[loopx-agent]

Repeated complete quota context can dominate solver input even after same-packet deduplication. This accepted design makes exact subtree reuse a shared quota/Turn projection concern, with automatic full recovery whenever the host cannot establish that the baseline remains in the active model context.

The English and Chinese RFCs specify:

• complete first delivery, fresh control facts every time, and exact reuse of unchanged requirement bodies;
• stable source identity and separation of changing authority metadata from unchanged text; explicit removals, atomic array replacement and complete-target reconstruction;
• host-owned continuity binding, including a final pre-dispatch check for compaction races, and baseline advancement only after intact model-input delivery;
• recovery from the same immutable capture without another admission, debit or human confirmation, while preserving existing source-freshness and permission refusals;
• the current native CLI limitation: session identity and capture-release ACKs do not establish active-context retention. Unsupported transports remain full; no incremental benefit is claimed for an all-full run.

Delivery order remains basic packet reductions → existing integration test → reviewed incremental implementation. #7165 supplies an independently reviewable basic reduction; this design does not hold that experiment waiting for a new protocol. The design review is accepted; incremental runtime implementation is still pending. Phase one must qualify both the actual benchmark solver and the productized Codex App heartbeat/continuation/recovery path using the same owner. Benchmark success alone cannot close phase one. An all-full App fallback remains safe but does not prove incremental productization.

Phase-one acceptance now explicitly includes real packaged/installed App readback, warm incremental delivery, first use, restart/resume, compaction/pruning, lost baseline and same-capture recovery. Missing host integration is an engineering gap, not a runtime owner question or additional admission gate. Design acceptance does not grant merge, install, release or new experiment authority.

Validation: the phase-one clarification at b2d0cba8fb65586a383bb5d8394d0032b986394b targets main cde8f0354ca1738f3b6ed8a175f01283cdb3e2db; paired RFC and public-boundary checks pass. Prior design validation at 8d6714e6e13d4d6401fa7b3270f06d3a012285f4 (base 035a077767d1e9fb9b321e35181ff9ffd81f00ef): bilingual staged and committed checks passed; whitespace, maintainability and public-boundary checks passed. Quick premerge retains the target-main semantic-vocabulary failure (goal_boundary: 32 Python modules, budget 30), reproduced on the unmodified base. No budget was relaxed; no runtime or live-model qualification is claimed.

中文：首包完整，温态发送变化子树；当前控制事实每轮完整给出。基线丢失、压缩或交付不完整时，从原 capture 自动恢复全量，不新增准入、扣额或人工问询。现有 CLI 的会话 ID/传输 ACK 不足以证明模型保留，需验证真正的模型输入组装边界。设计已 review 通过，一期须同时覆盖 benchmark solver 与产品化 Codex App，验收实际温态增量和压缩/恢复后的同 capture 全量回退；运行时实现仍待交付。

[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)

Exact head: b2d0cba8fb65586a383bb5d8394d0032b986394b；accepted baseline: 62851ab1543c140337d46d8cd45b4e92d9b4ac7a。

动机

反复读取普通 quota 上下文的 benchmark solver，以及产品化 Codex App heartbeat/续跑调用方与维护者。

此前即使同包内的重复已减少，后续回合仍可能重读未变的完整要求；直接省略正文又可能在压缩或重启后丢失约束。

本次明确完整首包、受连续性证据约束的子树复用和自动全量回退，并将两个真实调用方都纳入一期；当前实际输入、模型 token 和执行结果没有因此改变。

不实现增量 runtime，不修改当前默认、预算、冻结实验或授权，也不要求本人额外确认。 两个 host 的模型输入连续性接入、温态采用、实际打包/安装读回及完整成本收益仍待验证；裸 CLI 自动保持全量。

改动思路

已有 typed interaction/work-context owner 负责投影、精确比较与完整重建；host 负责模型输入顺序与连续性观察；原执行 owner 保留权限与效果。磁盘文件、session ID、旧 hash 或 stdout 前的 release ACK 都不能证明活跃上下文仍保留基线。

本 PR 交付完整双语设计与反例矩阵；基础包缩减先进入既有集成测试，增量一期须同时验证 benchmark solver 和产品化 App，各自包含真实 host adapter 与完整回退。 benchmark 可以先落地，App 未资格时一期继续开放。缺少 host 接入点时自动全量，记录工程缺口；不能增加 owner 问询、第二次准入或扣额。基本缩减仍先进入既有实验，没有被新协议阻塞。这里评审的是设计内容；其“已接受”措辞不构成合并、运行、安装或发布授权。

具体改动

spec_ref: docs/architecture/rfcs/agent-loop-effect-interpreter-v0.md；spec_revision: 62851ab1543c140337d46d8cd45b4e92d9b4ac7a。以修改前的三个收敛验收流与原四项下一切片为依据，未用新增文案自证。

关键内容讲解

• three-existing-streams：implemented，原三条验收流完整保留；设计不结算 runtime 或收益。
• ordinary-real-caller：implemented，优先实际 solver 的普通 quota 输入，并明确产品化 App 的独立一期验收。管理包、CLI fixture 或另一个 host 的证据均不能代替它。
• complete-current-requirements：implemented，正文、晚到停止条款与来源完整保留；稳定 source key 分离变化控制元数据，数组整段替换、删除显式表达，重建须等于完整目标。未知或混合来源保持全量。
• fresh-authority-and-recovery：implemented，每次完整给出身份、动作、准入、审批/停止、claim/lease、未完成读取和结算/scheduler。guard 后若压缩，在同一顺序边界内、模型提交前再次核对并换为同 capture 全量；只在精确展示完整进入活跃输入后才推进基线。来源/权限变化仍按原规则拒绝或 fresh guard。
• complete-cost-and-acceptance：implemented，首次及补读模型输入、上下文/cache、完整成本、后续动作和验收结果一起资格。App 还需实际打包/安装读回、温态增量、首次/恢复/压缩/丢基线及可见失败恢复；冻结 campaign 不改。

中英文完整 diff 已逐条核对，未发现这些条款的语义分歧。实际 quota emitter、capture handoff和原生 session 解析仍只提供前置释放/身份事实；文中明确要求尚未资格的模型输入组装 adapter，没有把此 seam 当作已交付。

对主干的风险

最强反例是 guard 后压缩，而旧 epoch 被误当作仍保留。设计要求提交前同边界重检、自动同 capture 全量恢复；这只恢复展示，不能重新授予过时执行权。丢失/截断/乱序 delta 不推进基线，delta 更贵时也发送全量。全 full 的 pilot 是正确性回退，不能证明增量节省。

当前标准 premerge 的4项直接检查及11项范围检查全部通过，含双语、完整语义扫描、8项风险 smoke 和公开边界；零 manual holds。旧 8d6714e 的32/30失败及当时基线归因保留在历史，不能套成新 head 的失败或称其由本 PR 修复。未查询 CI，未运行模型、连续性原型、安装或成本实验。

我的整体评价

APPROVE 这个双语设计增量，没有发现阻断设计内容的问题。原长程验收与完整要求保留，设计体验改进限于自动恢复和明确两个真实调用方的完成条件。采用随包经验中核对首个有用结果与实际采用的建议；没有要求设计 PR 先实现全部 runtime，也没有把经验投递、文档或绿色检查当作收益证明。

检查两个既有 RFC 的内部段落；未改变渲染组件、页面布局或首屏。 视觉未验证范围：none within this documentation-only presentation boundary。两个实际 host adapter、App 打包/安装采用和完整成本收益仍开放，批准不提供额外执行或合并权限。

English verdict: APPROVE — b2d0cba8fb65586a383bb5d8394d0032b986394b. The bilingual design preserves the three original acceptance streams, complete requirements, fresh authority and automatic same-capture recovery. Phase one explicitly requires both real benchmark and productized App consumers; actual adapters, installed adoption and benefit remain unqualified. All4 direct and11 selected premerge checks pass at this head. No runtime, token or merge authority is inferred.