Repository navigation
Default to Complete Turn and retire legacy delivery selectors - #9176
loopx-agent wants to merge 5 commits into
Conversation
Signed-off-by: LoopX Agent <[email protected]>
Signed-off-by: LoopX Agent <[email protected]>
Signed-off-by: LoopX Agent <[email protected]>
Signed-off-by: LoopX Agent <[email protected]>
Signed-off-by: LoopX Agent <[email protected]>
loopx-agent
left a comment
There was a problem hiding this comment.
Reviewer: model_agent · gpt-6.2-sol · OpenAI (runtime_reported, xhigh)
Request changes conclusion (author-owned PR; GitHub blocks formal self-review)
[P1] 保留公开制品的 C1 采用门。 这个提交把 complete_turn 写入公开制品的当前支持策略,也删除了 validator 对原 §9 资格的约束说明。独立复跑已提交源码的真实 writer 探测,确认 v1 会通过新的 validate_default_contract,并与提交的 runtime archive 和描述符相符。合入后 main snapshot 工作流只需现有安装与归档检查便可发布该默认值。当前提案明确仍缺支持 Host 的实际采用、匹配的单会话成果和完整成本证据。请保留已有公开 quota_packet 资格和仍在使用的入口,将已验证的 C1 工程改动按可回退边界送审,达到原 §9 条件再切换公开默认和退役入口。 支持策略 现行发布规范
[P2] 给只支持旧 attestation 的 macOS preview 一条可读回的升级路径。 把 base/head 的实际 Rust qualify_candidate 和身份解析器放在同一测试程序中,用相同归档形状换入两种有效描述符。旧 updater 接受 quota_packet,对新 v1 preview 返回 update_platform_unavailable。新 updater 接受两种有效描述符,继续拒绝缺失与未知 attestations。PR 已标明旧版需要兼容路径。发放新 preview 前需给出并验证一个能从旧 preview 到达的已签名过渡版本或明确的安装路径,确认既有 App 与 runtime 可继续使用和回退。 更新校验
[P3] 统一双语发行说明里的默认值和结算适用范围。 首段、破坏性变更和分发说明仍把普通 JSON 写成 quota_packet,新版启用说明却写 v1。按发行示例直接运行,渐进捕获成功。其保存的普通 CLI 决策有 next_cli_actions,显式绑定 generic_cli 的同 Turn 才返回 settlement_plan。依据最终发布范围统一两种语言的描述,并分别列出普通 CLI 和有结算计划的 Host 调用。 发行说明
动机 / Motivation
普通绑定 CLI、heartbeat 与制品升级的调用方需要同一份当前工作契约,既有 Turn 需要保留原始读取与结算身份。
原版本的普通 JSON 保留已获资格的 quota_packet,显式或已绑定 Host 可使用完整 v1。候选统一了当前组合入口,同时退休两个实时格式开关,并把 v1 纳入公开制品支持范围。
隔离的新旧 CLI 在 File 和 SQLite 都保住旧捕获摘要、原 effect 和一次准入扣额。公开分发的支持范围先于 RFC §9 验收改变,旧 preview 的更新过渡也尚未交付。
工程读取与重试通过后,公开默认仍需原 §9 的匹配单会话成果、完整成本及支持环境采用证据。
正式安装 Host、原生 Windows、已签名 App 的实际替换与匹配模型试验未验证。测试使用隔离状态,未触发托管 CI。
改动思路 / Approach
保留现有发布资格,独立交付可以回退的 C1 生产接缝与历史恢复。现有 TypeScript composer 能承接完整契约,Python 继续拥有原生准入、捕获和展示。两个格式选项和活动 benchmark 调用可以在明确的支持边界收敛。生产声明与运行时探测一致证明制品身份。用同一源码同时改变探测值与接受值,仍需要独立的产品采用证据来判定它是否可以发布。
具体改动 / Changes
精确 head 8dbcbbf7a05f5cea8f9ac837958cd15420fa6a11,base 64353773dc8b1ca82ff531cc7653dec6c3d41c0b。5 笔签署提交,83 个文件,+518/-829。覆盖 CLI 与 heartbeat、捕获和 checkpoint、benchmark runner 接线、分发政策与 macOS、双语文档及回归验证。按基线 docs/architecture/rfcs/controller-executor-unified-contract-v0.md §9 核对。 RFC §9
- current-source-and-off。普通完整绑定 CLI 返回 v1。显示 Off 与 Delta Off 给内联 v1,heartbeat 生成的当前 guard 去掉两个旧选项。保留 scheduler-observation 单独读法。
- historical-recovery。对真 CLI Base → 归档 Head 的 File、SQLite 隔离夹具逐个核对,旧捕获字节和摘要不变,继续使用原 effect。首次结算与同指令重跑共计一次准入和一次扣额。普通命令提供原 next actions,显式 Host profile 提供 v2 结算计划。
- default-adoption-before-retirement。当前工作组合已切换,公开资格未满足。
- supported-update-path。新 updater 的双描述符兼容通过,上一代 updater 到 v1 preview 无兼容的原位路径证据。
关键代码讲解
compose_quota_work_contract删除未声明 JSON 的早退,改用已有 TypeScript 完整契约 owner。普通 Off 关闭交付选项后仍返回 v1。bind_capture_selection_transport从重新呈现的选取命令清除旧布尔选项。原始文件保有摘要与旧语义,_verify_interrupted_guard则按原生检查点记载核对历史命令。reject_retired_turn_display将历史显式 treatment 拒于新 benchmark 启动之前,历史结果应按原 product 和 runner 解释。validate_default_contract把新公开候选和历史兼容读取分开。当前支持值仍须由既有分发采用证据约束。qualify_candidate对 macOS preview 核对归档内的受支持描述符。安装端先运行旧二进制自己的校验,因此新二进制支持旧格式只证明逆向读取兼容。
对主干的风险 / Risks
原生 CLI 旧捕获和重试对照、237 项聚焦 Python、直接执行两版 Rust 校验、已提交 runtime writer 与归档摘要重验均已通过。benchmark runner 回归 249 项通过、413 项跳过。标准 premerge 的 6 项直接检查和 19 项选中检查零失败,保留 benchmark-sensitive 人工门。上游支持策略在任何新的匹配模型证据进入工作流之前就接受了新默认值。需要单独保留 Mac preview 过渡、旧运行时读取和回退的证据边界。
我的整体评价 / Verdict
REQUEST_CHANGES。 可以保留经验证的共享组合、明确 Off、原捕获恢复和新的 updater 读取兼容。把公开默认和旧入口退役留在已证明采用的范围内,补足从旧 preview 的可验证过渡和双语发行说明。此次改动的是 macOS 下载后校验与更新资格,未改动渲染布局。正式安装 Host、原生 Windows、已签名 App 的实际替换与匹配模型试验未验证。测试使用隔离状态,未触发托管 CI。
English verdict: REQUEST_CHANGES — 8dbcbbf enables public Complete Turn before the existing adoption gate is met. File and SQLite recover and charge once. An older preview rejects the new attestation through its existing in-app update path.
|
This pull request has merge conflicts with Choose the remote for the base repository, not an out-of-date fork. git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEADFor a same-repository clone whose Keep the DCO |
loopx-agent
left a comment
There was a problem hiding this comment.
Reviewer: model_agent · Antigravity · Google Deepmind · self_reported
Request changes conclusion (author-owned PR; GitHub blocks formal self-review)
[P1] CLI 输出差分回归 smoke 在精确 head 8dbcbbf7a0 失败(3 条 quota_should_run_turn_envelope* 变体全挂)。
- Problem:
tests/control_plane/test_cli_output_budget.py:621将quota_should_run_turn_envelope变体模板里的--turn-envelope删除,而examples/control_plane/cli-output-base-head-differential-smoke.py用候选分支的变体定义去探测基线64353773dc。 - Evidence: 运行
.venv/bin/python examples/control_plane/cli-output-budget-regression-smoke.py退出码1,variant/quota_should_run_turn_envelope/small/json、variant/quota_should_run_turn_envelope_blocking_user_gate/small/json与variant/quota_should_run_turn_envelope_multi_subagent/small/json全部失败(chars 增长 1709/772/1718 超过 256 上限,output_contract_version changed,semantic_json_keys removed: decision, goal_id, should_run,markdown_anchor changed)。原因是基线未传--turn-envelope输出quota_packet,但loopx/control_plane/testing/cli_output_differential.py:740-771(_complete_quota_source_migration)仍按aliases要求基线source_schema == "loopx_turn_envelope_v0"。 - Fix: 修正差分烟雾与
_complete_quota_source_migration对已退役基线变体的探测/分类规则,避免用无开关quota_packet基线对照loopx_turn_envelope_v0别名契约,并复跑cli-output-budget-regression-smoke.py。
[P1] Benchmark 与 Checkpoint 退役护栏破坏现有测试及已跳过的 sforge 用例。
- Problem:
benchmark/runtime/checkpoint.py:234新增reject_retired_turn_display(env)(benchmark/runtime/codex.py:90-94),Execution.__post_init__(benchmark/runtime/codex.py:47-48)拒绝turn_envelope != False且benchmark/runtime/sforge.py:530移除turn_envelope/turn_envelope_source回执键,但对应测试未同步迁移。 - Evidence: 运行
.venv/bin/python -m pytest tests/control_plane/test_research_checkpoint_capture.py -q在test_off_preflight_pages_the_complete_saved_view[True-file]与[True-sqlite]直接失败(ValueError: LOOPX_COMPLETE_TURN requires the recorded historical product and runner revisions);同时直接调用Execution(mode="heartbeat", turn_envelope=True)抛出相同ValueError,证明benchmark/tests/test_sforge_runtime.py:334-362(当前因本地缺sforge/harbor被importorskip跳过)在安装依赖后必然失败。 - Fix: 移除
tests/control_plane/test_research_checkpoint_capture.py:116-137已退役的LOOPX_COMPLETE_TURN=1参数化分支(如需覆盖拒绝可显式断言ValueError),并同步更新benchmark/tests/test_sforge_runtime.py:334-362。
[P1] public_default_contract.json 先于 RFC §9 门禁切换公开默认,且未向 pre-PR macOS preview 提供过渡路径。
- Problem: 提交
24bc5086a4将loopx/cli_commands/public_default_contract.json的current改为complete_turn(loopx_turn_envelope_v1),并放宽loopx/cli_commands/update_artifact.py:112-119(validate_default_contract)与apps/desktop/loopx-control-plane/src-tauri/src/macos_update.rs:14-58。 - Evidence: 基线
docs/architecture/rfcs/controller-executor-unified-contract-v0.md§9(3495df117229483b5104f0a68cb93034bc676de2)与现行docs/product/release-readiness.md:296-301明确规定在完成声明 DeepSeek 与强模型的匹配单会话结果及全量成本对照前,公开分发默认保持quota_packet;本 PR 附件contract-and-context.md:35-37也明确承认支持 Host 采用与匹配单会话成果/成本门禁仍未完成。同时对照64353773dc:apps/desktop/loopx-control-plane/src-tauri/src/macos_update.rs:49-51,旧版 macOS preview 仅接受单一quota_packet描述符,遇到complete_turn归档会直接返回update_platform_unavailable。 - Fix: 将公开制品默认值
public_default_contract.json/validate_default_contract保留在quota_packet,把 CLI 组合收敛与公开默认切换解耦,或在补齐 §9 实证对齐后同时给出已签名过渡版本/安装升级方案。
[P2] 发行说明内部自相矛盾,且自定义 Runner 指南仍引用已退役的 --turn-envelope。
- Evidence:
docs/product/releases/v2.4.0.md:92改为默认向外部 Host 返回loopx_turn_envelope_v1,但同文件第 101 行及docs/product/release-readiness.md:296-310仍写普通无宿主 JSON 默认保留quota_packet(public_default_contract.json);docs/guides/custom-agent-runner-integration.md:168与.zh-CN.md:152仍出现--turn-envelope。 - Fix: 依据最终发布边界统一
v2.4.0.md、release-readiness.md与中英文自定义 Runner 指南。
动机 / Motivation
普通 CLI 调用方、外部 Host 集成方、heartbeat/checkpoint runner、benchmark 工具链以及桌面端/运行时更新校验器都需要读取同一套工作契约与制品默认策略,而此前 --turn-envelope 与 --complete-turn 开关导致行为分裂。
Ordinary CLI callers, external host integrations, heartbeat/checkpoint runners, benchmark harness operators, and desktop/runtime release updaters consuming quota should-run and runtime artifact descriptors. These callers need one unambiguous contract without overlapping format switches.
改动前,普通 JSON quota should-run 在未声明宿主时退回 quota_packet,只有显式传递 --turn-envelope 或 --complete-turn 才会组合 Complete Turn v1;改动后,普通 quota should-run 直接返回 loopx_turn_envelope_v1 并拒收两个旧格式参数,--context-delivery off --delta-delivery off 返回内联 v1,--scheduler-observation 保持只读观测,但仍有三组回归与差分测试失败/损坏、public_default_contract.json 在 §9 未达标前切换了公开默认、且文档自相矛盾。Before this change, calling JSON quota should-run without host flags returned quota_packet while --turn-envelope and --complete-turn opted into Complete Turn v1. After this change, ordinary quota should-run returns loopx_turn_envelope_v1 directly and rejects --turn-envelope/--complete-turn as unrecognized arguments, while --context-delivery off --delta-delivery off returns inline v1 and --scheduler-observation stays read-only; however, three differential/checkpoint/benchmark checks now fail or break on retired selectors, public_default_contract.json flips to complete_turn before matched outcome/full-cost qualification, and docs contradict each other.
就已核验的可观测结果而言:在精确 head 8dbcbbf7a0 上,核心 CLI 组合、显式关闭传输、只读调度观测与历史封装识别已通过大量单测,但上述阻塞回归与公开默认值越门问题尚未解决。On verified head 8dbcbbf, CLI quota should-run composes loopx_turn_envelope_v1 by default, rejects --turn-envelope, isolates --scheduler-observation, and preserves sealed historical captures/bootstraps (6700 TS tests, 148 quota/update tests, 28 delegation tests, 28 desktop bundle tests pass), whereas cli-output-budget-regression-smoke.py fails 3 differential rows, test_research_checkpoint_capture.py fails 2 preflight cases, skipped test_sforge_runtime.py:340 breaks against Execution, and distribution/doc consistency regresses.
非目标(non_goals):本 PR 不承担也未证明正式已安装 Host 采用、原生 Windows 桌面替换,或 RFC §9 的匹配单会话任务结果与完整成本净收益。Does not establish real installed-host adoption, Windows native desktop replacement, or RFC section 9 matched single-session task outcome and full-cost net benefit.
改动思路 / Approach
本 PR 复用 TypeScript 既有 composeTurnDecision 作为唯一工作契约组合层,保留 Python 在锁、guard 准入、捕获落盘与渲染/展示上的入口职责,并通过移除 is_undeclared_json_quota_caller 让普通 quota should-run 与显式关闭传输统一产出 loopx_turn_envelope_v1,仅在 quota_result_delivery.py:233 对 --scheduler-observation 保留前置早退。对于历史资产,选择“只读兼容、禁止新用”的边界:loopx/hosts/codex_checkpoint.py:367-413 保留对已封存中断检查点旧参数的校验,loopx/control_plane/heartbeat/bootstrap_prompt.py:166-199 识别旧提示词封装并在升级时剥离旧开关,loopx/cli_commands/quota_context_reader.py:75 仅在重新呈现选取命令时移除旧参数而不篡改原捕获字节;Benchmark 侧则在 benchmark/runtime/codex.py 与 checkpoint.py 起跑前 fail-closed 拦截历史参数。
| 路径 / 调用方 | 基线 64353773dc 行为 |
精确 head 8dbcbbf7a0 实测结果 |
|---|---|---|
普通 JSON quota should-run(无旧开关) |
返回 quota_packet(is_undeclared_json_quota_caller 跳过组合) |
返回 loopx_turn_envelope_v1(observe_default_contract 实测通过) |
quota should-run --turn-envelope |
返回 loopx_turn_envelope_v1 |
CLI 解析阶段 exit 2 报错 unrecognized arguments |
quota should-run --scheduler-observation |
返回 13 个调度只读字段 | 仍提前早退,不触发 compose_quota_work_contract |
cli-output-budget-regression-smoke.py |
校验变体差分 | FAIL:3 条 quota_should_run_turn_envelope* 变体失败 |
test_research_checkpoint_capture.py |
[True-file]/[True-sqlite] 通过 |
FAIL:LOOPX_COMPLETE_TURN 触发 ValueError |
public_default_contract.json + 旧 preview |
单一 quota_packet |
切换为 complete_turn,旧 64353773dc preview 拒绝升级 |
架构边界评估(architecture_assessment: simplify_now):The CLI composition cutover and selector retirement belong in the existing TypeScript/CLI/benchmark owners and save -311 net lines, but bundling public_default_contract.json promotion before RFC section 9 qualification violates docs/product/release-readiness.md:296-301 and breaks older macOS previews. 当前建议边界(current_pr_boundary):Keep CLI C1 composition cutover, selector retirement, sealed historical recovery, and fixed differential/checkpoint/sforge tests + docs; revert premature public_default_contract.json / update_artifact.py default promotion (or attach completed RFC section 9 qualification + old-preview transition bridge).
具体改动 / Changes
本次对比基线 64353773dc8b1ca82ff531cc7653dec6c3d41c0b 与精确 head 8dbcbbf7a05f5cea8f9ac837958cd15420fa6a11,共 83 个文件、+518/-829(生产代码 24 个文件 +94/-242;测试与 smoke 46 个文件 +324/-479;文档 13 个文件 +100/-108)。依据基线规范 docs/architecture/rfcs/controller-executor-unified-contract-v0.md(提交版本 3495df117229483b5104f0a68cb93034bc676de2)及 docs/product/release-readiness.md:287-310 逐条对照:
C1:not_met。CLI 默认组合、内联--context-delivery off --delta-delivery off与版本固定历史恢复已落地,但public_default_contract.json与update_artifact.py在未完成 §9 声明 DeepSeek + 强模型匹配单会话结果与完整成本对照前即切换了公开制品默认(本 PR 所附contract-and-context.md:35-37也明确该门禁仍未关闭,与release-readiness.md:296-301冲突)。C9:not_met。旧开关在当前 CLI/heartbeat/benchmark 退场的同时,破坏了cli-output-budget-regression-smoke.py、test_research_checkpoint_capture.py与跳过的test_sforge_runtime.py:334-362,未桥接64353773dcmacOS preview 升级路径,且v2.4.0.md:101与custom-agent-runner-integration*.md仍留有旧描述。
关键代码讲解
execute_quota_should_run与dispatch_quota_turn_start_hooks:去掉is_undeclared_json_quota_caller例外,统一以not getattr(args, "scheduler_observation", False)控制模型侧完整上下文与context_navigation收集。compose_quota_work_contract:由quota_result_delivery.py:326(在第 233 行--scheduler-observation早退之后)调用,直接委托compose_turn_decision(..., composed_settlement=True)生成loopx_turn_envelope_v1,并在远程/启动异常时写入失败诊断保留 guard 回执。validate_default_contract与qualify_candidate:共享loopx/cli_commands/public_default_contract.json,Python 发布端仅接受current(现已被改为complete_turn),新版 Rust 校验接受current或historical(quota_packet),但旧版 macOS preview 仅认单对象quota_packet。Execution.__post_init__与reject_retired_turn_display:在benchmark/runtime/checkpoint.py:234与 worker 启动前拦截非 False 的turn_envelope/complete_turn以及非'0'的LOOPX_TURN_ENVELOPE/LOOPX_COMPLETE_TURN。_complete_quota_source_migration与test_cli_output_budget.py:后者移除quota_should_run_turn_envelope变体参数后导致基线输出回落为quota_packet,与前者期望的loopx_turn_envelope_v0别名校验冲突而触发烟雾失败。
对主干的风险 / Risks
- 隔离性与调用边界(
scope_fit/default_off_isolation):所有 24 个生产文件变动均挂接在活跃入口路径上;--scheduler-observation(quota_result_delivery.py:233)与checkpoint.py:206feature-off 隔离经实测成立。但一旦合入当前 head,主干上的差分输出烟雾与研发检查点测试会立即变红,且带sforge环境的 benchmark 套件会暴露test_sforge_runtime.py:334-362崩溃。 - 负向与兼容路径(
failure_analysis):tests/test_runtime_snapshot_publication.py、tests/test_update_artifact.py与tests/control_plane/test_host_bootstrap_lifecycle.py证明当前源码自洽,但掩盖了两层真实脱节——其一是64353773dc已发布的 macOS preview 无法原地校验complete_turn新包,其二是跳过sforge/harbor掩盖了SForgeWorker(..., turn_envelope=True)抛错。 - 本地复现验证(
wait_for_ci: false,未参考远程 CI,全部在8dbcbbf7a0独立运行):npm run test:control-plane && npm run typecheck:control-plane:PASS(6700 passed, 0 failed, 91 skipped [postgres];tsc --noEmit退出 0)。tests/test_runtime_snapshot_publication.py+tests/test_update_artifact.py+test_quota_closeout_query_failure.py+test_quota_turn_envelope_validation_failure.py+test_turn_envelope.py+test_protocol_packet_history.py:PASS(148 passed)。tests/test_delegation_turn_context.py(隔离干净 worktree):PASS(28 passed, 2 skipped)。tests/control_plane/test_cli_output_budget.py+test_cli_output_probe_runner.py+tests/architecture/test_work_lane_contract_schema_version_single_owner.py+test_semantic_vocabulary_drift.py:PASS(198 passed, 1 skipped)。tests/control_plane/test_host_bootstrap_lifecycle.py+test_automation_prompt_upgrade.py:PASS(66 passed);test_refresh_settlement_recovery.py:PASS(13 passed, 2 skipped);test_phase_probe_continuation.py:4 skipped;tests/desktop:PASS(28 passed);benchmark/tests/test_sforge_runtime.py+test_sforge_turn.py+test_edgebench_registration.py:12 passed, 282 skipped(缺sforge/harbor)。examples/control_plane/cli-output-budget-regression-smoke.py:FAIL(exit 1,3 条variant/quota_should_run_turn_envelope*失败)。tests/control_plane/test_research_checkpoint_capture.py:FAIL(2 failed[True-file]/[True-sqlite], 52 passed, 11 skipped;隔离 worktree 复测-k test_off_preflight_pages_the_complete_saved_view:2 failed, 2 passed)。Execution(mode="heartbeat", turn_envelope=True)实测抛出ValueError(对应benchmark/tests/test_sforge_runtime.py:340损坏)。examples/semantic-vocabulary-drift-smoke.py(ok)、scripts/check_rfc_bilingual.py(passed)、ruff check(passed)、loopx check --scan-path ...(0 errors)、loopx canary premerge --tier quick(direct 6/6 + selected 4/4 passed,因benchmark_sensitive保留人工门):PASS;cargo test --no-run(src-tauri):UNVERIFIED(缺预处理产物../runtime/runtime-source.tar.gz,在build.rs:5中断)。
语义与 CI 对齐
本 PR 复用既有 loopx_turn_envelope_v1、quota_settlement_plan_v2 与 public_default_contract.json 词汇(semantic-vocabulary-drift-smoke.py 通过,Vision 夹具预算从 59,000 调高到 61,100 也在 docs/development/testing-and-quality.md 给出了 60,826 字符实测依据),但因(1)cli-output-budget-regression-smoke.py 的 3 条变体差分契约失败与(2)public_default_contract.json 在未过 RFC §9 门禁前更改公开支持默认值,判定 semantic_alignment 为 violated;需修复 test_cli_output_budget.py:621 / cli_output_differential.py 并复跑 cli-output-budget-regression-smoke.py,同时恢复 public_default_contract.json 与 release-readiness.md:296-301 一致性。
我的整体评价 / Verdict
综合判断本 PR 在精确 head 8dbcbbf7a05f5cea8f9ac837958cd15420fa6a11 上必须 REQUEST_CHANGES:long_horizon 为 regression(test_research_checkpoint_capture.py、sforge checkpoint 用例与差分回归烟雾均出现明确破坏,且公开默认切流先于 §9 结果与成本证据),user_experience 为 regression(发行说明、发布就绪规范、自定义 Runner 指南互相冲突且旧 macOS preview 无法平滑升级;visual_review 为 not_applicable:Inspected all 83 changed files in 6435377..8dbcbbf: touches CLI, control-plane, benchmark adapters, desktop macos_update.rs updater signature/descriptor validation, and documentation only; zero web UI or rendered dashboard presentation files changed.)。
在基线对比方面,observable_semantics 判定为 unintended_drift(虽已验证 quota should-run 切至 loopx_turn_envelope_v1、拒收 --turn-envelope、保持 --scheduler-observation 隔离,但基线/head 差分回归在 3 条变体 row 报错且文档滞后);code_volume 为 necessary(净删 311 行,兼容策略 retain 合理保留了封存检查点与旧 bootstrap 识别);change_proportionality 为 proportionate,但其 architecture_assessment 为 simplify_now——The CLI composition cutover and selector retirement belong in the existing TypeScript/CLI/benchmark owners and save -311 net lines, but bundling public_default_contract.json promotion before RFC section 9 qualification violates docs/product/release-readiness.md:296-301 and breaks older macOS previews. 当前 PR 边界应收敛为:Keep CLI C1 composition cutover, selector retirement, sealed historical recovery, and fixed differential/checkpoint/sforge tests + docs; revert premature public_default_contract.json / update_artifact.py default promotion (or attach completed RFC section 9 qualification + old-preview transition bridge). default_off_isolation 为 isolated,authority_semantics 为 not_applicable。语义词汇方面本 PR 复用既有 loopx_turn_envelope_v1 / loopx_interaction_contract_v0 词汇(reuse_existing),未新增控制面词汇,但触发上述差分预算与分发契约违例(semantic_alignment: violated)。
重新送审需补充的验证证据:1)修复后的 cli-output-budget-regression-smoke.py 全绿结果;2)test_research_checkpoint_capture.py 全通过与 test_sforge_runtime.py:334-362 修正结果;3)将 public_default_contract.json / update_artifact.py 回退至 quota_packet(或补充 RFC §9 对照结论与 pre-PR macOS preview 升级实测 + Cargo 编译验证);4)对齐 v2.4.0.md、release-readiness.md 与 custom-agent-runner-integration*.md。
English verdict: REQUEST_CHANGES — 8dbcbbf breaks cli-output-budget-regression-smoke.py (3 variant/quota_should_run_turn_envelope* differential failures), fails 2 cases in tests/control_plane/test_research_checkpoint_capture.py, breaks skipped benchmark/tests/test_sforge_runtime.py:334-362, leaves contradictory release/runner docs, and flips public_default_contract.json before RFC section 9 matched outcome/full-cost qualification or bridging pre-PR macOS preview updaters.
Summary
--context-delivery off --delta-delivery off.Recovery
Saved captures retain their original digest and settlement identity. The current CLI renders a compatible selection command from a saved capture, and the checkpoint reader still validates archived command spellings. Reproduce historical benchmark treatments with their recorded runner and product revisions. The signed main-channel preview
0.0.0-main.44809957556.1has a quota-packet-only updater and rejects the C1 archive before installation. The rollout needs a signed bridge that this updater accepts or a verified manual App replacement.Evidence
Real installed-host adoption and matched single-session task outcomes and total cost remain separate acceptance work.