cocolord · Nov 6, 2026 · Nov 6, 2026 · Nov 6, 2026 · Nov 6, 2026 · Nov 6, 2026
@@ -824,6 +824,91 @@ Approval needs a reviewed decision, not implementation progress or silence.
   installation, R1 managed minimal or D1–D7. No preset, frontend operation,
   permission or persisted state changes; protected execution keeps its owner.
 
+### 2026-11-06 — Selected version loading and matched costs
+
+The independent proposal `8a79d8602e699e51452ca1f43dc9b9527c81aa1b` at baseline
+`5a58635028fbef1316dc66531daf5fdb655e9b13` selects the existing version registrar
+and handler through `cli_runtime`. It reuses the canonical global grammar,
+format precedence, usage hooks and controller guard. Rejected parsing retains
+full-CLI fallback and diagnostics; explicitly bound package or full-CLI exports
+retain their registrar/handler patch points without triggering lazy exports
+on ordinary queries. Public imports, identity and types remain
+unchanged. No capability, provider, package, permission, preset or TypeScript
+decision owner is added. The separate DSH/helper and negative cost proposal is
+[#8185](https://github.com/loopx-project/loopx/pull/8185); its E7/E8 observations
+are distinct from this version-path comparison and are not a dependency.
+
+The initial `55e3ef45a` comparison retains 120 ordinary-query observations
+(source median 1.546 → 0.059 seconds; wheel 1.554 → 0.072). A later independent
+case finds that it ignores a patched package-facade handler, returning 0 instead
+of 37. Its passing premerge did not cover that branch. The failure and original
+inputs/data remain retained; they do not qualify facade compatibility. The
+proposal repairs both package registrar and handler adoption, then freezes the
+new exact-head comparison below.
+
+- **Real caller:** source and freshly installed monolithic wheels at both
+  immutable revisions exercise the actual `loopx` console. All 27 cases per
+  distribution have identical stdout, stderr and exit status, including
+  global/local/repeated formats, explicit paths, help, unknown/abbreviated
+  options, missing/empty values and fallback. Independent unknown-command/
+  option cases reject with exit 2 and no result. Version works with Node absent
+  in all four variants; outer-controller read behavior and full-CLI patch points
+  remain compatible.
+- **Loading census:** separate parent instrumentation records 1,157 → 10
+  LoopX modules in both distributions, with the candidate contained in the
+  baseline set. Source read-opens are 1,286 → 28 and wheel 1,289 → 29.
+  Each version capture has zero write-opens, subprocesses, observed local IPC
+  connections and covered socket bytes. Stdout remains 81 bytes; stderr is empty.
+- **Frozen comparison:** all four environments have verified provenance,
+  CPython 3.13.5 and identical 57 dependency versions. Stdlib, dependencies and
+  all 1,421 LoopX sources have valid timestamp bytecode in owned cache prefixes;
+  post-measurement/profile source/header inventories have no missing/invalid
+  caches and match the preflight. Every sample has fresh Python and disables
+  bytecode writes; disk cache is uncontrolled. No broker/warm-broker comparison
+  is made. Order is source then wheel, AB on even pairs and BA on odd pairs.
+  Both corrected groups complete 30 pairs: 120 accepted observations in 62.246
+  seconds. Including the first study, aggregate sampling is 166.191 seconds,
+  below the unchanged 1,200-second cap. All observations are retained.
+
+| Distribution | Baseline seconds, median / p95 | Proposal seconds, median / p95 | CPU seconds, median (base → proposal) | RSS MiB, median (base → proposal) | Median paired proposal minus baseline, seconds |
+| --- | --- | --- | --- | --- | --- |
+| Source | 0.980 / 1.056 | 0.055 / 0.075 | 0.520 → 0.037 | 90.17 → 16.81 | -0.9242 |
+| Monolithic wheel | 0.948 / 1.147 | 0.056 / 0.077 | 0.511 → 0.038 | 89.89 → 16.60 | -0.8993 |
+
+p95 uses nearest rank; paired differences and differences of arm medians are
+distinct statistics. Median latency falls about 94% in both source and wheel
+on this macOS arm64 environment; CPU and target RSS also fall. One-minute host
+load ranges from 8.25 to 9.06. This qualifies a material version-query
+improvement under frozen conditions, without adopting portable D6 budgets or
+claiming real managed-Turn/long-horizon savings. Usage covers the target and
+waited descendants; no child is launched in the census. Native detail and
+complete authoring/model/maintenance costs remain unknown.
+
+The heartbeat-to-final-freeze interval is about 2,404 seconds, including
+overlapping recovery, initial study, correction, dependency/build/install,
+test, cache and parity work; overlapping phases cannot be summed. Cache
+preparation is recorded separately. Corrected console/facade parity takes
+82.759 seconds, instrumentation 2.539 and four parent profiles 4.155. Earlier
+parity/instrumentation/profile costs of 136.239 / 3.830 / 7.089 seconds remain
+retained. Cross-study timing levels are not treated as a code comparison.
+Existing `performance-diagnosis` recipes show export loading and full parser
+construction in baseline sampled stacks; weights remain separate from repeated
+latency. Monolithic packaging is verified, without split-install or installed
+host-upgrade qualification.
+
+Seven selected-loading cases fail on the baseline and pass on the proposal.
+The package registrar/handler case passes on the baseline, fails on the initial
+candidate and passes on the correction; canonical diagnostic and full-CLI
+patch-point cases pass in both baseline and proposal. All 161 entrypoint cases and
+focused lint pass. Owner-scoped mypy with `--follow-imports=skip` fails with
+33 identical diagnostics at base/proposal after line normalization; this is
+neither a pass nor the full configured type check. Adoption requires final-head
+validation, independent review and merge readiness. I3's minimum one minor
+release window remains in force; no D3/expiry decision is made. R0's wider census,
+R1/minimal, D2/D5/D6, hard-lease/context/model and split-install acceptance remain
+open. The bounded refactor reuses its existing parser/handler; larger runtime
+and package changes retain their accepted owners.
+
 ## Appendix B: Decision log
 
 | Date | Decision | Owner / approval | Alternatives | Normative sections changed |
@@ -840,6 +925,7 @@ Approval needs a reviewed decision, not implementation progress or silence.
 | E4 | Selected CLI loading seam | `9cae3b28939f72fa4aadab5cfe49e19bcbdab372` | Linked entrypoint/CLI in Section 4 | Source-audited | No new timing claim |
 | E5 | DSH composition example | Pinned upstream commit | [`minimal.patch.yml`](https://github.com/deepseek-ai/deepseek-harness/blob/f3803c96974ad4a94195fb97e5c637e9c2a3afe3/packages/bundle/web-app/presets/minimal.patch.yml), [`standard.patch.yml`](https://github.com/deepseek-ai/deepseek-harness/blob/f3803c96974ad4a94195fb97e5c637e9c2a3afe3/packages/bundle/web-app/presets/standard.patch.yml), [`sdk-minimal`](https://github.com/deepseek-ai/deepseek-harness/blob/f3803c96974ad4a94195fb97e5c637e9c2a3afe3/packages/bundle/sdk-minimal/README.md) | Source-audited | Shared preset composition and a self-contained SDK bundle; not a LoopX host qualification or permission recommendation |
 | E6 | Pi core/context example | Pinned upstream commit | [`agent` core](https://github.com/badlogic/pi-mono/blob/f14813e96510a7cf7562a120e809c6a134af073d/packages/agent/README.md) | Source-audited | Agent/tool/event core with `transformContext`, `convertToLlm`, `prepareRequest` and `finishTurn` seams; no LoopX authority or performance claim |
+| E9 | Selected version loading and matched cost | `5a5863502` / `8a79d8602`; macOS arm64 / CPython 3.13.5 | Version checkpoint above; existing console/version and performance-diagnosis owners | Corrected two groups × 30 pairs; 120 observations; public-facade/parity/cache readback pass; material query improvement | Initial 120 observations and facade failure retained; public-safe aggregates; full Turn/minimal, split install, host upgrade and complete costs unqualified; E7/E8 remain with #8185 |
 
 ## Appendix D: Rejected or superseded alternatives
 

@@ -599,6 +599,69 @@ R0 可以先于目录移动交付，但不代表内核合格。R3/R4 可选，
   不关闭完整 R0/V10、M1、拆包安装、R1 managed minimal 或 D1–D7。不改变 preset、
   前端操作、权限或持久化状态；受保护执行仍由既有 owner 所有。
 
+### 2026-11-06 — Version 按需加载及匹配成本
+
+独立候选 `8a79d8602e699e51452ca1f43dc9b9527c81aa1b` 对比基线
+`5a58635028fbef1316dc66531daf5fdb655e9b13`，由 `cli_runtime` 选择既有
+version registrar 与 handler，复用规范全局语法、format 优先级、usage hook
+及 controller guard。解析拒绝保留完整 CLI fallback 与诊断；已显式加载完整
+CLI 或绑定 package 导出的调用保留其 registrar/handler patch 点；普通查询不会
+触发 lazy exports。公开 import、身份及类型不变。
+不新增 capability、provider、package、权限、preset 或 TypeScript 决策 owner。
+独立 DSH/helper 及负面成本提案仍在 [#8185](https://github.com/loopx-project/loopx/pull/8185)；
+其 E7/E8 与本次 version 对照分开，也不构成依赖。
+
+初始 `55e3ef45a` 对照的 120 个普通查询观察继续保留：source 中位数
+1.546 → 0.059 秒、wheel 1.554 → 0.072 秒。随后独立用例发现它忽略已替换的
+package-facade handler，返回 0 而非 37；此前通过的 premerge 未覆盖该分支。
+原输入、数据和失败保留，不将其视为 facade 兼容资格。提案补回 package
+registrar 与 handler 的采纳后，另行冻结下列精确 head 对照。
+
+- **真实调用：** 两个不可变版本的源码及新安装单体 wheel 均经实际 `loopx`
+  console 执行。每种发行形态的 27 个用例，stdout、stderr、退出状态完全一致，
+  覆盖全局／局部／重复 format、显式路径、help、未知／缩写参数、缺失／空值及
+  fallback。独立未知命令／参数 oracle 以 exit 2 拒绝且无结果。四个组合在缺少
+  Node 时均可查询 version；outer-controller 读行为及完整 CLI patch 点兼容。
+- **加载清点：** 独立 parent 插桩记录，两种发行形态均为 1,157 → 10 个 LoopX
+  模块，候选是基线的子集。Source read-open 1,286 → 28，wheel 1,289 → 29。
+  每个 version 捕获的 write-open、子进程、观测到的本地 IPC 连接及覆盖 socket
+  字节均为零。Stdout 保持 81 字节，stderr 为空。
+- **冻结对照：** 四个环境均核验来源，使用 CPython 3.13.5 及相同 57 个依赖
+  版本。标准库、依赖及全部 1,421 个 LoopX 源文件，在独占 cache prefix 具有
+  有效时间戳字节码；测量／profile 后的源码及 header 清点与预检一致，无缺失／
+  无效缓存。每次采样是新 Python 进程，禁止字节码写入；磁盘缓存不受控，不比较
+  broker／warm broker。顺序为源码先于 wheel，偶数对 AB、奇数对 BA。两组各
+  完成 30 对，共 120 个有效观察，采样 62.246 秒。计入首轮后合计采样
+  166.191 秒，低于不变的 1,200 秒上限；全部观察保留。
+
+| 发行形态 | 基线秒数，中位数 / p95 | 候选秒数，中位数 / p95 | CPU 秒数，中位数（基线 → 候选） | RSS MiB，中位数（基线 → 候选） | 配对候选减基线的中位数，秒 |
+| --- | --- | --- | --- | --- | --- |
+| Source | 0.980 / 1.056 | 0.055 / 0.075 | 0.520 → 0.037 | 90.17 → 16.81 | -0.9242 |
+| 单体 wheel | 0.948 / 1.147 | 0.056 / 0.077 | 0.511 → 0.038 | 89.89 → 16.60 | -0.8993 |
+
+p95 使用 nearest rank；配对差中位数与两 arm 中位数之差不同。本次 macOS
+arm64 条件下，source 与 wheel 中位延迟均约下降 94%，CPU 与 target RSS
+也下降。一分钟宿主负载为 8.25–9.06。本片证明冻结条件下的明显版本查询收益，
+不采纳可移植 D6 预算，不声明真实 managed Turn 或长程节省。用量覆盖 target
+及等待子进程，清点中未启动子进程；native 细节及完整编写／模型／维护成本未知。
+
+从本次唤醒时间戳到最终冻结约 2,404 秒，包含有重叠的入口恢复、首轮实验、修正、
+依赖／构建／安装、测试、缓存及一致性验证；重叠阶段不可求和。缓存准备单独记录。
+修正后的 console/facade 一致性为 82.759 秒、插桩 2.539 秒、四个 parent
+profile 4.155 秒；首轮 136.239 / 3.830 / 7.089 秒成本继续保留。
+不同轮的耗时水平不作为代码对照。既有
+`performance-diagnosis` recipe 在基线采样栈中显示 export loading 及完整 parser
+构建；栈权重与重复延迟分开。已验证单体打包，不代表拆包安装或已安装宿主升级。
+
+七项 selected-loading 用例在基线失败、候选通过。Package registrar/handler
+用例在基线通过、初始候选失败、修正后通过；规范诊断及完整 CLI patch 点在基线／
+候选均通过。完整入口 161 项及聚焦 lint 通过。Owner-scoped mypy 使用
+`--follow-imports=skip`，基线与候选在行号归一后均有 33 项相同错误，不是通过或
+完整配置的类型检查。采纳仍需最终 head 验证、独立评审及 merge readiness。
+I3 至少一个 minor release 的窗口继续有效，不作 D3／到期决策。R0 更广清点、
+R1/minimal、D2/D5/D6、hard-lease／context／model 及拆包安装验收仍开放。
+有界重构复用既有 parser/handler；更大 runtime/package 改动保留在其已接受 owner。
+
 ## 附录 B：决策日志
 
 | 日期 | 决策 | 负责人 / 批准 | 备选 | 变更的规范章节 |
@@ -615,6 +678,7 @@ R0 可以先于目录移动交付，但不代表内核合格。R3/R4 可选，
 | E4 | 所选 CLI 加载接缝 | `9cae3b28939f72fa4aadab5cfe49e19bcbdab372` | 第 4 节链接的入口/CLI | 源码审计 | 无新性能测量 |
 | E5 | DSH 装配范例 | 固定上游 commit | [`minimal.patch.yml`](https://github.com/deepseek-ai/deepseek-harness/blob/f3803c96974ad4a94195fb97e5c637e9c2a3afe3/packages/bundle/web-app/presets/minimal.patch.yml)、[`standard.patch.yml`](https://github.com/deepseek-ai/deepseek-harness/blob/f3803c96974ad4a94195fb97e5c637e9c2a3afe3/packages/bundle/web-app/presets/standard.patch.yml)、[`sdk-minimal`](https://github.com/deepseek-ai/deepseek-harness/blob/f3803c96974ad4a94195fb97e5c637e9c2a3afe3/packages/bundle/sdk-minimal/README.md) | 源码审计 | 共享预设装配与自包含 SDK bundle；不代表 LoopX 宿主资格或权限建议 |
 | E6 | Pi core/上下文范例 | 固定上游 commit | [`agent` core](https://github.com/badlogic/pi-mono/blob/f14813e96510a7cf7562a120e809c6a134af073d/packages/agent/README.md) | 源码审计 | Agent/tool/event core，以及 `transformContext`、`convertToLlm`、`prepareRequest`、`finishTurn` 接缝；无 LoopX authority 或性能声明 |
+| E9 | Version 按需加载及匹配成本 | `5a5863502` / `8a79d8602`；macOS arm64 / CPython 3.13.5 | 上述 version checkpoint；既有 console/version 及 performance-diagnosis owner | 修正后两组各 30 对，共 120 个观察；公开 facade／一致性／缓存读回通过；版本查询明显改善 | 首轮 120 个观察及 facade 失败保留；仅公开安全汇总；完整 Turn/minimal、拆包安装、宿主升级及完整成本未取得资格；E7/E8 保留在 #8185 |
 
 ## 附录 D：被拒绝或被取代的备选
 

@@ -71,7 +71,7 @@
 	"refresh-state", "checkpoint-context", "read-only-map", "reward", "operator-gate",
 }
 _SELECTED_COMMANDS = _STATUS_COMMANDS | GLOBAL_READ_COMMANDS | PROJECT_LIFECYCLE_COMMANDS | {
-	"todo", "quota", "task-lease", "change-window", "delegation", "turn", "doctor", "commands",
+	"todo", "quota", "task-lease", "change-window", "delegation", "turn", "doctor", "commands", "version",
 	"authority-archive", "coordination-shadow", "extension", "slash-commands",
 	"manager-inbox",
 	"agent-context", "agent-directory", "agent-capabilities",
@@ -396,6 +396,10 @@ def _build_selected_parser(command: str) -> LoopXArgumentParser:
 		from .help_surface import register_command_reference
 
 		register_command_reference(subparsers)
+	elif command == "version":
+		from .cli_commands.version import register_version_command
+
+		register_version_command(subparsers, add_subcommand_format)
 	elif command == "performance-diagnosis":
 		from .capabilities.performance_diagnosis.cli import register_performance_diagnosis_commands
 
@@ -710,7 +714,13 @@ def _dispatch_selected(args: argparse.Namespace, raw_argv: list[str]) -> int:
 	if guard_result is not None:
 		return guard_result
 	registry_path, _registry_was_configured = resolve_cli_registry(args, raw_argv)
-	if args.command == "capability":
+	if args.command == "version":
+		from .cli_commands.version import handle_version_command
+
+		result = handle_version_command(
+			args, output_format=output_format, print_payload=print_payload,
+		)
+	elif args.command == "capability":
 		from .cli_commands.capability import handle_capability_command
 
 		# Keep the explicit full CLI's handler and protected operations unchanged.
@@ -766,6 +776,16 @@ def main(argv: list[str] | None = None) -> int:
 
 
 def _run_command(command: str | None, raw_argv: list[str]) -> int:
+	if command == "version":
+		version_exports = sys.modules.get("loopx.cli_commands")
+		# Inspect bound exports without triggering the lazy facade. Explicit
+		# package/full-CLI imports retain their registrar and handler patch points.
+		if "loopx.cli" in sys.modules or (
+			version_exports is not None
+			and ("handle_version_command" in vars(version_exports)
+				or "register_version_command" in vars(version_exports))
+		):
+			return _run_full_cli(raw_argv)
 	if command in _SELECTED_COMMANDS:
 		try:
 			args = _build_selected_parser(command).parse_args(raw_argv)