[loopx-agent]

A batched outcome review could pass refresh-state --dry-run, then fail its due Stride review during execution because the refinement had already changed the referenced context. Preview now uses the acceptance owner’s prospective contract, matching the existing execution order and detecting the stale basis before any commit.

Goal And Delivered Outcome

• Basis: reproduced preview/execute mismatch in the existing outcome-refinement → replan journey; intended base main (started at cdbdc14ea3963e6f6aec4d574a1b88baff965132).
• Before → after: both File and SQLite returned successful preview for a batch that execution rejected with stride_review_basis_mismatch; both now reject that preview with no refinement, run or debit. Ordinary batching remains accepted.
• Owner: the typed acceptance owner still computes the contract and the existing replan owner qualifies it. Python transports the prospective facts only for dry-run planning; execution rejects that internal preview input and reads committed authority.

Author Declaration

• Written by: model_agent — OpenAI Codex (GPT-6 family).
• Implemented against: docs/architecture/rfcs/goal-outcome-refinement-v0.md, “State, reads and recovery,” at cdbdc14; existing due-review context semantics in docs/reference/protocols/goal-vision-replan-contract-v0.md at that revision.

Criterion	Disposition	Implementation / evidence
Sequential refinement and partial-effect recovery	implemented	Preserve the original receipt and operation; native tests retain one refinement and one current-Turn debit after recovery.
Exact due-review context	implemented	Preview the post-refinement facts; stale decisions remain refused. Read fresh facts before authoring a new decision.
Retained intent and ordinary continuation	implemented	Original requirement/source survive, feature-off batching remains accepted, and no Todo completion or stop rule is added.

Scope And Continuation

Complete author implementation for this bounded preview repair. The project and self-repair guides now order changed milestone facts before a fresh same-Turn due-review context, and distinguish an applied refinement receipt from a failed ordinary writeback. The English/Chinese RFC checkpoint is updated together.

This retains sequential composition; it does not make the two effects atomic or automatically rebind an old decision. Installed adoption, model efficiency and long-horizon outcome benefit remain unqualified. Revert this commit to roll back the source change; it introduces no migration, configuration or persisted-state format.

Bounded refactoring: reuse the current owner-produced plan and planning call; an extra decision helper, cache or transaction framework is unnecessary. No broader refactor is included.

Validation

• Tested revision: ed9153989aec852d46e345b02f7462c3faeafb3c (same product/test blobs as the working-tree runs below).
• Run state: finished for the listed local tests; the standard premerge run and its environment recovery are complete; qualification remains held below.
• Inputs: synthetic, disposable canonical File and SQLite authorities. No model, solver, benchmark or evaluator job was launched.

Check kind	Result	Evidence / limitation
real_entrypoint / real_backend	passed	tests/control_plane/test_outcome_replan_composition_cli.py: 4 pass; preview isolation, exact-context refusal, fresh-context recovery, partial receipt, retained requirement/source, ordinary batching and single-debit replay. Repeated on the committed head.
regression_parity	passed	The same two due-Stride test oracles fail on immutable cdbdc14 because preview incorrectly succeeds; both pass with this change. The baseline was a separate checkout using its own imported source.
integration	failed	Outcome refinement/diagnostics, effective-Turn cadence and long-chain closeout modules: 70 pass, 3 fail. All three unchanged test_managed_turn_counts_advancement_without_monitor_backlog cases fail at the initial missing-canonical-authority fixture read on the immutable baseline as well. These remain failures.
unit	passed	tests/test_repair_pattern_lookup.py: 15 pass; exact error-code lookup finds the new ordering/recovery pattern.
static	passed	Changed Python compile, diff checks, staged/exact-head bilingual RFC checks and maintainability ratchet. Diff-scoped semantic advisory finds no supported new vocabulary carriers; it does not cover dynamic forms.
static	failed	Existing interaction-pattern catalog smoke omits the built-in context_delivery namespace in IP-030; identical assertion on baseline and candidate. Unrelated to the changed namespace/registry owners.
static	passed	All 7 changed public files pass the isolated publication scan. An initial ambient-registry check separately found two unrelated existing Goals without canonical authority; it did not find a public-file leak.
manual	blocked	Exact-scope change-quality remains non-passing because failed broader validation is retained. No optional-baseline exemption or merge-readiness claim is made.
static / integration	qualification held	Standard catalog-selected premerge: all 6 direct checks and 16 of 17 selected checks passed. The remaining semantic-vocabulary check initially lacked local TypeScript dependencies; after npm ci --ignore-scripts, that exact check passed. The recorded change-quality failure still holds qualification; the original premerge failure is retained.

The required broader failures and pending independent review/CI hold merge qualification. They do not make this completed author proposal a Draft. No checks or gates were relaxed, and no CI workflow was manually dispatched.

Frontend / Visual Evidence

UI impact: none. This is the existing CLI preview composition; the Goal outcome editor continues to use its existing native acceptance API. No new capability, setting, screen or visual state is introduced; screenshots are not applicable. Packaged/installed UI and PostgreSQL were not newly qualified by this source repair.

Type And Area

• Bug fix
• Documentation update
• Test update
• Control plane
• Capability / skill guidance

Direction: existing S2/S3/S7 recovery and evidence-driven replan boundaries. This is not a new shared-authority migration or benchmark treatment.

The PR is offered for maintainer review; no self-merge, installation or release is requested or performed.

[loopx-agent]

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

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

当前精确头 ed9153989aec852d46e345b02f7462c3faeafb3c 的代码结论为 APPROVE，未发现阻塞性代码缺陷。严格 CQR 与 premerge 仍为失败，不代表可以合并。

动机

已把成果审阅和本轮结果一起提交的 CLI Agent。

旧版预览提示可以提交，执行却在成果已保存后拒绝旧审查决定；修复后预览提前拒绝，Agent 按原轮次读取新事实再完成一次结算。

真实 File 和 SQLite 已验证无副作用预览、保留部分提交、原身份恢复和一次扣额。

不引入通用事务、自动改绑旧决定、安装升级或模型效率承诺。

安装宿主采用、真实模型决策与完整长程成本仍由既有采用工作分别验收。

改动思路

预期成果由既有 TypeScript 验收 owner 计算，Python 仅把它交给原规划器；实际执行仍读取已提交事实。

当前切片完成原 CLI 预览与执行的事实顺序对齐，顺序提交和原回执恢复的责任保持原归属。

“先把两步做成通用事务”成本更大，也改变原回执责任；仅改文档又留下错误的预览成功。本补丁复用已经存在的预期成果投影，只在 dry-run 的内存规划中使用。启用 Stride 的到期审查仍必须绑定它实际读过的事实；失败后保留原操作，重新判断，不能只给旧决定换一个上下文编号。

具体改动

规格：docs/architecture/rfcs/goal-outcome-refinement-v0.md#state-reads-and-recovery，spec_revision cdbdc14ea3963e6f6aec4d574a1b88baff965132，使用修改前已接受的文字。条款映射：sequential-composition 已实现，执行仍先保存成果，随后写回失败时返回原回执；retained-intent 已实现，原要求、来源和不确定审阅完整保留；ordinary-continuation 已实现，关闭 Stride 的两类后端均沿普通批量路径结算，没有增加用户确认。

关键代码讲解

• handle_refresh_state_command 保存已有 owner 的返回值，仅在 dry-run 把完整预期契约交给规划器；紧随其后的普通结果仍保留精简成果回执，异常不会伪装成全部回滚。
• refresh_state_run 在执行态拒绝该内部参数；预览时只替换规范规划快照中的成果投影，原数据、归属与提交来源不改。
• 既有 TypeScript refineGoalAcceptanceOutcome 仍决定细化、保留和重放；HTTP、原生会话工具与 Goal 成果编辑 API 继续用它。本次七路径 +227/-2，其中生产两文件 +22/-2、测试 +166、文档与技能 +39；没有另一套决定 owner、状态或配置。

对主干的风险

最强风险是把预览事实当作执行权，或者失败后换操作编号重复写入。真实 File/SQLite 的四条场景通过：预览不改契约、索引和扣额；实际拒绝保留 applied 回执；同一操作重放返回 replayed；读取同 Turn 新事实后完成一次写回和扣额，再重放不追加。相同观察器在不可变基线上保留两项旧缺陷失败和两项普通通过，当前四项通过，没有把基线失败改名为成功。

较广的五模块及修复目录测试在 base/head 均为 81 过、3 败。对三项失败另取完整真实 CLI 响应，均在第一次 quota 读取时因 fixture 缺规范 authority 拒绝，发生在本次细化/refresh 路径之前。仅归一合成夹具根目录和经校验的 rollout 事件编号、时钟后，完整错误、拒绝、无写入和恢复字段相同。根 Ruff 的三项 F811 输出与对应不可变基线逐字相同，因果文件未变。它们可独立归因为已有问题，但仍是失败。

语义与 CI 对齐

changed Ruff、配置指定的 17 文件 mypy、TS typecheck、双语 RFC、diff 与语义 advisory 通过；标准 premerge 的 6 项直接检查和 17 项选择检查通过。严格 CQR cqr_85b11c75489c180143d1 已物理读回，保留 required 验证失败，故总门为 quality_invalid_receipt。未设可选基线豁免，未取、轮询或等待托管 CI。当前复用既有词汇；新增私有规划参数不增加共享协议，错误与义务保持通用语义，文档没有把机器拒绝说成可选建议。

我的整体评价

这是可独立验收的 CLI 恢复增量：long_horizon 与 user_experience 在该边界得到改善，原轮次、承诺和一次效果保留，也避免把版本/上下文辨认责任交给人。机制与问题相称，继续保留既有 receipt 与顺序恢复即可，不需要新抽象或人为后续任务。真实模型是否遵守指引、安装宿主是否采用，以及完整费用和长程收益均未资格化。

已检查 CLI、原生成果工具和 Goal 成果编辑 API；本差异不改变渲染页面或布局，视觉评审不适用。 视觉未验证范围：none。

APPROVE 仅是当前代码判断；现有 required 红灯、CQR/premerge hold 与维护者合并权限仍须分别处理。

English verdict: APPROVE — ed91539. Real File/SQLite CLI cases verify effect-free prospective preview, exact-context refusal, original-operation recovery and once-only settlement; the immutable base retains two stale-preview defect failures. Three broader fixture failures and three Ruff failures match the actual bases. Required failures remain failed: strict CQR and premerge hold merge qualification. No installed/model or total-cost benefit is claimed.