，须由原提案的维护者重新资格验证后再判断合并。

这是修复既有回归覆盖的有用增量，而不是新的产品功能。不能只删掉失败的完整比较，或在测试里代造另一条可用命令：这都会放过实际生成命令的错误。所用历史经验只帮助确认测试准备之后仍有真实 Host 和读回结果，不增加本次内部测试修复之外的交付义务。

改动思路

在既有测试中复用一个源码环境的 CLI 路径，让两种生产入口使用相同绑定；之后验证生成命令的可执行文件与绑定参数，并原样执行该参数列表。当前改动只有一个测试文件，7行增加、4行删除；不新增生产所有者、状态、CLI 选项、版本分支或烟测模块。最小修复已经留在原集成回归内，继续保留完整上下文、完整要求和失败路径断言。这样测试真正验证调用者将收到的命令，维护者也不需要维护第二份命令重写逻辑。这个简单改动用文字即可说明，无需额外流程图。

具体改动

1. guard 与默认 heartbeat-prompt 共用 cli_bin，保留原有 envelope["work_context"] 的完整相等比较。独立在不可变主线 385771e56371315976ebe1166a3c07fea8b539df 重跑两个受影响案例，File、SQLite 均在完整 sources 命令绑定处失败；当前完整模块10例通过，并继续执行比较之后的断言。
2. 生成命令前缀断言 要求实际可执行文件和 --cli-bin 都等于调用者指定值。要求来自已有源码环境绑定和当前原生读取合同，而非把头部实现本身当成正确性定义。
3. read_detail 与缺失任务调用直接执行原始 argv，并保留同一 cwd、registry、runtime 与 PYTHONPATH。旧写法替换 argv 的首部，可能掩盖生成的可执行文件错误；新写法实际走完整命令。原有 Task/Goal 变化后的新鲜读取、Host prompt、缺失任务和来源文件读取失败仍保留。

合同依据为更改前 385771e56371315976ebe1166a3c07fea8b539df 的 docs/heartbeat-automation-prompt.md，对应原章节 Two Prompt Layers、Static semantics retained across hosts：使用本轮当前合同，向 Host 提供完整要求、验收及停止条件，并让来源失败可见。另核对同版本 docs/development/testing-and-quality.md 的源码环境说明；测试没有改写这些义务。检查相邻 selected-work、heartbeat、TurnEnvelope 和 Todo 覆盖后，没有新增同形副本；最近20项公开PR与完整30项当前队列也没有给本次修复建立重复脚手架的理由。

对主干的风险

最强回归是生成命令指向错误的 CLI，而测试偷偷用已知可用的模块调用替代。当前前缀断言与原样 argv 执行共同覆盖它；独立基线反例证明原失败存在，修复没有降低完整要求。完整模块使用真实 File、SQLite、CLI 子进程与 Host 构造，但 Goal 和材料为隔离合成数据，不证明实际模型采用或全库无回归。

本次独立验证：完整模块10例、仓库CI配置范围的 Ruff、配置内17个文件的 mypy，以及固定基线的 CLI 输出预算检查通过。标准 premerge 的5项直接检查通过，测试文件未选中 catalog/risk smoke；其最终聚合仍为 FAILED / quality_stale_receipt。当前实际目标主线已到 7129efcdef7a7110bd49b760b7de8359544889ec，质量回执必须按最新 base/diff 重新资格验证，不能用此评审或五个直接检查冒充合并通过。全库Python资格未重跑，旧失败证据仍只是历史证据，不声称已修好或已被独立归因为无关。当前配置不查询、不轮询、不等待远程CI。

我的整体评价

APPROVE，仅针对 7e4e04d475ae943a8ad7e386c6902020546b664f 的完整单文件测试修复。 原有集成回归已经拥有该边界，在原处修复绑定并执行真实命令，比另建测试层更直接。后续当前要求变更、旧观察保留和来源失败仍被验证，不新增重试、扣额或生命周期规则。测试准备仍复用已知 registry、runtime、CLI 和任务，没有新增用户配置、确认或补录；实际产品调用路径没有行为变化。

仅修改 tests/control_plane/test_selected_work_requirements.py；检查了生成命令、Host 和读回调用，生产 CLI、页面、Lark 与可见提示词均无改动，因此不需要新截图。 没有声称新增产品页面或实际模型验证。完整仓库资格及最新范围 CQR 仍是原提案维护者的后续门；本评审不自合并，不清除其他讨论，不把测试修复等同于整个长期目标完成。

English verdict: APPROVE — 7e4e04d475ae943a8ad7e386c6902020546b664f. The existing regression now uses one explicit source CLI binding and executes the emitted argv unchanged without weakening full-context, Host or fresh-readback assertions. The immutable baseline reproduces both provider failures; the current module passes all10 cases, with configured local static/budget checks passing. Merge remains held by stale exact-scope quality qualification; the full suite and remote CI are not claimed.