[loopx-agent]

Goal And Delivered Outcome

The read-only loopx turn plan command eagerly loaded process cleanup and extension recall code. In a fresh process, the preview first loaded the cleanup helper through the public-boundary scanner, then reached the same helper through independent preference and validation imports. Load each owner when its operation selects it.

On the same synthetic File and SQLite Goals, the source checkout retained 661 → 653 loaded Python modules after turn plan. The import trace also fell from 663 to 654 entries. One omitted entry was a second reference to an existing module. Both revisions returned the same route, selected Todo and iteration context. The preview kept the Goal and registry intact.

Author Declaration

Written by model_agent, OpenAI Codex.

Implemented against the Turn preview isolation assertion added by #8185 and the bounded Git probe cleanup from #8152.

Criterion	Disposition	Owner and evidence
Preview keeps unselected process and recall owners unloaded	implemented	The existing fresh import assertion and the extended real CLI test detect the original failure.
Git timeout and cancellation stop the owned process group	implemented	The public-boundary scanner loads cleanup at retirement. Existing timeout, ownership and cancellation tests pass.
Validation retains process cleanup and input policy	implemented	Caller validation selects cleanup before launch. Real timeout, cancellation and validation receipt tests pass.
Public preference APIs and configured providers resolve their original owners	implemented	The five public exports retain object identity. Existing configured extension tests pass.

Scope And Continuation

The patch covers the Python import paths of the CLI preview and their selected operations. It can be reviewed and reverted independently of the complete Turn cutover. The comparison measures loaded modules in fresh source-checkout processes.

Validation

• Base 8e86f173f08bd55eb833d8cdf42360f93bf989fe
• Tested revision c65486ef2af5e5d46ce31517bd284507a0b8888a
• Run state finished with synthetic fixtures

Check kind	Result	Public-safe evidence
real_entrypoint	passed	Fresh loopx turn plan on isolated File and SQLite fixtures. Both reject the baseline's eager imports and pass at this head. The public CLI scan agrees for both revisions on clean files and tracked private-state refusal.
integration	passed	214 passed, with one Windows stdin case skipped on macOS. Two additional configured semantic-preference extension cases passed.
static	passed	Focused Ruff and explicit public-boundary scan. The scan reports the same non-literal credential-pattern advisory as the baseline.
static	failed	Standard premerge passed six direct checks and 16 of 17 selected checks. The existing first-connect smoke reads should_run from the progressive navigation header and fails identically at base and head. Following the returned reader yields should_run=true and effective_action=normal_run at both revisions.

Evidence covers local source-checkout execution. Installed application and model-level performance remain unmeasured.

Frontend / Visual Evidence

UI impact none. The CLI output and app presentation are unchanged by this import-path change.

中文：默认 Turn 预览延迟加载进程清理及语义偏好召回模块。File 和 SQLite 实测各少加载 8 个 Python 模块，计划和公开边界扫描结果一致。原 Git 与校验进程的超时、取消测试继续通过。