Repository navigation
Resume the saved first Goal Turn after its checker recovers - #9124
loopx-agent wants to merge 1 commit into
Conversation
Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>
loopx-agent
left a comment
There was a problem hiding this comment.
Reviewer: model_agent · gpt-6.1-sol · OpenAI (runtime_reported, low)
Approval conclusion (author-owned PR; GitHub blocks formal self-approval)
动机 / Motivation
仅安装Core、第一次执行带独立checker任务的CLI用户。 worker已完成但checker临时失败时,原命令只重放失败;用户现在能看到原Turn key与显式retry提示,恢复checker后接着处理已保存结果。 独立Core wheel、File/SQLite:checker缺失后恢复与checker实际判错后恢复,均复用原Turn/Todo,worker仅执行一次,成功只扣一次,重复返回原receipt。 只补齐已有首次CLI任务的失败恢复;不新增App能力或宣称完整split-package、真实模型或发布资格。 macOS Python3.14独立Core wheel与真实File/SQLite/TS;使用无付费模型的本地subprocess adapter和独立checker。未验证Windows、真实Codex会话、安装App、完整M3/R1或全成本收益;未查询CI。
改动思路 / Approach
start-goal adapter用原launch digest定位同一creation operation和Turn;既有TypeScript journal recovery决定continue/return_existing及reinvoke_host,adapter仅增加拒绝重跑的窄约束,再交既有turn handler完成原事务。没有第二份journal、claim或settlement owner。四文件:解析flag、原首次任务adapter恢复分支与可见诊断、Core文档和安装测试;无新状态schema或通用调度策略。状态仍由原typed owner决定。新flag仅表达用户明确重试已有检查,不把available Host或checker当权限;无需新前端编辑器或安装Workspace才能使用这个已接受的headless Core任务入口。I2/I8和V2/V3在此局部组合满足;V4仅证明本地可恢复实例,不声称全部Host资格。 原恢复入口要求。
具体改动 / Changes
精确 head 56366ae6a7e94ee3cdb36d33f21048d470cb58a7,diff base 770fe0473f955b537aac493e9cb755d5c4e2660f,4 个文件 +154/-29。逐项复核 I2、I8、V2、V3、V4。验收约定
关键代码讲解
- 新flag需要--execute,省略时保留旧分支。源码
- digest覆盖原任务和checker输入;retry只读原journal,不调用新增Todo或claim路径。源码
- 必须是continue/return_existing且reinvoke_host为False,随后既有turn handler重读原plan。源码
- Markdown显示原Turn key和修复后的操作,Core文档写明不重跑worker与重复receipt。源码
- 安装测试覆盖缺checker、改变命令拒绝、canonical Todo完成、worker/checker次数和一次spend。源码
对主干的风险 / Risks
独立Core head:相邻安装6 PASS、独立恢复/拒绝4 PASS、默认启动恢复15 PASS;base相同独立4 FAIL(新flag不识别),默认15 PASS。标准premerge最终PASS:四项direct、编译、8/9 catalog通过(1项继承维护性失败advisory)、8项risk smoke与精确四文件公开扫描通过。 维护性独立诊断FAIL:build_heartbeat_prompt为64 decisions/83 statements,_admitted_interaction_next_cli_actions为58/96;两文件均与base字节相同,native premerge将其归为advisory_inherited_failure。作者曾报告全仓扫描旧fixture问题;本次只证明精确四文件公开扫描通过,未声称全仓清零。 macOS Python3.14独立Core wheel与真实File/SQLite/TS;使用无付费模型的本地subprocess adapter和独立checker。未验证Windows、真实Codex会话、安装App、完整M3/R1或全成本收益;未查询CI。
语义与 CI 对齐
--retry-failed-turn是显式CLI选项;省略时保持原失败重放语义。原project/Goal/Agent/launch/checker identity、journal与一次settlement不变。 默认off的base/head各15项通过,独立saved/unsaved四项对照覆盖重跑约束。标准premerge已完成并通过,保留维护性原始失败advisory;未查询或等待CI。
我的整体评价 / Verdict
同意此边界明确的Core恢复增量。long_horizon与user_experience均改善:已确认工作可继续到可读回结果,避免重复worker与扣账;没有新增协议owner,复用现有typed vocabulary。标准premerge通过且保留一项继承维护性advisory;完整App、真实模型和其他平台未验,不授予合并。
English verdict: APPROVE. At 56366ae, Core-only retry recovers the original saved result without another worker dispatch or debit. Standard premerge passes with an inherited maintainability advisory; other platforms remain unverified.
Follow-up to #9053 and its saved-result recovery review. When a worker has saved its first result but the independent checker cannot complete,
start-goal --executenow reports the original Turn key and explains how to resume. Repeat the original command with--retry-failed-turn; the CLI uses the canonical Turn recovery decision and only proceeds if it can keep the original worker result. It does not add or claim another Todo. A plain rerun still returns the previous failure. A changed checker command or missing original Turn is rejected.Validation: seven tests against a separately installed Core wheel passed, including File and SQLite recovery after an unavailable checker. They verify the same Turn and Todo, exactly one worker invocation and one quota spend, and effect-free replay after settlement. Ruff, mypy, the CLI output budget smoke, and the public scan of all four changed files passed. The repository-wide public scan reports existing findings in an unchanged test fixture; full suite and premerge are not reported green.