[loopx-agent]

Ordinary quota projection can relocate a complete inline acceptance owner while leaving an existing acceptance_contract_ref pointing to the removed field. Retarget only aliases whose owner actually moved, including aliases visited before their owner. Complete requirements, source versions, original numeric types and caller-owned input remain intact; unknown or unmoved references retain their existing reader semantics.

The existing compatibility projector and its nested copy-spine helper own this repair. It adds no RPC, source read, schema, authority or capture rewrite. Parent #7430 remains unfinished: producer/consumer cutover, historical-reader exit and original-quota outcome/full-cost qualification stay with that work.

Written by: Codex (OpenAI GPT-6).

The branch includes the exact shared commits from #7699, including its one-line parsed-module type correction, to qualify the combined source. Prefer reviewing/merging that independent validation prerequisite first; then reconcile this branch with actual main before its own exact-head merge review. Both PRs currently target main. No duplicated scanner implementation is introduced.

Final combined head: 45834a3ff77896720f2cfa8a8eb2ea881bc0f1c2, tested against main a36e01c32b0a290a96466c766629a0b01850f823. Inputs are synthetic/public fixtures.

Check	Result	Evidence / limit
Reference regression / cost parity	passed, retained	Three immutable-base dangling-reference cases resolve on the repair, retaining complete criteria, valid Turn signatures, idempotence, unknown/unmoved aliases and numeric distinctions. The projector and these tests are byte-identical to 3769a605. Its same-fixture default/all-details call-count and serialized-input/result/output controls remain applicable; they do not measure complete costs.
Source integration	passed	451 cases pass on 4501ca198 before the sole final inventory annotation. CLI projector and reference-test blobs are unchanged at final head. After that annotation, 142 semantic production/binding/inventory cases pass again on final 45834a3ff.
Installed wheel	passed	Final isolated Python 3.11.17 wheel: 56 cases pass, all 1,898 product files match source. File/SQLite CLI selection, replan, recovery and admission use the installed package. Both changed module hashes and an installed registered-return probe pass; no foreign product-module origins remain. This is a disposable installation.
Complete static observations	passed	Independently frozen baseline/final analyzers inspect the same 1,769 SourceFiles and preserve all 137 production observations across 53 nonempty cases, including unknowns. Both corpus arms have zero nonempty literal observations; tests cover positive literals.
Static / docs	passed	Changed-file Ruff; normal-import mypy on both semantic modules; fresh-cache configured mypy (17 files); final vocabulary advisory and public-file scan. Full docs governance passes on 4501ca198; every changed documentation blob is unchanged at final head.
Standard premerge / author quality	passed	Final458: 6 direct checks and all 17 selected checks pass, no warnings, skips or tracked side effects. Full semantic scan: 48.065 s under the unchanged 120 s limit. Exact-scope author quality is valid; this is not merge authority.
Full CI / independent merge review	not queried / separate	No CI query or manual trigger. Required CI, independent review, exact-head readiness and maintainer merge authority remain separate.

The previous 3769a6051011f8abe8c3f42dbb2258333d3fcabf gate remains a historical failure: six direct and 13/14 selected checks passed, but semantic vocabulary scanning timed out at 120.713 seconds under its 120-second limit, and that exact-scope quality result is non-passing. The new head must earn its own result; no timeout, fixture or assertion was relaxed.

The paired RFC retains the rejected standalone late TS-sharing experiment: on the same 100-requirement fixture, projection parameters grew 102,680 → 344,010 bytes, while visible output stayed 102,189 bytes and all-details added a call. That relocation is excluded. Integrated C1 should absorb sharing at the existing source-owning composition boundary, then retire the compatibility algorithm with its callers.

The helper's current component samples do not establish a whole-Turn or long-horizon benefit. #7687 retains its own historical required-check failures and is not qualified by this combination. Original quota outcome/full-cost acceptance remains open.

No live installation, model/benchmark campaign, release, approval or self-merge is included. No authority-store behavior is changed and no PostgreSQL qualification is claimed. UI impact: none; the existing CLI/context-reader journey remains in place, with no viewport or interaction change. Every commit retains DCO sign-off; RFC language files change together.

[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)

Exact head: 45834a3ff77896720f2cfa8a8eb2ea881bc0f1c2；整项不可变 base：a36e01c32b0a290a96466c766629a0b01850f823。

动机 / Motivation

通过普通 quota 输出和恢复 reader 读取 Goal 完整要求的 Agent。

投影把完整验收 owner 移到 work context 后，已有别名可能仍指向被删除的位置；当前版本让本次实际迁移的引用跟随完整 owner，包括先访问别名、后处理 owner 的情况。

现有 CLI 与 TurnEnvelope 可以继续读取完整要求、来源、失败和修订；未知或未迁移的引用保持原有语义，调用者输入不被改写。

本 PR 不交付统一 controller 切换、live 安装、模型采用或完整成本净收益，也不增加状态、RPC、权限或配置。 C1 的生产者/consumer 切换、历史 reader 退役和 original-quota 成果/完整成本资格仍沿既有 #7430 保持开放。

改动思路 / Approach

这是决定完成后的引用展示修复；复用当前 compatibility projector，比把所有验收副本再传过一条 TS bridge 更合适。 当前 PR 修复现有 caller 的别名可解析性，并集成已单独评审的静态扫描 prerequisite；不把这两个有界结果宣称为 C1 完成。

不改会保留可复现的 dangling reference；只修 summary 的一个地址又会漏掉提前访问的 alias。这里在原 helper 中记录真正迁移的 owner，复用一个局部 copy-spine helper，然后第二遍修正已有引用。不解析任意输入链，也不按文本或相同 id 猜测两个来源相等；完整 JSON 的修订、类型、失败和 source 都参与比较。

原决定、准入、capture 和一次结算仍归各自 owner。Python 只在已决定的显示响应上改路径；现有 typed reader 继续消费同一份完整值。C1 后续应在 source-owning composition 吸收该规则并随 caller 退役兼容算法，不能把本次修复固化成第二套 authority。

具体改动 / Changes

完整九文件 +213/-38：三个生产模块 +73/-38、两个测试文件 +95、四份文档/按需指导 +45。共享 scanner prerequisite 的四个源码/测试/指导 blob 与已独立评审的 1dc6cb85d1203eeea1236d70c3127782fd375379 完全一致，且该提交是当前分支祖先；此次仍重新判断整个组合，而不是继承整项批准。

关键代码讲解

• _reference_inline_goal_acceptance 只为本次实际删除的 inline owner 建 relocation map；第二遍修正先前见过的 alias。嵌套 writable_parent 只复制被改的父链，原始输入不变，原 carrier 顺序和尺寸判断不变。
• compact_quota_should_run_cli_payload 保留原调用位置和已有 coarse reference_context 请求；没有新增 RPC。不同或不可用来源不会被强行归并，未知/未迁移引用保持既有 reader 语义。
• _field_mentions 复用完整 SourceFile 键下的已解析名称事实；不会用 substring 过滤代替 AST，也不再存一份 AST。语法错误、escaped string、normalized identifier 和 hash collision 反例保持。
• scan_python_production 在 qualified enum、builder 和 registered return 已解析后才跳过无关环境。没有字段文本不等于没有 producer，unknown 也不能被消掉。parse_python 的一行 ast.Module 注解纠正实际返回类型，不改 parser。

规范依据为 docs/reference/required-work-context.md，spec_revision：a36e01c32b0a290a96466c766629a0b01850f823。complete-required-context 与 same-capture-recovery 均 implemented：完整来源、修订、失败、原始要求和恢复仍可到达。并以同版本 semantic-vocabulary RFC 核对 I2、I7、I8、I11、I12、I13：固定 grammar、tracked-source 范围、角色、registered producer 和 unknown 边界保持。既有 controller/executor RFC 的 C1 仍 deferred 于 #7430，不把新加段落当作自己的 oracle。

对主干的风险 / Risks

未发现当前阻塞性问题。 同一组冻结反例通过真实公开 compactor：不可变 base 的三个旧 alias 地址断言失败，当前 head 三个通过；完整 owner、input immutability 和签名同时检查。另有 34 组同输入 default/all-detail、不同类型/修订/source/failure、unavailable/unknown 和 source-prefix 0/4/80 对照，完整结果及跨语言调用参数逐字相等，无归一化，原 default 一次、all-detail 零次的投影调用保持。最初使用各自即时 fixture clocks 的比较失败只暴露 observed_at/served_at 不同；该运行按“非同输入”保留，后续回放的是完整冻结输入，未抹掉时间字段造绿。

当前完整相关源码模块 497 passed。普通 wheel 的 1,898 个现有产品文件与源码逐项相等，三个变更模块哈希一致，实际载入的产品模块位于隔离安装环境；installed registered-return probe 保留无显式 field 的生产证据。该 wheel 上 28 项投影/恢复用例通过，另两项真实 File/SQLite CLI 用例在修正测试 staging 后通过。fixture 的源码相对定位仅指向实际 wheel；未替换产品实现、mock backend 成功结果或修改断言。最初缺 examples 及 source-only 路径的两次失败保留。另一次包来源预读正确挡住 ambient package，随后显式隔离 PYTHONPATH；没有把其他安装的行为算给本 wheel。

首次 wheel 构建也正确拒绝陈旧 frontend assets；按正常文档重建后完成打包，没有绕过来源检查。全部安装只在 disposable 环境，未升级 live Host。

标准 premerge 6 direct +17 selected 全过，零 warning、skip、timeout 或 manual hold；完整 native semantic check 45.588 秒，原限额 120 秒。普通 import 的两模块 mypy、fresh-cache 配置的17文件 mypy、Ruff、完整 docs governance 和双语 RFC 检查通过。最初猜错测试文件名的 exit4/no-collection 保留，纠正后的完整选择才计入497。作者旧3769的120.713秒失败及 #7687 原失败没有改写成通过；当前结果既不解释旧 timeout，也不替别的 head 合格。未查询、轮询或等待 CI。

语义与 CI 对齐

复用既有引用、parser 和 production 词汇，没有新增共享 Enum、actor 或权限状态。共享 scanner 的原 1,769-source 完整 API/corpus、反例和 profile 证据保留实际旧版本，通过四 blob 相等和 caller/grammar 检查复用；本轮同时验证当前完整 native tree。59 项动态 unknown 仍是 unknown，4096 个 cache entry 也不是 byte bound。原资料中的 component CPU 降低不等于 whole-command、整轮或模型净收益。

成对 RFC 保留了“后移共享反而增加 projection 参数”的原作者受控负结果，并明确那不是完整流量、延迟或科学结果；该实验数值未在本轮重新创造或升级为独立净收益证明。当前自身对照只证明没有新增 bridge 调用/参数变化。局部 helper 收敛已应用；更大的 source-sharing 迁移留给既有 C1 owner，未新建框架或跟踪任务。

我的整体评价 / Verdict

APPROVE 当前完整组合，delivery 为 justified_increment。long_horizon 为 preserved：原要求、来源、后续恢复和执行身份保持；user_experience 为 improved：已有命令能读回完整验收，不增加配置、重新填报或确认回合。三项故障恢复与其余完整对照共同支撑这一判断，不用通过数量、缓存字段或字节数代替收益。

视觉审阅不适用：变化位于 quota CLI/context JSON 投影和静态分析器，没有网页、App 视口或交互布局改动。 已安装 Host/App/Lark 的自然采用、完整 C1 切换、模型遵从与长期净收益未验证。

本结论没有继承父 PR 的效益资格；#7699 共享提交先后关系和任何后续 base/head 变化仍须合并前重新核对。当前 GitHub 同账号限制采用 COMMENTED 回退，代码批准不等于正式 GitHub APPROVED 或合并授权，也不授权安装、dismiss 或实验。

English verdict: APPROVE — 45834a3. Three immutable-base alias regressions pass through the current public projector; 34 exact-input observations and bridge parameters match without normalization. Independently verified 497 source tests, installed 28+2 projection/native cases with source provenance, and 6 direct/17 selected checks. C1 cutover, live/model adoption and whole-Turn net benefit remain separate; no merge authority.