[loopx-agent]

Quota context sharing currently owns exact Task-title and Explore-guidance rules in Python while Turn/replan readers relocate the same context in TypeScript. A same-response Explore reference can therefore outlive the path it points to when a reader moves the context.

This PR routes the existing quota callers through one coarse call to the existing TypeScript owner and restores referenced guidance before Turn/replan relocation. It removes the duplicate Python algorithms while retaining current quota/Turn schemas and entrypoints. It is an independently usable prerequisite for #7430; the complete-contract producer/consumer cutover remains there. The paired controller RFC now distinguishes an acceptance target from a mandatory single PR.

Behavior and boundaries

• Exact Task-title sharing retains source identity, full requirements, Unicode length semantics and the local display-only prefix omission. Unverified or different sources remain inline.
• Explore guidance shares exact inline durable-writeback text within the same response. Short text remains inline when a reference costs more; separate guidance owners can each be shared. Readers restore full instructions before relocation. Dangling, chained, conflicting or mistyped references reject explicitly.
• Both sharing operations use one transport call and preserve independent detail selectors; explicit full detail or absent context sources requires no sharing call.
• Original source objects, settlement commands/identity and negative evidence remain intact. No new schema, host entrypoint, packet-format toggle, model job or authority-store behavior is introduced.
• Rollback is local: revert this PR to restore the prior projection/read behavior. C1 remains open for the actual cutover, remaining consumer retirement and original-quota outcome/full-cost qualification. The remaining bridge adds projection cost relative to the original Python implementation; no end-to-end benefit is claimed.

Validation

• 138 passed on the updated implementation: quota projection, shared context references, real staged replan CLI and Turn envelope suites. Existing staged CLI coverage exercises isolated File and SQLite flows, original capture/readback and settlement identity.
• TypeScript Turn/context suites: 70 passed; control-plane typecheck passed.
• Ruff, diff check, docs-governance smoke, bilingual RFC staged/exact-head checks and project-scoped loopx check passed; public-boundary scan clean for all seven changed files. Diff-scoped semantic advisory found no supported new closed-vocabulary carriers; this is not proof of semantic parity.
• Initial matrix: 67 passed / 2 failed because an existing attachment fixture omitted the now-required finding status. The fixture now supplies tentative, allowing its intended missing-settlement-identity refusal to be tested; product validation was not weakened. An initial global-registry check reported an unrelated unmigrated Goal; the explicit project-scoped check passed.

Cost evidence and limits

A controlled warm projection experiment rotates three arms over the same synthetic 26.9/32.7/86.7 KB inputs (60 samples each): original base Python, the prior two-call shape, and the new one-call shape. The two-call arm uses the prior Python caller with its internal operation names mapped to the same current TypeScript owner; this isolates call count and is not a full-revision comparison. All outputs are exactly equal and inputs remain unchanged.

With the existing CLI request scope already resolved, median projection times are:

Input	Original Python	Two calls, same owner	One call
26.9 KB	0.062 ms	2.277 ms	1.196 ms
32.7 KB	0.071 ms	2.651 ms	1.441 ms
86.7 KB	0.156 ms	3.469 ms	1.879 ms

The intervention reduces this added projection cost by about 46–47%, but remains slower than the original Python projection. Direct calls outside a request scope also repeat runtime source fingerprinting and have materially higher, variable costs. An earlier post-change run overlapped tests and is excluded from the controlled comparison; profiler sampling cannot attribute child-process CPU. These are synthetic component timings, not default-provider CLI latency, model token/quality results, or complete-cost qualification. Raw packets/profiles stay private. No performance budget or acceptance criterion is relaxed.

No packaged/installed adoption, model-effectiveness comparison or complete-cost measurement is claimed. This changes existing internal composition and recovery, not a new frontend journey. CI and independent review remain pending.

[loopx-agent]

Reviewer: model_agent; runtime_reported; gpt-6.1-sol; OpenAI; reasoning_effort=max

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

动机 / Motivation

使用普通 quota CLI、捕获后的重规划入口及 TurnEnvelope 的 Agent，会反复读取同一任务正文和 Explore 结果指导。同一份完整说明在任务源、操作摘要和结算附件中重复出现时，原 Python 入口自行选择可引用的副本，TypeScript 读者另行搬运上下文；以后修改同一规则需要同时核对两处。本版让普通入口实际调用同一个 typed owner，并在读者搬运上下文前恢复完整指导。普通 CLI 已实际采用共享实现；同一冻结输入的十八份投影或公开 reader 输出与旧版完全相等，原任务正文、负证据和结算身份均保留。这是已被调用的内部前置切片；不宣称完成 C1 的统一 producer/consumer 切换，也不宣称减少模型 token、入口延迟或长期费用。C1 的完整 schema、各最终 Host 输入、writer 切换和旧 reader 退出仍由原 C1 工作承接；本 PR 不将这些未验项目标为完成。

这里的 owner 指负责解释规则的唯一实现，引用只是同一捕获内对原文的定位。实用增量在于以后修改同一引用规则时只维护一处，同时已有普通入口立即采用；完整源和原结算继续由既有 owner 负责。未发现 blocking finding，下面的批准只针对这个已采用切片。

改动思路 / Approach

同一引用规则归现有 TypeScript owner，Python 仅保留显示整理和一次传输；没有新增 authority、配置或持久状态。当前交付是普通 quota CLI 已采用的引用 owner 与完整读者恢复，C1 切换及实际模型收益保持在原后续验收。

普通 CLI 完成原决策后，compact_quota_should_run_cli_payload 调用一次既有 TypeScript 事务通道；Task 详情与 vision 详情分别控制两种引用，两者都显式展开或没有 source carrier 时跳过。原来只省略重复显示标签的 Python 逻辑仍留在显示层，不参与权威源匹配。

规则	旧路径	当前 head
Task 标题引用	Python 选择完整当前 Task 的相同正文	TS currentTaskTitleOwner/referenceCurrentTaskTitle；要求完整、唯一、已匹配且非 done/archive，按 Unicode code point 保持原阈值。
Explore 指导引用	Python 查找相同结算附件文本	TS referenceExploreGuidance；仅 exact-text 同 packet 的 inline durable-writeback Explore owner，短文或不同文不引用。
完整读者	TS 各自搬运上下文	resolveExploreGuidance 先恢复原文，再交原 replanWorkContext；坏引用明确拒绝。

这是源码导出的归属表，不当作执行证据。真实执行证据见后文。通用完整性约束归原 typed owner，Python 保留 IO；Explore optional attachment 仍是可选能力，既有结算和 required validation 不变。仓库 #5944 经验适用于核对“是否有实际调用者和有用结果”，不要求这个内部前置 PR 交付整个父路线图或凭回执宣称模型收益。

具体改动 / Changes

审查完整 head 56559929aad42dc86df8b0035092595a48ba0ac8，不可变 base 71057d44f0d50a3636fcd8796f7c98717ce41742，共七条路径、+339/-87：生产 +113/-83、测试 +182/-1、文档 +44/-3，无生成文件或新模块。删除 Python 两段匹配算法，在原 TS 文件加入共享匹配/恢复，并改接实际 caller；新增181行用例覆盖同源/不同源、短文、Unicode、多 owner、坏引用和详情分支。另一测试只补齐原有 status=tentative 必填夹具，使它真正到达原本要验证的缺 Todo 身份拒绝，未放松产品要求。

规范依据：完整阅读 不可变 controller/executor RFC、同版本 compaction 协议，并核对现有整体路线图 S2/S6/S7 的完整来源、单 owner 和成本边界。spec_revision=71057d44f0d50a3636fcd8796f7c98717ce41742。各 criterion_id 的本 head 处置：

criterion_id	处置
§5 one complete work contract	deferred：完整 producer/final Host/cutover 不在本 PR，由原 C1 承接，未宣称完成。
§5 typed owner and full requirements	implemented：现有 TS owner，普通 CLI 实际采用，完整 Task/证据保留。
§7 recovery and authority	implemented（本切片）：原 source、Task/Turn、可执行结算与 once-only 读回保持；无新执行权。
§8 full-first and differential context	implemented（引用边界）：精确同 capture、恢复前校验、源不变及负例成立；不代表整个差分系统新资格。
§9 matched outcomes and full cost	deferred：实际 Host/模型/完整费用和受控延迟未验，不把字节或测试数算收益。
§11 implementation and first cutover	implemented（阶段说明）：双语把可用前置与 first cutover 区分，首次切换仍需 producer+consumer 资格，不将 serializer/存活状态算完成。

关键代码讲解

1. cli_projection.py:770::_reference_context 只做 carrier/详情选择并调用原 transport，不复制“哪个 source 合法”的业务判断。适用普通投影新增一次粗粒度 bridge 调用，这有序列化成本，不能宣称提速。
2. turn_envelope.ts:275::currentTaskTitleOwner/referenceCurrentTaskTitle 绑定 selected Todo 与唯一完整 source；只给相同 text/title 建别名，保留历史 caller-owned reference。Array.from(JSON.stringify(title)).length 避免 emoji 被 UTF-16 双计，匹配旧 Python 字符阈值。
3. turn_envelope.ts:244::referenceExploreGuidance 比较的是完整相同字符串；不同 owner 文本、短指导、缺附件和非 Explore source 保留 inline。它不把“可发现”当启用、不推导用户意图，也不新建 state。
4. turn_envelope.ts:222::resolveExploreGuidance 检查结构化路径必须指向 inline durable-writeback Explore 附件，拒绝悬空/错 step/错 capability/chain/inline+ref。对象经过复制，删除 alias 不修改原输入；完整 reader 在搬运 source 前恢复原文，保持测量反馈、负证据和原结算。

对主干的风险 / Risks

没有当前 blocking finding。 最强风险是引用搬运后丢掉执行前义务，或为了减少重复文字而扩大权限/删除未知。本次逐条核对了原 Task 完整正文、先读 measurement_feedback、保持原 task lease 到 refresh、raw logs 留本地、capture optional、实际被评价 candidate/版本范围、负证据/unknown、原 validation→writeback→spend 顺序和原 Task/Turn 绑定；不是只比 reason、schema 或大小。它们在同源别名与恢复后保持原文。六种显式非法引用拒绝，短文/不同文/无 owner 保留 inline；done/archive/不完整/歧义 Task 不借摘要伪造完整源。

PASS：独立运行145项相关 Python、73项 TypeScript、control-plane typecheck、改动 Python Ruff；标准 premerge 6 direct + 17 selected 全过、零 manual hold，包含双语 RFC、semantic/maintainability/public-boundary。最终七路径 CQR cqr_a1f680f99f458f97b12c 为 PASS，物理规范回执在首次公开结论前完整读取。

真实路径比较：同一私有合成 harness 在固定旧 base 和当前 head 各运行10个 File/SQLite CLI 场景，覆盖 Explore off/on、显式 detail、捕获 replan 与 envelope，原 Task/claim/status、full source 和 settlement identity 保留，未新增 Spend。另把同十份冻结完整 capture 交给两版实际 CLI projection dispatcher 和每版8次公开 quota replan-view：共18份可比完整输出逐字段/字节相等，没有任何规范化。独立 native fixture 路径和时钟不同，完整 fingerprint 分别保留，不把它们冒作逐字相同。

回归敏感性：在真实 CLI projection seam 刻意删掉完整 Explore guidance/ref，两种真实后端均被原 full-graph oracle 抓住；正常 head 十例通过。另8个真实 CLI 反事实只改投影输入、保持 authority 决策不变：去重复 inline 副本而保留合法同源引用、加入80条无关 source、倒转其显示顺序、标明 carrier 不完整，均保留完整 Task/图中义务、负证据、原身份和零额外扣额。这些 seam 变异不冒充 storage 故障或真实模型执行。done/archive title 属显示资格，未改生命周期或授权 gate，已有相关测试保持 inline。

补充探针的早期失败已保留并归因：多行文本被既有 CLI 空白规范化、未先进入原重规划准入、误以为 replan packet 必有 selected-Todo source，以及把文件字节 SHA 当 typed canonical JSON hash；还有一次把只返回 review/writeback 数据的公开 reader 当作 work_context reader。更正夹具/读取方式后使用当前完整对照，未删除产品断言或把这些探针错误归为 PR 缺陷。没有查询或等待 GitHub CI。原 PG authority/schema/writer 未改，此处不新宣称 PG qualification；完整仓库、最大 packet/持续负载、正式安装 Host 与模型行为未新验。

语义与 CI 对齐

复用现有 *_ref 词汇与 TypeScript transaction owner；reference_context 是现有通道的本地 operation，不是新 actor/权限协议。regex 只校验结构化 JSON 路径，不按自然语言 substring 决定业务状态。Explore 关闭的两份完整冻结输出与旧版完全相等，整个 shared builder 的 schema/指导/原决策/身份都核对过，不能仅靠“没有 feature object”推断隔离。没有默认开启新能力或新增机器义务；optional 指导与 enforced bad-reference refusal 分开。RFC 阶段文字是明确提案修订，父级 C1 与 matched-cost 资格仍保留。

我的整体评价 / Verdict

这是 justified_increment，完整源及可恢复引用在现有入口实际采用，long_horizon 与 user_experience 均为 preserved；局部维护收益来自删除重复决策，不从测试数、字节、单次耗时推长期效果。同一引用规则归现有 TypeScript owner，Python 仅保留显示整理和一次传输；没有新增 authority、配置或持久状态。当前交付是普通 quota CLI 已采用的引用 owner 与完整读者恢复，C1 切换及实际模型收益保持在原后续验收。

代码量为 bounded 的原 owner 扩展，没有新模块、设置或用户操作轮次；保留真实 full capture/原 reader 的兼容义务，未来 producer cutover 继续由原 C1 验收。额外 bridge 序列化、最大输入和完整性能/费用是明确未验边界；PR 也未作提速或 token 节省承诺。visual_review=not_applicable：七路径没有 rendered UI，相关视觉未验范围为 none；安装和实际模型采用仍未资格。

English verdict: APPROVE at 5655992: the ordinary CLI adopts the existing typed reference owner, full instruction/identity semantics survive native and identical-input comparisons, and required local checks pass. This qualifies the bounded prerequisite only; full C1 cutover, installed Host/model adoption and performance/cost benefit remain unverified.