[loopx-agent]

Long multiline Tasks can reach the quota CLI twice: the complete selected-Todo read and agent_lane_next_action.title. The existing title deduplicator compares the whitespace-normalized, bounded next-action text against the complete title, so it misses this case.

Reuse the complete, current, TypeScript-verified selected-Todo source already in this response. Replace only an exact duplicate lane title with title_ref; retain all requirements, source metadata, control fields and original captures. Missing, incomplete, ambiguous, mismatched or different sources retain their title. Explicit --include-detail agent-todos retains the inline title.

This is ordinary CLI serialization, applying with differential delivery off and on. The native differential switch remains default-off, and this change does not enable tracing or native transport. The existing Python display adapter consumes the typed owner's verification; it adds no decision, persistence or permission owner, capability, dependency or switch.

长、多行 Todo 的短摘要经过折叠和截短，旧去重无法识别它与完整标题的重复。本次仅把当前包内已核验、精确相同的标题副本改为 title_ref，完整要求与控制信息保留。普通全量路径也受益；差分开关继续默认关闭。详见同包引用合同。

Additional applicability calibration at the same reviewed head: all 31 previously selected historical benchmark packets have identical compact output under the two renderers. Reusing two retained single-line case Task text/title/note records inside isolated Goal/runtime fixtures, all 16 exact-ID native cells (32 cold/warm wakes; File/SQLite × off/on × base/head) consume complete input, reconstruct warm deltas, and retain the original Turn/debit checks. Across all eight same-mode base/head pairs, packet bytes and native request counts are unchanged (22 → 22 off; 19 → 19 on). Both renderers also give identical complete projections for all 32 fresh captures; no new title_ref is emitted because the legacy single-line path already avoids that duplicate. Entire-request token proxies vary −0.70% to +0.29% with identical serialization; this variation is not attributed as a saving or regression. These are Task-shape transport checks, not reruns of the original cases or solver outcomes.

因此，下方 16%–20% 仅属于 240 行多行任务的压力口径；本次两份真实单行任务文本对照没有额外 packet 缩减。原多行结果保留，但不能据此外推当前 benchmark 的收益。没有新增模型账单、solver 效果或安装采用结论。

Matched native-worker replay against clean ea1421812, using File/SQLite, one synthetic 240-line Task, two same-session wakes, the same CLI prefix, and unchanged 3 KiB first / 4 KiB continuation pages:

Provider / mode	Complete request-input proxy change	Native requests, two wakes
SQLite / off	−19.26%	30 → 26
File / off	−19.31%	30 → 26
SQLite / on	−16.03%	23 → 20
File / on	−17.26%	23 → 20

Cold packet size falls about 18.8% (57.0 → 46.3 KB). Counts use o200k_base over entire request JSON, including repeated history, tool definitions, instructions and every supplementary read. All eight revision/provider/mode cells consume complete bytes, preserve the original Turn and one inter-wake debit, and reconstruct warm deltas. Eight same-capture comparisons recover the entire baseline projection after expanding the exact reference.

This deterministic loopback qualifies transport, not model reasoning, billed tokens/cache pricing, solver score, installed Desktop adoption or IO/latency savings. Differential cold setup overhead remains; these are measured reductions against the same-mode baseline for this workload.

Validation on 22c3f1069:

• 109 focused contract / File / SQLite CLI cases passed.
• 162 context-consumer cases passed, including nine actual native loopback cases; 43 optional cases skipped.
• Configured strict mypy (17 source files), Ruff, TypeScript control-plane typecheck, public-boundary scan and bilingual diff check passed.
• Standard premerge: all six direct and 14 selected checks passed. Exact-scope CQR cqr_ad1eb99c20e356728786 is valid.

Known inherited failure: tests/control_plane/test_turn_completion_source.py has four callback-count IndexError failures on both clean base and head (24 pass / 4 fail each). Separate identical-command runs retain identical full failure signatures and fixture/environment digests. The failed short-Task captures have identical complete projections under both renderers; this new alias is inactive there. CQR records this author attribution as an unresolved warning. These tests remain failed: no full-suite or required remote-CI pass, independent review, merge readiness or installation is claimed.

The bounded companion refactor extends the existing exact-title helper and same-response reference pattern. Larger renderer migration remains separate from this serialization fix.

[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)

未发现阻塞问题，代码结论 APPROVE，绑定 22c3f1069c7f3081beeff479ff16404e0b972288。完整测试并非全绿：四个相邻完成回调测试在不可变原基线、当前目标分支和本 head 都失败，下面保留其独立归因；这不是远端 CI、合并就绪或合并授权。

动机

反复读取长、多行任务的 CLI Agent。

长任务原来同时出现在完整来源和下一动作标题中；现在只把已核验且精确相同的标题副本改为同包引用，完整任务仍可直接还原。

File 与 SQLite 的 16 个实际 CLI 基线/head 对照在精确展开引用后完整等价；长标题路径少传一份正文。

本 PR 不改变任务准入、验收、结算、来源权限或差分默认值，也不证明模型用量、费用或长期成果改善。

评审依据是改动前的 docs/reference/protocols/quota-cli-hot-path-compaction-v0.md，版本 ea1421812bab0c12da858a4216c6b9270ce16572。其 Ownership Boundary 要求先计算完整决定再投影视图，不能重算路由或丢弃行动权威；Qualification Contract 要求完整/紧凑还原、形状与边界的确定性验证，模型行为和真实成本另行资格化。这两个条件在当前范围有独立调用证据，不以作者的压缩率或测试数作为验收。

改动思路

同包精确引用复用现有 Python CLI 序列化层；完整来源核验继续由既有 TypeScript owner 负责，无新增决策或持久状态。

当前范围限于默认 CLI 包的一个重复标题及其显式明细回退；更大的渲染器迁移不属于此修复。

旧规则只比较 lane 的有界摘要与标题，遇到折叠换行或截断后的长任务时，完整标题仍重复传输。新分支先保留旧的短标题去重，再消费当前完整 work-context 的核验事实。只有唯一 selected-Todo 来源、正确任务身份、完整且已核验的正文，以及精确相同的 text/title 才能产生引用；不从前一捕获、相似文本或摘要猜测内容。

这次不适合再加缓存、接收确认或独立传输 owner。现有 TypeScript 已核对来源身份、完整正文摘要、状态和 claim；Python 只选取包内相等的 scalar。既有引用模式与显式明细已足以处理有用边界。用前后路径文字对照即可说明，未加入需要额外渲染验证的示意图。

具体改动

三条路径合计 +168/-10：cli_projection.py 的既有 helper 增加 +51/-10，测试增加99行，协议增加18行中英说明。_deduplicate_agent_lane_next_action 保留短文本旧行为，长标题则在引用确实比正文省字节时生成 title_ref。compact_quota_should_run_cli_payload 把同一完整包传给 helper，--include-detail agent-todos 仍直接保留标题。后续的现有压缩与 promotion 没有移走引用目标，原输入对象也没有改写。

独立验证包括：

• 对原 merge-base 与精确 head，从同一 canonical fixture 快照运行16个公开 CLI 场景，覆盖真实 File/SQLite、长正文、不同 title、短正文、显式明细、Markdown、完成任务、未 claim 任务，以及超过展示上限的12个其他 Agent 项。精确展开新引用后，完整输出一致；保留 provider revision，只归一化观察时钟、派生 freshness age 与随机 rollout event ID。
• 长标题、不同标题和超限无关项三类，在两后端都移除一份完整副本；原正文、控制字段、来源元数据、停止条件及权限没有缩短。旧版本的去重收益断言会失败，当前可从包内还原完整原文。未 claim 的候选没有获授权来源时保留旧路径，不把缺来源当成可去重证明。
• projection/settlement 的204个用例通过；另一组实际上下文 full/delta/recovery、Host 和捕获消费者86通过、30条件跳过。17种来源边界覆盖 missing proof、partial、failed、ambiguous、wrong identity、已有引用、done/archive、布尔身份及短文本等拒绝/保留路径。跳过项不冒充原生宿主或模型保留证明。
• 标准 premerge 的6项直接检查和14项选中检查通过、无 manual hold；配置内17源文件严格 mypy 与 TypeScript 控制面 typecheck 通过。未查询、轮询或等待远端 CI。

相邻 test_turn_completion_source.py 的同一命令在不可变 ea1421812bab0c12da858a4216c6b9270ce16572、当前目标 2b01f9450fc40472821c54e71c1da27b7cd5d91b 和当前 head 均为24通过、4失败。完整失败身份与 traceback 仅归一化临时目录和 monkeypatch 对象地址后相同：File/SQLite 两组回调列表索引在202/477行报 IndexError。这些短任务不进入新 alias，失败所依赖的来源/完成逻辑不在三文件改动里；独立长任务/完整包证据覆盖本次不变量。这四项仍是 failed，不宣称完整套件或主干所有检查已绿，也不要求本 PR 修无关的 fixture 回调契约。

对主干的风险

最强风险是错误引用掩盖了任务中的验收或停止条款，而 helper 单测仍通过。这里用实际公开 CLI 还原整个包来区分这个风险；仅“字段还在”或预算变小不足以证明语义等价。source 不完整或身份不符时保持标题；显式 detail 保持旧内联方式；原捕获不变，回退不产生新授权或 spend。

行为变化已经在协议和 PR 中明确：这是普通 CLI 序列化，同样作用于差分关闭与开启路径。alias 本身不是新的 opt-in 功能；现有上下文差分仍默认关闭，安装、命令存在、来源可读或一个 title_ref 都不能将其开启。现有实际消费者 full/delta/缺基线完整恢复覆盖开启侧，16组普通 CLI 对照覆盖关闭侧。本次验证没有调用付费模型，不把字节下降换算成缓存价格、tokens、IO/时延或长期研究收益。

标题原本就可能被旧 deduplicator 省略，现有 JSON/Markdown 消费者不依赖其必需性；现在增加可解析的同包来源。没有新增持久状态、协议版本分支、配置副本或权限 owner。较大渲染器迁移可以继续按原边界推进，不是这次小修复的前置条件。

我的整体评价

当前目标范围成立，goal_achieved / APPROVE。它消掉了真实可重现的重复传输，同时保留原任务、来源和行动/结算含义；没有把“更短”当作认知改善。repository experience 提醒检查实际用户结果，这里具体落实为完整 CLI 消费、来源反例、后续可读和零新增人工轮次，没有继承历史案例的批准结论。

同包精确引用复用现有 Python CLI 序列化层；完整来源核验继续由既有 TypeScript owner 负责，无新增决策或持久状态。 当前范围限于默认 CLI 包的一个重复标题及其显式明细回退；更大的渲染器迁移不属于此修复。

已检查 quota CLI、Turn/Host 上下文和 Markdown 调用；没有页面、控件或布局改动，视觉截图不适用。 没有新增视觉范围；没有重新认证安装态 App 或模型理解。

四个 baseline 等价的相邻失败仍保留为独立风险。真实模型理解、安装态采用和完整任务成本没有资格化；代码批准、CQR、GitHub 的旧 review/thread closeout、即时 merge-readiness 与维护者 merge 权限仍分别判断。

English verdict: APPROVE — 22c3f1069c7f3081beeff479ff16404e0b972288. Sixteen real File/SQLite public-CLI comparisons reconstruct the complete baseline packet from the exact same-response alias; focused/context coverage, native premerge, strict configured mypy and TypeScript typecheck pass. Four adjacent completion-callback failures reproduce with identical full diagnostics at immutable base, current target and head and remain failed. No model-cost, installed-adoption or merge authority is claimed.