[loopx-agent]

Quota/heartbeat, governed Turn and benchmark workers currently expose overlapping work contracts and lifecycle concepts. This proposal converges them on one complete contract and shared governed lifecycle, then introduces an optional controller/executor session split with a natural-language executor interface.

Proposed behavior

• One emitted schema and semantic instance across product paths, with shared full/delta/detail/recovery and original settlement identity. Retire active quota-versus-envelope format choice after consumer qualification; retain recorded versions for historical replay.
• The controller is the only model-facing Core client. Executor tasks/results carry natural-language work and complete sources; host adapters enforce role/session/attempt isolation. Existing typed owners retain admission, validation and settlement authority.
• Qualify unified single-session execution before the optional two-session profile. App heartbeat, managed execution, external workers and benchmark consumers reuse owners but each proves its actual host boundaries.
• C1–C9 identify concrete replacements, dependencies, acceptance and rollback. The source audit names both production composition callsites and downstream final-input readers. Independent obligation checks span composition, reconstruction and final input; shared captures alone do not prove unification.
• Recommend evolving the existing envelope version while retaining complete decision root paths, with actual host-operation binding. Explicitly migrate context-surface versions and reuse paths; do not inherit old baselines by Turn identity. Exact wire naming and complete mode/field mapping remain open before cutover.
• Keep freshness, relevance and settlement identity separate. A rejected stale review requires current facts and reassessment before an original-identity retry; another actor's changed acceptance is not automatically irrelevant.
• Synchronize the paired controller, effect, session, manager and roadmap RFCs, indexes and related protocol/context/heartbeat/benchmark guides. Repair the existing LayerFS pair's mirror declaration. Brief/result sizing is language-neutral exploratory guidance, not a runtime output cap.

Validation and limits

• Exact documentation head: 97d83bca9898671f2d624fcb513cf5d45f8b1ad5. Full docs-governance, staged/committed bilingual checks and diff hygiene pass. Changed files were scanned for private paths/state and internal links with no findings. Target is main; branch incorporates 99921829f2c5, with latest inspected target e20e4b684d04e4faed8ca3437880c9bbd70070a0.
• A synthetic layout probe used real isolated File/SQLite quota captures and the existing TypeScript codec. Preserved root paths enabled warm delta; nesting all facts while retaining the old surface caused delta_not_smaller full fallback. Both layouts reconstructed exactly, missing-baseline recovery preserved the target, epoch changes required full, original captures were unchanged and no spend occurred. Retention was synthetic. This rejects unmapped relocation under the existing codec, not nesting in general; no model or installed-host benefit is claimed.
• Earlier deterministic source/projection probes found distinct complete sources with equal compact action signatures. This rules out promoting the bounded projection unchanged; it does not establish a current authority-loss bug.
• The earlier risk-based premerge at 7e55f3618 passed 4 direct and 19 selected checks, with manual_review_required overall because the full PR includes benchmark-sensitive documentation. That run predates the final paired design clarifications and is not an exact-current-head merge gate. New-head CI, exact-head review and merge readiness remain separate.
• Related runtime proposals fix(turn): preserve work context and gate unsupported consumer versions #7420 and fix(replan): return current facts for same-obligation recovery #7426 supply consumer-preservation/version checks and stale-review recovery. Their evidence does not qualify a unified writer or close C1. Shared production composition, explicit writer/consumer cutover, historical-reader retirement, differential-context and all-host acceptance remain required.
• The paired C1 checkpoint now records unfinished runtime candidate WIP: compose one complete Turn contract before capture #7430: complete producer/final-input transport, original-Turn inline successor settlement, root-action semantic refusal and real selected-work inputs. Its final focused suite is 72 passed; six staged-reader/guidance failures also reproduce on the preceding candidate and remain open. Unknown optional facts and historical routing are retained. These checks do not qualify all consumers, native continuity, installed App adoption or model benefit.
• Documentation/design only: no default promotion, installation, model experiment, benchmark job or merge is performed by this PR.

中文：本提案统一契约、上下文交付与受治理生命周期，再引入可选 controller/executor 分离。新增布局建议保留完整事实根路径，随 wire 迁移显式升级 context surface；隔离 File/SQLite 探针验证了沿用旧 surface 却搬迁根路径会失去差分收益。该证据只支持布局决策，不代表模型收益、统一 writer 或 C1 完成。双语和 docs-governance 检查通过；实际入口替换与各宿主资格仍待交付，合并由维护者另行决定。

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; reasoning_effort=max; declaration_source=runtime_reported.

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

动机

Problem and delivery boundary. Maintainers connecting heartbeat, managed Turn and benchmark workers need one migration plan that preserves the work and its authority across those entrypoints. A shared capture can still feed different execution formats and final prompts, so an implementer can mistake local transport parity for complete convergence; this proposal identifies the actual callers, final consumers, replacement order and recovery boundaries. The verified result is a coherent bilingual design and C1–C9 cutover plan, with current behavior and unimplemented adoption distinguished. This PR implements no unified wire format, dual-session execution, default promotion, installed adoption or measured model benefit. Shared composition, versioned context cutover, role isolation and real App/benchmark adoption remain future implementation and qualification work.

这份提案解决的是迁移方向不够具体的问题。共享捕获文件，并不意味着各宿主最终收到同一份完整契约；本次把真实调用链、替换顺序和验收责任写清楚，避免将局部传输成功当作整体交付。

改动思路

Ownership and alternatives. Retain the domain authorities and converge composition before capture/encoding. An optional controller is the sole model-facing Core client; the executor receives ordinary work and authorized materials, under host-enforced isolation. This role is distinct from the deterministic Turn Loop Controller. One logical Agent still has at most one business executor. A pointer-only patch would retain contradictory permanent-format direction; implementing a new serializer in this documentation PR would add an unqualified third path.

The following source-derived table describes the proposal at 70d7e190012fdc4d635ab8b114ddd2ff5d2d990e; it is explanatory, not runtime evidence. Mermaid rendering was not inspected, so a plain table is used.

Concern	Current boundary	Proposed qualified boundary
Producer	Quota emission and direct Turn build separately	One complete composition owner, both real entrypoints
Consumer	Ordinary/staged native and structured prompts differ	Complete reconstruction and source-clause checks at actual final input
Recovery	Original capture and receipt retain their owners	Same-capture full fallback; fresh authority and original settlement remain distinct
Session roles	Existing single native binding	Explicit role/generation/ordered queue/isolation before optional split

先统一单会话的真实输入，再验收角色分工；两者不形成新的控制面。缺失基线、来源丢失、迟到结果和控制器退出都有明确的恢复或停止责任，不能靠 prompt 中删去工具名来证明隔离。

具体改动

关键内容讲解

• New RFC, sections3–5 and8–10: retains complete original requirements and source revisions, separates semantic/content/byte/settlement integrity, scopes session identity, preserves unknown fields and requires full recovery when reuse is unsupported. C1 names both producers and downstream final prompts; its first slice requires a complete real journey while wider host retirement remains open.
• Effect-interpreter convergence and session modes: update future design direction while retaining current active readers, historical captures, host effects and one-active-executor semantics. Manager handoff reuses existing request/revision/attempt owners; it does not introduce a second message ledger or authority rank.
• Roadmap, status, benchmark and protocol references: consistently label this design-only, retain current wire contracts, distinguish model controller from pure disposition, and keep token ranges experimental. The paired LayerFS edit clarifies the existing semantic-mirror maintenance rule. All22 changed paths and both full new language versions were reviewed.

Independent pre-change specification. docs/architecture/rfcs/agent-loop-effect-interpreter-v0.md at 99921829f2c5fa2ad35a024c72abb88717003ef3 supplies the existing M7.4 acceptance, rather than using the new RFC as its own oracle:

Criterion	Disposition at this head
M7.4-source-obligations	implemented in the design: source/actor/reads/lease/replan/settlement/scheduler clauses retained and assigned concrete owners
M7.4-real-host-adoption	deferred: C1/C2/C3/C8/C9 require real caller, recovery and packaged-host qualification before retirement
M7.4-measured-default	deferred: C7/C9 require matched complete costs and independently checked outcomes before promotion

条款核对覆盖了中英文中的主体、顺序、权限、必读来源、接续和停止条件。上述 deferred 是本提案明确留下的实现验收，不是已通过的运行证据。

对主干的风险

Falsification and validation. The strongest risk is declaring two equal compact action projections to be a complete-source contract. An independent synthetic probe confirms that changed diagnostic tails and an unknown root field can keep the same projected action hash while retaining different complete-source hashes. The proposal explicitly recognizes this distinction and does not label it current authority loss. Source inspection also confirms that quota emission captures before display projection, direct Turn builds an envelope separately, and capture IO is already shared; therefore a new capture service is unnecessary.

The existing real CLI full/delta/same-capture recovery test passed all4 File/SQLite and caller-profile cases. These checks support the current codec boundary; the author's synthetic layout observation and these checks establish neither future final-input adoption nor model savings. All953 relative file targets in changed Markdown exist.

语义与 CI 对齐

This extends design vocabulary through the bilingual RFC; it adds no runtime enum, registered wire version or default. Local risk-based premerge passed4direct and19selected checks, including RFC pairing, semantic advisory and public-boundary checks. Its overall manual_review_required is solely the benchmark_sensitive maintainer hold. GitHub CI was not fetched, polled or used. This approval does not clear that separate merge hold.

主要风险留在未来迁移阶段：须证明最终输入保留来源义务，并通过真实隔离、恢复和完整成本对照。当前协议仍有效，历史回执不可改写；本次没有启用双会话或改变现有结算行为。

我的整体评价

APPROVE — justified design increment at 70d7e190012fdc4d635ab8b114ddd2ff5d2d990e. Long-horizon execution is preserved: original obligations, current authority, replay and complete-cost acceptance remain binding. User experience is preserved: no new setup or confirmation round, and future controls reuse the existing conversation/activity/stop journey. No rendered presentation changes: only Markdown design/protocol content and indexes were changed; no App, Lark or CLI implementation changed. Unverified scope: future role runtime, installed adoption, model compliance and measured benefit; none is claimed delivered here.

The volume is justified by the bilingual design and coordinated replacement of conflicting direction, with existing owners reused and no speculative production abstraction. Design clarity improves; runtime effectiveness or efficiency is still unmeasured. All existing PR discussion arrays were empty when read. Published as COMMENTED because the shared GitHub account authored the PR; this records the review conclusion, not formal GitHub self-approval or merge authorization.

我认可这份设计切片。实现、宿主采用和默认提升仍按其各自出口验收；不把文档通过、编解码无损或自动检查通过等同用户成果。

English verdict: APPROVE — 70d7e190012fdc4d635ab8b114ddd2ff5d2d990e; complete bilingual design and caller/authority/recovery boundaries reviewed, current codec checks passed, future implementation and benchmark-sensitive maintainer merge hold remain explicit.

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; reasoning_effort=max; declaration_source=runtime_reported.

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

动机

Problem and delivery boundary. Maintainers connecting heartbeat, managed Turn and benchmark workers need one migration plan that preserves the work and its authority across those entrypoints. A shared capture can still feed different execution formats and final prompts, so an implementer can mistake local transport parity for complete convergence; this proposal identifies the actual callers, final consumers, replacement order and recovery boundaries. The verified result is a coherent bilingual design and C1–C9 cutover plan, with current behavior and unimplemented adoption distinguished. This PR implements no unified wire format, dual-session execution, default promotion, installed adoption or measured model benefit. Shared composition, versioned context cutover, role isolation and real App/benchmark adoption remain future implementation and qualification work.

这份提案解决的是迁移方向不够具体的问题。共享捕获文件，并不意味着各宿主最终收到同一份完整契约；本次把真实调用链、替换顺序和验收责任写清楚，避免将局部传输成功当作整体交付。

改动思路

Ownership and alternatives. Retain the domain authorities and converge composition before capture/encoding. An optional controller is the sole model-facing Core client; the executor receives ordinary work and authorized materials, under host-enforced isolation. This role is distinct from the deterministic Turn Loop Controller. One logical Agent still has at most one business executor. A pointer-only patch would retain contradictory permanent-format direction; implementing a new serializer in this documentation PR would add an unqualified third path.

The following source-derived table describes the proposal at 7ff693f36820363a45cf6a812d200edbd09aa573; it is explanatory, not runtime evidence. Mermaid rendering was not inspected, so a plain table is used.

Concern	Current boundary	Proposed qualified boundary
Producer	Quota emission and direct Turn build separately	One complete composition owner, both real entrypoints
Consumer	Ordinary/staged native and structured prompts differ	Complete reconstruction and source-clause checks at actual final input
Recovery	Original capture and receipt retain their owners	Same-capture full fallback; fresh authority and original settlement remain distinct
Session roles	Existing single native binding	Explicit role/generation/ordered queue/isolation before optional split

先统一单会话的真实输入，再验收角色分工；两者不形成新的控制面。缺失基线、来源丢失、迟到结果和控制器退出都有明确的恢复或停止责任，不能靠 prompt 中删去工具名来证明隔离。

具体改动

关键内容讲解

• New RFC, sections3–5 and8–10: retains complete original requirements and source revisions, separates semantic/content/byte/settlement integrity, scopes session identity, preserves unknown fields and requires full recovery when reuse is unsupported. C1 names both producers and downstream final prompts; its first slice requires a complete real journey while wider host retirement remains open.
• Effect-interpreter convergence and session modes: update future design direction while retaining current active readers, historical captures, host effects and one-active-executor semantics. Manager handoff reuses existing request/revision/attempt owners; it does not introduce a second message ledger or authority rank.
• Roadmap, status, benchmark and protocol references: consistently label this design-only, retain current wire contracts, distinguish model controller from pure disposition, and keep token ranges experimental. The paired LayerFS edit clarifies the existing semantic-mirror maintenance rule. All22 changed paths and both full new language versions were reviewed.

Independent pre-change specification. docs/architecture/rfcs/agent-loop-effect-interpreter-v0.md at 99921829f2c5fa2ad35a024c72abb88717003ef3 supplies the existing M7.4 acceptance, rather than using the new RFC as its own oracle:

Criterion	Disposition at this head
M7.4-source-obligations	implemented in the design: source/actor/reads/lease/replan/settlement/scheduler clauses retained and assigned concrete owners
M7.4-real-host-adoption	deferred: C1/C2/C3/C8/C9 require real caller, recovery and packaged-host qualification before retirement
M7.4-measured-default	deferred: C7/C9 require matched complete costs and independently checked outcomes before promotion

条款核对覆盖了中英文中的主体、顺序、权限、必读来源、接续和停止条件。上述 deferred 是本提案明确留下的实现验收，不是已通过的运行证据。

对主干的风险

Current-head change and reference verification. Since the prior reviewed 70d7e190012fdc4d635ab8b114ddd2ff5d2d990e, only the two new RFC language versions changed; all 20 other changed-file contents are byte-identical. Both complete incremental paragraphs were reviewed. They point to unfinished Draft #7430, inspected at ef68665e77df25ff6ef3f343b19f4d354bab0dd2. Its source composes before quota and direct Turn capture and sends the complete contract to Codex stdin. The updated source routes capability reentry through the canonical CLI channel, preserves original hash-bound review/writeback readers, and derives the one native-launch binding from recorded settlement identity. I independently ran 11 complete-contract tests, including fresh-obligation and prior-Todo replan reader/tamper cases, plus 8 current source-quota/health refusal cases and 2 actual bash/zsh bootstrap cases: all pass, using isolated File/SQLite fixtures and synthetic executables. These are source-path checks. Cache-suite consistency, staged parity across all consumers, structured/external migration and packaged host continuity remain unqualified; the author's earlier broader failures remain historical evidence. No full review, ready-state change or approval of that Draft is implied. The bilingual checkpoint accurately retains these limits; C1 and model/cost benefit remain unqualified.

Falsification and validation. The strongest risk is declaring two equal compact action projections to be a complete-source contract. An independent synthetic probe confirms that changed diagnostic tails and an unknown root field can keep the same projected action hash while retaining different complete-source hashes. The proposal explicitly recognizes this distinction and does not label it current authority loss. Source inspection also confirms that quota emission captures before display projection, direct Turn builds an envelope separately, and capture IO is already shared; therefore a new capture service is unnecessary.

Prior exact-source evidence is reused explicitly: the4real CLI full/delta/same-capture recovery cases and the action/source-hash probe passed at 70d7e190012fdc4d635ab8b114ddd2ff5d2d990e. Source, test and dependency paths are unchanged by the two-file documentation increment; those checks were not rerun or attributed to the referenced Draft. These checks support the current codec boundary; the author's synthetic layout observation and these checks establish neither future final-input adoption nor model savings. All953 relative file targets in changed Markdown exist.

语义与 CI 对齐

This extends design vocabulary through the bilingual RFC; it adds no runtime enum, registered wire version or default. At 393c7b66273718070beb9cea6bb10c2eb9b00426, local risk-based premerge passed 4 direct and 19 selected automatic checks; its overall manual_review_required was solely the benchmark_sensitive maintainer hold. At this exact head, the paired-RFC and diff-hygiene gates pass again; the fully reviewed two-file increment adds only the public Draft checkpoint above. The remaining premerge evidence is reused with exact unchanged source/test/dependency and 20-file content checks, rather than claimed as rerun. GitHub CI was not fetched, polled or used. This approval does not clear the separate maintainer merge hold.

主要风险留在未来迁移阶段：须证明最终输入保留来源义务，并通过真实隔离、恢复和完整成本对照。当前协议仍有效，历史回执不可改写；本次没有启用双会话或改变现有结算行为。

我的整体评价

APPROVE — justified design increment at 7ff693f36820363a45cf6a812d200edbd09aa573. Long-horizon execution is preserved: original obligations, current authority, replay and complete-cost acceptance remain binding. User experience is preserved: no new setup or confirmation round, and future controls reuse the existing conversation/activity/stop journey. No rendered presentation changes: only Markdown design/protocol content and indexes were changed; no App, Lark or CLI implementation changed. Unverified scope: future role runtime, installed adoption, model compliance and measured benefit; none is claimed delivered here.

The volume is justified by the bilingual design and coordinated replacement of conflicting direction, with existing owners reused and no speculative production abstraction. Design clarity improves; runtime effectiveness or efficiency is still unmeasured. The previous exact-head review was read back; no new human comment or inline thread was present. The current verdict is independently reassessed after the changed head. Published as COMMENTED because the shared GitHub account authored the PR; this records the review conclusion, not formal GitHub self-approval or merge authorization.

我认可这份设计切片。实现、宿主采用和默认提升仍按其各自出口验收；不把文档通过、编解码无损或自动检查通过等同用户成果。

English verdict: APPROVE — 7ff693f36820363a45cf6a812d200edbd09aa573; complete bilingual design and caller/authority/recovery boundaries reviewed, current codec checks passed, future implementation and benchmark-sensitive maintainer merge hold remain explicit.

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; reasoning_effort=max; declaration_source=runtime_reported.

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

动机

Problem and delivery boundary. Maintainers connecting heartbeat, managed Turn and benchmark workers need one migration plan that preserves the work and its authority across those entrypoints. A shared capture can still feed different execution formats and final prompts, so an implementer can mistake local transport parity for complete convergence; this proposal identifies the actual callers, final consumers, replacement order and recovery boundaries. The verified result is a coherent bilingual design and C1–C9 cutover plan, with current behavior and unimplemented adoption distinguished. This PR implements no unified wire format, dual-session execution, default promotion, installed adoption or measured model benefit. Completion and qualification of all consumer/version/context transitions, packaged host journeys and optional role isolation remain open.

这份提案解决的是迁移方向不够具体的问题。共享捕获文件，并不意味着各宿主最终收到同一份完整契约；本次把真实调用链、替换顺序和验收责任写清楚，避免将局部传输成功当作整体交付。

改动思路

Ownership and alternatives. Retain the domain authorities and converge composition before capture/encoding. An optional controller is the sole model-facing Core client; the executor receives ordinary work and authorized materials, under host-enforced isolation. This role is distinct from the deterministic Turn Loop Controller. One logical Agent still has at most one business executor. A pointer-only patch would retain contradictory permanent-format direction; implementing a new serializer in this documentation PR would add an unqualified third path.

The following source-derived table describes the proposal at 816e004a262f723b26a1df762ecb3727894bf0ae; it is explanatory, not runtime evidence. Mermaid rendering was not inspected, so a plain table is used.

Concern	Current boundary	Proposed qualified boundary
Producer	Quota emission and direct Turn build separately	One complete composition owner, both real entrypoints
Consumer	Ordinary/staged native and structured prompts differ	Complete reconstruction and source-clause checks at actual final input
Recovery	Original capture and receipt retain their owners	Same-capture full fallback; fresh authority and original settlement remain distinct
Session roles	Existing single native binding	Explicit role/generation/ordered queue/isolation before optional split

先统一单会话的真实输入，再验收角色分工；两者不形成新的控制面。缺失基线、来源丢失、迟到结果和控制器退出都有明确的恢复或停止责任，不能靠 prompt 中删去工具名来证明隔离。

具体改动

关键内容讲解

• New RFC, sections3–5 and8–10: retains complete original requirements and source revisions, separates semantic/content/byte/settlement integrity, scopes session identity, preserves unknown fields and requires full recovery when reuse is unsupported. C1 names both producers and downstream final prompts; its first slice requires a complete real journey while wider host retirement remains open.
• Effect-interpreter convergence and session modes: update future design direction while retaining current active readers, historical captures, host effects and one-active-executor semantics. Manager handoff reuses existing request/revision/attempt owners; it does not introduce a second message ledger or authority rank.
• Roadmap, status, benchmark and protocol references: consistently label this design-only, retain current wire contracts, distinguish model controller from pure disposition, and keep token ranges experimental. The paired LayerFS edit clarifies the existing semantic-mirror maintenance rule. All22 changed paths and both full new language versions were reviewed.

Independent pre-change specification. docs/architecture/rfcs/agent-loop-effect-interpreter-v0.md at 99921829f2c5fa2ad35a024c72abb88717003ef3 supplies the existing M7.4 acceptance, rather than using the new RFC as its own oracle:

Criterion	Disposition at this head
M7.4-source-obligations	implemented in the design: source/actor/reads/lease/replan/settlement/scheduler clauses retained and assigned concrete owners
M7.4-real-host-adoption	deferred: C1/C2/C3/C8/C9 require real caller, recovery and packaged-host qualification before retirement
M7.4-measured-default	deferred: C7/C9 require matched complete costs and independently checked outcomes before promotion

条款核对覆盖了中英文中的主体、顺序、权限、必读来源、接续和停止条件。上述 deferred 是本提案明确留下的实现验收，不是已通过的运行证据。

对主干的风险

Current-head change and reference verification. Since the prior reviewed 7ff693f36820363a45cf6a812d200edbd09aa573, only the two new RFC language versions changed; all 20 other changed-file contents are byte-identical. Every intervening paired paragraph was read. The referenced unfinished Draft #7430 was inspected at 66878872bbb6967740be59928f215057360af479. Its shared complete-input helper/structured renderer, Markdown source preservation, modern inner-version/binding validation and pre-capture successor-binding owner were traced separately. Complete source retains inactive facts without adding recall instructions or reflection result fields. A supported Todo-bound DSH route gets the original capture; an obligation-only DSH route still rejects before dispatch.

The earlier complete-contract/host-memory run at 07ca38f0351d887c31f9524c0be0d034b7e1bb68 was 125 passed. Four representative synthetic historical authority mappings produce byte-identical full renderer prompts against the prior candidate, with no normalization; renderer source remains byte-identical at the newer reference. That comparison proves formatting parity, not request admission, activation or model behavior.

At the newer 66878872bbb6967740be59928f215057360af479, independent complete-contract/staged-reader validation produced 50 passed, 6 failed. The new File/SQLite App-heartbeat and generic-CLI inline-successor cases pass: create from the captured contract, reject premature spend, settle the original identity once and leave the successor open without another guard/reader. The checkpoint preflight cases preserve complete input/full recovery across the File/SQLite, replan and context-delivery matrix; unsupported modern inner versions/bindings reject even with matching outer hashes. These use real isolated authorities and synthetic executables/runners; no benchmark job or model runs. Six legacy staged-suite cases still fail on compact Explore-template absence, compact settlement-plan absence and the old envelope writeback alias. All six exact cases/error categories also fail on the prior candidate head; they remain failed migration coverage, not reclassified passes. The author's 70-case external-worker result remains reported evidence; I did not independently rerun that selection. Its stated scope is scheduling/recovery, while model delivery through an opaque shell wake remains explicitly unqualified.

The bilingual checkpoint accurately separates source transport, inline lifecycle, activation and model/installed adoption. Remaining reader/cache/staged/all-consumer migrations, version admission and packaged continuity remain open; operation-SDK runtime checks remain untested. Prior broader failures remain historical evidence. No Draft review, ready-state change, C1 qualification or model/cost benefit is implied.

Falsification and validation. The strongest risk is declaring two equal compact action projections to be a complete-source contract. An independent synthetic probe confirms that changed diagnostic tails and an unknown root field can keep the same projected action hash while retaining different complete-source hashes. The proposal explicitly recognizes this distinction and does not label it current authority loss. Source inspection also confirms that quota emission captures before display projection, direct Turn builds an envelope separately, and capture IO is already shared; therefore a new capture service is unnecessary.

Prior exact-source evidence is reused explicitly: the4real CLI full/delta/same-capture recovery cases and the action/source-hash probe passed at 70d7e190012fdc4d635ab8b114ddd2ff5d2d990e. Source, test and dependency paths are unchanged by the two-file documentation increment; those checks were not rerun or attributed to the referenced Draft. These checks support the current codec boundary; the author's synthetic layout observation and these checks establish neither future final-input adoption nor model savings. All953 relative file targets in changed Markdown exist.

语义与 CI 对齐

This extends design vocabulary through the bilingual RFC; it adds no runtime enum, registered wire version or default. At 393c7b66273718070beb9cea6bb10c2eb9b00426, local risk-based premerge passed 4 direct and 19 selected automatic checks; its overall manual_review_required was solely the benchmark_sensitive maintainer hold. At this exact head, the paired-RFC and diff-hygiene gates pass again; the fully reviewed two-file increment adds only the disclosed structured-host/activation checkpoint above. The remaining premerge evidence is reused with exact unchanged source/test/dependency and 20-file content checks, rather than claimed as rerun. GitHub CI was not fetched, polled or used. This approval does not clear the separate maintainer merge hold.

主要风险留在未来迁移阶段：须证明最终输入保留来源义务，并通过真实隔离、恢复和完整成本对照。当前协议仍有效，历史回执不可改写；本次没有启用双会话或改变现有结算行为。

我的整体评价

APPROVE — justified design increment at 816e004a262f723b26a1df762ecb3727894bf0ae. Long-horizon execution is preserved: original obligations, current authority, replay and complete-cost acceptance remain binding. User experience is preserved: no new setup or confirmation round, and future controls reuse the existing conversation/activity/stop journey. No rendered presentation changes: only Markdown design/protocol content and indexes were changed; no App, Lark or CLI implementation changed. Unverified scope: future role runtime, installed adoption, model compliance and measured benefit; none is claimed delivered here.

The volume is justified by the bilingual design and coordinated replacement of conflicting direction, with existing owners reused and no speculative production abstraction. Design clarity improves; runtime effectiveness or efficiency is still unmeasured. The previous exact-head review was read back; no new human comment or inline thread was present. The current verdict is independently reassessed after the changed head. Published as COMMENTED because the shared GitHub account authored the PR; this records the review conclusion, not formal GitHub self-approval or merge authorization.

我认可这份设计切片。实现、宿主采用和默认提升仍按其各自出口验收；不把文档通过、编解码无损或自动检查通过等同用户成果。

English verdict: APPROVE — 816e004a262f723b26a1df762ecb3727894bf0ae; complete bilingual design and caller/authority/recovery boundaries reviewed, current codec checks passed, future implementation and benchmark-sensitive maintainer merge hold remain explicit.