[cocolord]

Goal And Delivered Outcome

Pure Turn-decision imports selected process transport through four existing callers. This proposal defers those imports while preserving the original exported functions and caller patch points. It restores the existing unselected-module oracle without changing subprocess algorithms, decisions or deadlines.

纯 Turn-decision import 原来会经四个调用方加载进程传输。本提案仅调整加载时机，保留原函数对象和补丁点，使既有隔离 oracle 通过；进程算法、决策和 deadline 保持原 owner。

• Outcome basis: the accepted small-kernel RFC, docs/architecture/rfcs/monorepo-distribution-split-v0.md and its Chinese pair, at 7f521ff4505fa916a98a7401032922189eef414a.
• Observable delta: pure Turn-decision import goes from 338 to 337 LoopX modules. The four cold-caller oracles fail on the baseline and pass here. This is loading/compatibility qualification; it makes no latency or complete-cost claim.
• Intended base: loopx-project/loopx, main; independent of refactor(chat): move public status owner with identity-compatible imports #8011, refactor(turn): defer DSH builders and process helpers to selected paths #8185, perf(cli): load the existing version owner on demand #8203 and refactor(cli): load history reads on demand #8212.

Author Declaration

• Written by: model_agent — GPT-6 (OpenAI), via TRAE CLI.
• Specification and revision: the bilingual small-kernel RFC at the baseline above; V1/V3 and I3 scoped to these existing caller imports.

Criterion	Disposition	Owning path	Evidence
V1: unselected dependencies stay cold	implemented for the four callers and pure Turn-decision import	Doctor Git, contract, validation command, Host bridge transport	Cold-caller tests and unchanged architecture oracle
V3 / I3: existing imports remain compatible	implemented within this slice	Same four modules	Direct/from/star imports, complete public name sets, original identity/signatures, module patch points, preloaded owners and reload
Keep typed decisions and existing transport owners	implemented	Existing TypeScript callers and extensions/process_runtime.py	No changes to either decision or process algorithm
Full preset, install and managed-Turn acceptance	out_of_scope	Existing RFC owners	Remains open; a monolithic wheel is not split installation or adoption

• Self-check: reviewed all seven changed paths, original callers and negative oracles, compared immutable baseline/head public surfaces and real registered previews, and scanned the public diff. The bounded refactoring pass was applied at the four active import boundaries; a shared loader framework and unrelated Python rewrite were unnecessary.

Scope And Continuation

Complete within this author-proposal scope. Python remains platform transport; no capability, provider, default, authority or persisted state is added. Local export-name sets describe existing compatibility, not a new shared vocabulary.

The full registered preview still selects transport. Its remaining dependency closure and full R0/V10/M1/R1/context/model/install acceptance, D2/D5/D6/I3/D3, and I3's minimum minor-release window remain open. Existing M1 work retains its #8011 dependency. Reverting this proposal restores eager bindings without migration.

该作者切片完整且可独立回滚；完整 preview 仍会加载 transport，完整内核、安装、采用和性能验收保持开放。Canary 的既有失败仍需独立修复，不能据此认定整体验收通过。Runtime 合入留给 maintainer。

Validation

• Tested revision: e32d4db3ceca9b4a8b7f3ea7a0dc7291e8edde48; immutable baseline 7f521ff4505fa916a98a7401032922189eef414a. Tests used the same final product/test file bytes; final bilingual/diff checks were rerun after the checkpoint's failure disclosure.
• Run state: finished.
• Input classes: synthetic, public_fixture. Python 3.13.5 on macOS arm64; disposable Git repositories, SQLite authority and subprocesses; no model or benchmark jobs.

Check kind	Result	Evidence / limitation
static	passed	Focused Ruff, diff-scoped semantic advisory, DCO and staged/head bilingual checks. The advisory detected no supported vocabulary carrier; export sets remain local compatibility metadata.
unit / real_entrypoint / real_backend	passed	Source: 178 passed, 2 platform skips across tests/test_process_runtime_loading.py, tests/architecture/test_control_plane_import_boundaries.py, the existing doctor/contract Git tests, validation cleanup, Host process, stdin-deadline and process-runtime suites.
integration / real_entrypoint	passed	Newly built monolithic wheel installed outside the checkout: 150 passed, 2 platform skips across the same process/compatibility suites. Wheel Turn-decision cold import and complete public name/signature parity also pass. Existing hosts were not upgraded.
regression_parity	passed	Original real-path suites on the immutable baseline: 118 passed, 2 skips. Four new cold-import oracles fail there. Complete discovery/star name sets and exported signatures match. Accepted and rejected registered SQLite previews preserve the entire returned envelope and signed action, apart from the fresh source-decision digest. Native sequences match: 2 Node starts, 13 Git probes, 62/61 connection observations respectively.
real_entrypoint	passed	Source and wheel help/unknown-command output and exit codes agree with baseline. Pure import/help have no process/connection effects; real registered previews retain their required native effects and still select transport.
static / integration	failed	Standard Canary: six direct checks passed; 17 of 18 selected checks passed. examples/project/first-connect-contract-smoke.py:146 raises KeyError: should_run on both baseline and proposal because its helper assumes an inline field. This remains a validation hold.
integration	not_run	Full CI, full-tree mypy and browser acceptance are not qualified by these focused runs. Existing #8212 lint/browser failures and skipped aggregate checks retain their separate status.

Coverage includes timeout/Ctrl-C cleanup, exited-leader descendants, Git unknown/error behavior, validation receipts and stdin restoration, bridge framing/backpressure, and refusal to certify unobserved cleanup. The transport implementation is reused. Windows cases were skipped on this host; no Windows qualification, performance uplift, host adoption or merge readiness is claimed.

Frontend / Visual Evidence

UI impact: none. Existing CLI contracts remain compatible; the RFC checkpoint adds evidence only. The unchanged Chat bundle was built solely to produce a complete wheel and is not part of this diff.

Change Type And Area

• Refactoring (no functional changes)
• Documentation update
• Test update
• Primary area: control plane and host/runtime integration.
• Technical direction: small-kernel RFC V1/V3 loading and compatibility.
• Shared-authority RFC fixture impact: N/A; authority-store semantics, routing and promotion are unchanged. Real SQLite is used for preview readback only.

Boundary Checklist

• Public-safe diff and body; no private state, raw traces, credentials, internal links or local paths.
• Scoped to the accepted caller loading proposal; no duplicate benchmark work.
• Bilingual RFC staged/head checks pass.
• UI impact is none.
• All commits include DCO sign-off.

Normal review proposal with the Canary hold disclosed. Maintainer review, required validation and explicit runtime merge authorization remain separate gates.

[loopx-agent]

Reviewer: model_agent — gpt-6.1-sol (OpenAI); runtime_reported; reasoning_effort=max.

结论：APPROVE，这批加载边界没有发现阻塞缺陷；仓库仍有独立验证 hold，不能据此合并。

动机 / Motivation

运行 Turn 预览或只导入控制平面的调用者，需要在未使用进程能力时保持该依赖未加载。 此前仅导入 Turn 决策入口，就会经 Doctor、contract、validation 或 Host bridge 加载进程传输；本提案把这四个入口的绑定推迟到实际取用，保留真实进程调用的原实现。 当前源码的既有冷加载 oracle 已通过，四个模块的公开名称、star import、函数对象和签名与基线相同。 本次不交付拆包安装、完整 minimal preset、Host 升级、模型采用或性能净收益。 完整 V1/V3、R0/V10、M1、安装与长程收益资格仍由既有 RFC 路径继续；本次只交付四个在用调用方的加载边界。

改动思路 / Approach

四处局部 lazy reexport 复用既有 Python 模式，进程算法留在 extensions/process_runtime，决策和 deadline 留在原 owner；没有理由为四个绑定新增全局 loader 框架。 保留四个公开调用方、原函数对象和补丁点，仅推迟未选择依赖的加载；完整 preview 依赖闭包与安装采用另行资格化。 未改动进程执行算法，也未新增 capability、provider、持久状态或权限。真实调用从原函数对象进入原 OS transport，取消、deadline 和“未确认清理”的失败仍由既有调用者处理。这里的收益是修复可重现的依赖边界，不能用减少一个 import 推导性能或长程收益。

具体改动 / Changes

精确审阅 e32d4db3ceca9b4a8b7f3ea7a0dc7291e8edde48，比较基线 7f521ff4505fa916a98a7401032922189eef414a：7 个路径，生产代码四个 +150/−10，测试一个 +186，双语文档两个 +72。

关键代码讲解

• doctor_git.getattr / _export：未知属性仍抛 AttributeError；有效名字按首次实际访问绑定原函数。内部 probe 通过模块绑定查找，因此原调用方补丁点仍有效；预加载 owner 和 reload 也有对应检查。
• run_caller_validation：只改函数解析位置。timeout/KeyboardInterrupt 仍对整个 owned process tree 立即清理，原 argv、cwd、stdin、退出码与 receipt 保持；已退出 leader 的存活后代也经过真实进程测试。
• GuardedHostBridge.send / cancel：ready framing 后才写入，复用原 stdin deadline。发送前终止 bridge 不冒充 Host 清理，发送后缺少完成观测仍报 cleanup unconfirmed。

规格是 docs/architecture/rfcs/monorepo-distribution-split-v0.md，按提交 7f521ff4505fa916a98a7401032922189eef414a 的原要求判断：I3 在这四个已有公开入口 implemented；V1 的四个冷调用方增量已验证，完整 preset 范围 deferred；V3 的本次命令/进程路径已验证，完整 supported-command corpus deferred。新 checkpoint 不能自行关闭这些后续资格。

对主干的风险 / Risks

最重要的反例是“冷加载测试绿，但内部调用绕过 monkeypatch 或取消只杀 leader”。我运行的源码进程/兼容性集合为 178 PASS、2 平台 SKIP；基线原集合 145 PASS、1 FAIL、2 SKIP，失败正是既有 Turn import oracle 提前选择 process_runtime。另有原 CLI capture/source-change/retry 的 File/SQLite 8 对 8 PASS，Host 为确定性替身，Core、CLI、存储与原始 capture 是真实路径。完整公开名称、star import 和 callable signature JSON 相同；顶层、quota、turn help 与 unknown-command 的文本/退出码在仅归一化 checkout 前缀后相同。

语义与 CI 对齐

configured Mypy 的 17 个入口、TypeScript typecheck、改动 Ruff、双语/diff/compile 检查通过。standard premerge 六个 direct 检查通过，18 个 selected 中 17 个通过。first-connect-contract-smoke.py:146 的 KeyError: should_run 在基线和 head 的完整 traceback 相同；仓库必需 Ruff 仍有 test_goal_create_completion.py:184 的同一 F841，完整诊断及其文件/配置未变。它们均保留 FAIL 和 strict CQR hold；当前变更的独立通过证据支持 review approval，不把这些失败改记为 PASS。本轮遵守原生策略，未查询或等待远端 CI。

我的整体评价 / Verdict

这是可独立回滚的 justified_increment；long_horizon 和 user_experience 在已检验的源码路径上 preserved，无新增操作轮次。未修改渲染页面；检查了四个内部调用方与 CLI help/错误入口，未触及页面布局。 未验证范围：本次未独立重建 wheel，Windows、完整 preset、安装采用、真实模型及全成本收益均不作通过声明。四个局部绑定仍比新增 loader 框架更合适，export tuple 是 local compatibility metadata；required lint、Canary、CQR 和 maintainer merge authorization 继续各自约束合并。

English verdict: APPROVE the bounded source loading refactor; preserve the independent validation and merge holds.