[loopx-agent]

Goal And Delivered Outcome

修复可复现的续跑诊断缺陷：原始 decision capture 缺失或被改动时，turn inspect-journal --retry-failed-turn 仍报告可以继续，而 turn run-once 会在调用 Host 前拒绝。现在普通检查与 --watch 均复用执行入口的只读校验，在需要重新调用 Host 时明确拒绝不可用捕获；恢复原字节后仍可继续原 Turn。

• Intended base: current main@7e03545abae842db8e15cf3b6f64a98312a0e4d8. The implementation and specification baseline is 8f6e88cb2a0fb76aba32bce979ba51293c1981f8; the current target and spec readback use the current base.
• Outcome basis: the reproduced defect above and docs/reference/effect-interpreter-packet.md, “Public read-only consumer”, at that revision. No separate public issue.

Author Declaration

• Written by: model_agent — OpenAI Codex.
• Self-check: reviewed the shared typed recovery decision, existing capture IO verifier, CLI/watch paths and native internal inspection callers. This changes the existing diagnostic path; it adds no operation or UI journey.

Criterion	Disposition	Implementation / evidence
Match execution preflight before Host reinvocation	implemented	executor.inspect_loopx_turn_journal consumes the existing typed action and reinvoke_host fields, then calls stage_turn_decision_capture under the journal lock
Refuse absent, altered and invalid original captures without effects	implemented	20 regression cases across v0/v1 and normal/watch modes; canonical journal and session bytes stay unchanged
Preserve saved-result recovery and terminal replay	implemented	File/SQLite real CLI cases recover validation without capture, invoke no additional Host and settle once
Restore original input and continue the original Turn	implemented	File/SQLite × historical v0/current v1 cases resume the same session with the original source, then replay without repeated effects

Scope And Continuation

Complete within this author scope: the diagnostic repair, durable regression coverage and owning reference documentation. Independent PR review, required CI, the uncompleted full-suite qualification and merge remain separate gates.

Bounded refactoring was applied at the same boundary: inspection now holds the existing journal lock through assessment and the read-only input check. It reuses the verifier and typed decision rather than adding another state machine or helper. The low-level require_decision_capture=False compatibility default matches the executor API; the public CLI explicitly enables strictness in both modes. No new semantic vocabulary, persisted state or configuration is introduced.

Validation

• Tested revision: 9732c420c6852d6c7ed482044c1955a5af42277d.
• Run state: finished — the local commands have ended; full-repository validation remains incomplete.
• Input classes: synthetic, public_fixture.

Check kind	Result	Public-safe evidence / limitation
unit	passed	python -m pytest -q tests/test_turn_inspection_session_host.py tests/test_loopx_turn_journal_inspection.py tests/test_loopx_turn_driver.py: 251 passed on the current-main head.
regression_parity	passed	The updated inspection suite against pinned base 8f6e88c: 20 expected negative regressions fail, 14 existing cases pass. Final candidate passes all 34 cases, including inspection after restoring original bytes
real_entrypoint / real_backend	passed	Eight isolated real CLI scenarios over the File provider and SQLite 3.53.4 (Node.js 24.21.0), with synthetic Codex transport. Historical input is generated by released v2.2.0; the candidate consumes both v0 and v1 records. Four scenarios verify refusal/restoration/same-session single settlement; four verify retained-result recovery and terminal replay without capture
static	passed	Repository Python lint roots plus changed CLI adapter; mypy: 17 source files; whitespace and bilingual RFC checks
static	passed	Changed-path public scan: zero errors/warnings. Full public-tree scan: zero errors, 49 advisory warnings outside the changed paths
integration	passed	Standard premerge against current target 7e03545aba: 6 direct and 14 selected checks pass, including the CLI hard budget, public boundary and exact-scope quality verification.
unit	failed	Full local repository run was interrupted (exit 130): 1 failed, 6017 passed, 841 skipped and 56 subtests passed. The failed case is tests/test_chat_agent.py::test_transient_recovery_requires_complete_native_terminal_and_effect_readback[native-error-before-terminal-info0]. An earlier run also had five import-path/environment failures; those five pass after clearing inherited Python path controls
regression_parity	passed	That Chat recovery case passes in a matched isolated base/head check with the same interpreter, arguments and environment. Its implementation and test are unchanged by this PR. This does not explain the full-run failure or establish full-suite success

The current-main merge exposed one stale inspection test guard that still monkeypatched the retired build_turn_envelope symbol. The test now patches only the current build_loopx_turn_plan entry and still verifies that inspection reaches no live or write path. The affected journal-inspection file passes 32/32 cases; the complete final focused cohort passes 251/251.

The end-to-end harnesses are temporary acceptance probes, so only aggregate results are reported publicly. Durable public regressions are in tests/test_turn_inspection_session_host.py; provider-level acceptance is not yet a checked-in reproducer. These are source-checkout correctness checks with synthetic Host output, not installed-runtime adoption or live-model qualification.

Frontend / Visual Evidence

UI impact: none. The existing CLI diagnostics and reference text change; no App, website, documentation chrome or first viewport changes.

Type Of Change / Area

• Bug fix
• Documentation update
• Test update
• Area: Control plane; Host/runtime integration.
• Direction: Core control-plane hardening. Shared-authority RFC fixture impact: N/A; no promotion, provider routing or compatibility projection changes.

Boundary Checklist

• Diff and PR text contain no private state, credentials, raw traces, internal links or local machine paths.
• No benchmark work is duplicated.
• The change stays within the reproduced defect.
• No paired RFC changed; the bilingual gate passes.
• UI impact is none.
• The commit includes a DCO sign-off.

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · max

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

动机 / Motivation

准备恢复一个失败回合、先用CLI检查是否可继续的操作者和Agent。 原Session仍可用时，缺失或被修改的原决策捕获未阻止inspect报告可以重调Host，随后run-once却拒绝；调用者会拿到相互矛盾的恢复建议。

原Host超时后保存了可恢复的Session，但同一Turn的原决策文件丢失。操作者先inspect，再尝试原Turn续跑。 同一冻结34项测试在当前main出现20个目标负例失败、14项原行为通过；候选相关251项通过。独立真实File/SQLite CLI四格均在缺失时exit1、恢复原字节后exit0，再续原Session成功，终态删除捕获后仍只回放、不再执行Host。

验证覆盖当前源码、隔离真实File/SQLite、原CLI与合成Host超时/恢复；真实模型采用、安装版和长期运行收益未验证。

改动思路 / Approach

只在既有恢复判定为continue且reinvoke_host=true时，复用dispatch的原捕获验证器，并在原journal锁内保持同一检查快照；不publish、不新建捕获、不重跑guard。 4条路径 +144/−66；两个CLI调用、既有executor检查、参考说明及现有测试夹具。

原journal、Session和capture继续各自保有来源；原TypeScript owner决定何时需要Host，Python沿用stage验证provider文件。新增flag属于public adapter前提，没有新持久状态、权限或恢复服务。原Session失败先处理，避免capture错误覆盖更早的拒绝；保存结果与terminal继续原阶段，不为只读检查补造来源。

具体改动 / Changes

评审提交0fb8a777a1b1c6a2cb7f7dd4776c59e888df4270，当前main为5f132dd66585cd5a229a7c9f480e54b0db1cdb62，实际共有祖先8f6e88cb2a0fb76aba32bce979ba51293c1981f8。4条路径 +144/−66；两个CLI调用、既有executor检查、参考说明及现有测试夹具。 原公开consumer契约

• dispatch-parity：会重调Host的检查必须执行与dispatch一致的原捕获前提。 同一冻结34项测试在当前main出现20个目标负例失败、14项原行为通过；候选相关251项通过。独立真实File/SQLite CLI四格均在缺失时exit1、恢复原字节后exit0，再续原Session成功，终态删除捕获后仍只回放、不再执行Host。
• read-only：检查不改journal/Session/capture，也不执行或记账。 v0/v1×普通/watch×缺失、改写、symlink、malformed、目录20格拒绝且journal/Session字节不变；恢复原文件即可检查。已有保存结果和终态不需新的Host输入。
• recovery-boundary：缺失恢复原捕获，同Session续跑；缓存结果和终态回放不取得新输入。 同一冻结34项测试在当前main出现20个目标负例失败、14项原行为通过；候选相关251项通过。独立真实File/SQLite CLI四格均在缺失时exit1、恢复原字节后exit0，再续原Session成功，终态删除捕获后仍只回放、不再执行Host。
• legacy：v0原envelope和v1完整source均保留各自原来源。 v0/v1×普通/watch×缺失、改写、symlink、malformed、目录20格拒绝且journal/Session字节不变；恢复原文件即可检查。已有保存结果和终态不需新的Host输入。

关键代码讲解

• handle_turn_journal_inspection：普通CLI显式要求原捕获检查。 源码
• inspect_loopx_turn_journal：原锁内先Session/recovery再按reinvoke_host读捕获。 源码
• stage_turn_decision_capture：复用原source绑定和文件验证；stage无publish副作用。 源码
• recoveryResumePhase：host-result验证失败回Host；task-postcondition/cached结果保留原阶段。 源码

stage只读取并绑定原source；没有调用publish_for_host。validation失败是否回Host与原TypeScript recoveryResumePhase和executor的_host_attempt_requirements一致。候选增加的20格覆盖v0/v1、普通/watch、五类坏文件，独立CLI另覆盖File/SQLite和原Session续跑，避免只证明serializer或假观察。

对主干的风险 / Risks

最强风险是错误放行缺失来源，或反过来阻断不需要新输入的终态/缓存恢复。v0/v1×普通/watch×缺失、改写、symlink、malformed、目录20格拒绝且journal/Session字节不变；恢复原文件即可检查。已有保存结果和终态不需新的Host输入。 独立四格观察到缺失exit1、原文件恢复exit0；检查前后原journal与capture字节保持，再续跑使用resume，最终移除capture后replayed且Host总次数仍为2。

251项相关Python、独立真实CLI四格通过；冻结34项在main20失败14通过。首次并发premerge中semantic和CLIbudget在120秒超时，原失败保留；同一两项串行通过，随后标准串行premerge原限制内6项直接+14项选中全部通过。

作者早先全仓测试exit130，报告1failed/6017passed/841skipped/56subtests，涉及native-error-before-terminal-info0。孤立base/head各通过不能解释该整跑失败。本轮保留未知原因；当前未改chat_agent原生parser，251项相关路径和原生CLI恢复对本次不变量提供独立证据，不声称全仓绿色或历史失败已修复。没有查询远端CI、dismiss、合并或生产安装。

我的整体评价 / Verdict

当前范围内支持批准：长程恢复建议与原执行前提一致，用户避免矛盾重试。原验证器和typed recovery词汇被复用，代码规模与问题相称；锁内复用检查是本次有界整理，没有新持久owner。历史v0和v1均保留，public强制门与低层兼容默认的边界已明确，未获得Host或结算执行权限。

检查了CLI普通/watch输出与调用链；差异没有页面、首屏、样式、设置或可见控件变更，既有CLI检查就是本次用户入口。 本轮未新增浏览器截图或安装版视觉验证。

验证覆盖当前源码、隔离真实File/SQLite、原CLI与合成Host超时/恢复；真实模型采用、安装版和长期运行收益未验证。

English verdict: APPROVE at 0fb8a77. Inspection now checks the original capture before Host retry, preserves read-only behavior, and retains cached-result and terminal recovery. Historical full-suite failure attribution remains unknown.