fix(turn): preserve effect uncertainty and original journal recovery on errors

[huangruiteng]

Goal And Delivered Outcome

• Outcome basis / optional anchor: Overall roadmap R2 recovery; TypeScript control-plane single-owner boundary.
• Goal/source and gap: an executing turn run-once can raise after Host execution or canonical writeback. The CLI exception renderer replaced every effect with false, even when the original journal already contained Host and validation checkpoints. A caller could mistake a failed reply for non-execution and start another task.
• Observable before → after: failures retain the original error and resume identity; uncertain current invocation effects are null. A separate allowlisted journal_observation projects the original Turn checkpoints and recovery decision through the existing TS journal owner. prepared and malformed attempts remain unknown, never invented commit receipts.
• Issue/task and intended base: user-requested, reproduced reliability repair; main. Related to the existing recovery protocol in fix(turn): recover ambiguous settlement effects without duplicate execution #3412. It does not reimplement provider recovery or loosen permission rules; the inspector now shares the original writer's prepared-intent validation.
• 中文：修复执行后异常却声称“没有执行”的误导。区分本次调用的不确定效果与原 Turn 的持久观察，保留原恢复身份，避免重复模型调用；不放松验收、租约或扣额门禁。

Scope And Continuation

• Complete within this scope: original-journal observation on unexpected execution errors; JSON and Markdown error/read-only-inspection entry points; strict Python transport validation; TS effect classification; shared original writer/inspector intent validation; protocol and roadmap disclosure.
• The TS owner remains authoritative. Python only performs existing read-only registry/journal I/O, forwards the projection, and renders invocation uncertainty; no second policy, store, scheduler, receipt, or provider call is added.
• This is partial R2 / heterogeneous delegation, not complete runtime discovery, provisioning, caller admission, fresh-task derivation, or product qualification. Existing follow-up work owns original configuration readback, model/effort/runtime controls, frontend and Lark delivery, and repository-aware lease conflict diagnosis. Structural validation is not semantic acceptance or financial utility.
• 中文：此 PR 只完成异常读回切片；异构委派的准入、任务绑定与前端/Lark 产品闭环仍开放，结构验证通过不等于父 Agent 语义接纳。

Validation

• Tested revision: 49ac1fcccc5f372e70b99347bd25a7dc7dbf04c5
• Run state: finished
• Input classes: synthetic, public_fixture, authorized_private_read_only

Check kind	Result	Public-safe evidence / limitation
unit	passed	55 TS journal/effect/settlement tests: checkpoints, shared writer/inspector invalid-intent matrix, valid historical/prepared/terminal-closeout cases, foreign identity and acknowledgement uncertainty.
integration	passed	190 final-head Python tests across Turn CLI, journal inspection, execution, lane fencing and disposition.
real_backend	passed	Existing promoted File-provider crash fixture: canonical Todo commits before the journal checkpoint, error observation reports unknown writeback, and resume retains one commit/Host invocation. No SQLite/PostgreSQL effect code changes.
real_entrypoint	passed	CLI JSON and Markdown inspection/error roundtrip against the backend; The actual unknown-intent inspection checks journal byte equality and forbids all Host/status/writeback/spend paths.
regression_parity	passed	The identical File crash fixture and final-head test blob were executed against immutable base fd96e5e2574272262b9ea604a96581a0d20e94d1 and the final head. Base fails specifically at False is None after the real canonical commit; head passes and verifies same-Turn recovery without another Host invocation. Only the independent test oracle was copied into the isolated base; its production source was unchanged.
regression_parity	passed	The same real CLI oracle blob 24b5bcfeb1b507e2dd088b3bc6dc0be75340a9c0 fails on prior PR head 39684b7659a05d59b131d4e87a5cdf45e656faf3 at True is False for journal consistency, then passes on current head. Unknown prepared intent now gives blocked/no-reinvoke/unknown, matching the original writer denial. Only the oracle was copied into the isolated old-head checkout; no production changes or journal writes.
manual	passed	Authorized read-only inspection of an already committed managed Turn agrees with its checkpointed effects and last recovery's no-reinvoke fact. No model call, settlement, permission change, or new financial validation was performed. Evidence stays private.
static	passed	TypeScript typecheck, targeted Ruff, DCO/author+committer identity and git diff --check; no private runtime artifacts in the branch.
static	passed	Standard diff-selected loopx canary premerge --from-git-diff --git-diff-base fd96e5e2574272262b9ea604a96581a0d20e94d1 --tier standard: all 19 checks passed, no failures/skips/warnings/manual holds; validation is not merge authority.

• Coverage and gaps: unchanged success/replay invocation effects and existing settlement gates remain covered. Full repository tests and live external models are not rerun; the read-only observation cannot establish the missing provider commit receipt or semantic acceptance. Frontend/Lark product journeys remain follow-up work.

See validation disclosure guidance.

Frontend / Visual Evidence

• UI impact: none
• Before: N/A — no dashboard/web/desktop layout or settings changed.
• After: N/A — this bounded repair changes CLI JSON and Markdown feedback only.
• States and viewports shown: CLI execution-error and read-only inspection, including unavailable/unknown observations; no packaged browser/Lark qualification claimed.
• Source data: synthetic
• Attention review: adds the original recovery/effect facts next to the actual error instead of presenting false certainty. The broader configuration/editor/Lark companions remain open; API-only completion is not product completion.

Type of Change

• Bug fix
• New feature
• Documentation update
• Test update

LoopX Area

• Control plane (goals, todos, quota, scheduler, registry, runtime)
• Host or runtime integration
• Public docs or presentation surface (protocols)

Technical Direction

• Core control-plane hardening; roadmap R2 bounded recovery, following the existing TypeScript journal owner. No provider promotion or enablement default change.

Shared-authority RFC fixture impact

• Production-scale fixture schema: no authority-store fixture, Todo, lease or settlement schema change. Inspection adds allowlisted recorded_effects; exceptional invocation effect booleans can now be null.
• Semantic dimensions changed: observation scope/uncertainty, plus stricter inspection of invalid prepared intent. Inspector and writer now share the original pure TS predicate; intents already denied by the writer block inspection recovery instead of suggesting continuation. Original writer admission, identity/fence, persisted journal schema, valid replay, provider receipts and side effects are unchanged.
• Provider conformance arms run: real isolated promoted File crash/replay fixture and synthetic transport/TS projections. No new store operations.
• Read-only legacy/file/PostgreSQL three-arm rehearsal: not applicable; no authority provider promotion, runtime routing or legacy compatibility projection change. Existing read-only journal owner and original provider recovery paths are reused.

Boundary Checklist

• Neither diff nor public artifacts disclose private state, credentials, raw traces, prompts, internal links or local machine paths.
• No duplicated benchmark or provider recovery owner.
• Scope is the reproduced exception-readback defect; parent delivery remains open.
• UI impact marked none with affected CLI journey and unqualified companions disclosed.
• Every commit has a verified GitHub noreply author/committer and DCO sign-off.

[huangruiteng]

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

Exact head: 49ac1fcccc5f372e70b99347bd25a7dc7dbf04c5; immutable base: fd96e5e2574272262b9ea604a96581a0d20e94d1。
这是作者自评审，不是独立 approval，也不单独授权合并。按当前 capability policy revision 11 完成全 PR 复核，不继承旧 head 的结论。Goal 配置不等待 CI，本次未查询或轮询远端 CI；用户指定的独立审阅仍是合并前置条件。

动机

真实执行可能已调用 Host、验证产物，甚至完成 canonical Todo 提交，随后在本地 checkpoint 失败。基线 CLI 异常 renderer 却把所有 effects 写成 false，把“没有收到成功答复”误说成“没有执行”。纠正这种反馈可以减少恢复时重复的模型调用和业务效果，同时保留原 Turn 的恢复身份，不取消验收门禁。

上一轮独立审阅还发现更具体的反例：原身份和 phase prefix 合法，但 journal 含未知 prepared step。旧 PR head 的 inspector 给出 consistent/continue/reinvoke_host=true，原 writer 却拒绝同一个 intent。因此本次不只增加 null，也要把 inspector 对齐现有 writer 契约。这里实证的是错误恢复建议，并未证明旧 writer 真的执行了重复效果。当前交付是独立可验证的恢复切片；异构发现、准入、fresh task 绑定、产品回传和父 Agent 的金融语义验收都没有按此 PR 宣布完成。

改动思路

只删除 false 更小，但会隐藏已有 checkpoint 和恢复入口。新造 provider 查询、Python 判定或另一恢复状态机会复制权威。这里复用已有 executor 的锁内只读 inspection、原 TS recovery owner 与 provider prepared-effect readback；新增 recorded_effects 每次从原 journal 派生，不维护新的持久状态。

独立审阅修复采用更小的 consolidation：把 writer 原有 prepared-attempt 校验提成同一个 TS 纯函数，writer 和 inspector 共用；不创造另一份 inspector 规则。仍使用原 settlement step、phase、status 与 effect identity。absent attempt 字段仍合法；present empty map 仍保留原 writer 的拒绝语义。unsupported step、多个 intent、错误引用和错误 phase/status 不给出可继续建议。原 writer 准入、身份、租约、quota、completion validation 和 valid replay 不放松；只有 inspector 对原 writer 已拒绝的 intent 更严格。Python 保持 I/O、严格 transport 与展示的角色。

具体改动

关键代码讲解

1. handle_turn_command（loopx/cli_commands/turn.py:93）区分是否已进入 executor；异常时用原 goal/agent/Turn key 只读观察一次。观察失败仍保留原错误和未知，不创建任务、不重新调用 Host。
2. preparedAttemptViolation（loopx/control_plane/turn_driver/turn_journal_attempt_contract.ts:18）复用原 writer 的单个 prepared step、精确 effect reference、next phase 和 status 契约，返回 typed diagnostic。原 private checker 被替换为共享调用，而不是复制新规则。valid terminal-closeout 的特殊 phase 也保持原语义。
3. interpretTurnJournalEffect（loopx/control_plane/turn_driver/turn_journal.ts:548）分别判断 lineage/phase 和 prepared intent。intent 错误追加原式诊断并令 journal_consistent=false，现有 recoveryDecision 据此 blocked、reinvoke_host=false；inspection 仍不产生 effects。
4. recordedTurnEffects（loopx/control_plane/turn_driver/turn_journal_effect_readback.ts:16）只投影原 Turn 的 lower-bound facts。可信 lineage 中已 checkpoint 的 true 可以保留；非法 intent 下未证明的效果均为 null，包括此前保存的 scheduler acknowledged=false。foreign/corrupt lineage 不借用事实。prepared 不能当作 committed。
5. build_turn_error_payload（loopx/cli_commands/turn_rendering.py:16）保留已知 hook effect，当前调用未确认的执行效果为 null，allowlisted 原 journal 放在 scope=original_turn。严格同包 Python adapter 与 JSON/Markdown 使用同一投影，不把原 Host 当成本次又启动。

整条 base-to-head diff 为 14 文件 +468/-53：生产 7 文件 +193/-50、测试 5 文件 +233/-3、双语协议/roadmap 2 文件 +42/-0。最后一次修复为 8 文件 +216/-60；新增共享纯校验 57 行，删除原 private 检查，派生 observation helper 55 行。不是语言迁移或配置框架扩张。测试涵盖真实 File 提交后 checkpoint 崩溃、shared writer/inspector 九种非法 intent、合法历史/准备/终结情况、不可用/异源观察、隐私 allowlist 和 CLI 渲染。

对主干的风险

最大风险是把原 Turn 的效果当成本次重新执行，或把 prepared 当成提交证明；另一风险是 observer 和 writer 再次漂移。明确 scope、nullable lower bounds、共享原谓词及真实 CLI 反例共同约束这些风险。原事实不授予新交易或新的 Host 调用，结构验证也不证明金融内容正确。只读异常路径增加一次锁/RPC；损坏或不可用观察仍可能缺失诊断，此时保留原错误和未知，不自动 fallback 执行。

语义与 CI 对齐

行为变化明确披露：基线异常 all-false 改成 nullable；旧 PR head 未知 prepared intent 的 continue 改为 blocked。后者对齐已存在 writer 拒绝，不改变其执行准入。valid success/replay 的当前调用 effects、原持久 journal/receipt schema 和默认启用配置不变。依赖异常 false 的外部脚本需要遵循双语协议识别未知，不做版本双轨或第二权限来源。

最终 head 验证：190 项 Python 回归（97.68s）、55 项 TS、typecheck:control-plane、Ruff、diff/身份/DCO 和 Goal-bound standard premerge 19/19 全通过，无 skip、warning、manual hold 或预算延长。既有真实 File oracle 在 immutable base 的 canonical 提交后准确失败于 False is None，修复路径保持一次 Host/提交。新增相同 CLI oracle blob 24b5bcfeb1b507e2dd088b3bc6dc0be75340a9c0 在旧 head 39684b7659a05d59b131d4e87a5cdf45e656faf3 失败 True is False（7.11s），新 head 通过（1.28s）。它检查 journal 字节不变，并禁止 Host、状态/业务写回与 spend 路径；不是用 mock 成功回执证明提交。

本 PR 的产品入口是现有 CLI 错误/inspection JSON 与 Markdown，已实际运行。没有 dashboard 布局、设置或配置字段变化，不需要新增 UI source；frontend/Lark 全体委派闭环仍由既有 companion 负责。本次未运行 packaged browser/Lark、外部模型或全仓测试，不以 API 或测试数量声称整体产品完成。

我的整体评价

APPROVE，限于该 exact head 的异常读回和原 journal 契约一致性修复。全 PR 判断是有价值、边界合理的 justified increment：持续工作中的恢复更可信，用户能看见真实错误及原身份；shared predicate 减少重复规则，TS 保持决策权威。独立审阅反例已修复并实际 old/new 证伪，不继承旧结论。通用 discovery/admission/task binding、跨仓库 lease namespace 与 frontend/Lark 读回仍未闭环，父 Agent 的证据语义拒绝也不能因此撤销。作者自检不替代用户指定任务的最新 head 独立批准，未批准不合并、不升级。

English verdict: APPROVE - 49ac1fc; truthful nullable error effects and original recovery now share the writer's TS prepared-attempt contract. The old-head public CLI counterexample fails and the new head passes without writes or reinvocation; 190 Python, 55 TS and 19 premerge checks pass. Author self-review is not independent approval; broader delegation/product qualification remains open.