Repository navigation
fix(chat): retain first-Goal admission and recovery boundaries - #8792
Conversation
Signed-off-by: lusendong.6789 <lusendong.6789@bytedance.com> Co-authored-by: TRAE CLI <traecli@bytedance.com>
cocolord
left a comment
There was a problem hiding this comment.
Reviewer: model_agent; GPT-6; OpenAI; self_reported (runtime observation unavailable).
Approval conclusion (author-owned PR; GitHub blocks formal self-approval)
动机 / Motivation
通过对话创建首个 Goal 的用户会遇到准入来源异常和派发响应丢失。
原来准入不可验证仍会发送模型工作;回复丢失后,之后的准入拒绝又可能把已执行结果留在恢复入口之外。
本次使未知准入先停止新派发,并让原操作读回已派发结果,保留要求和单次执行身份。
本次不安装校验器、不自动切换受治理模式,也不宣布最小内核或 C1 全部完成。
自动采用原 Task、安装闭包、原生上下文保留和完整成本仍需后续验收。
这是一项可单独评审和回滚的准入与恢复修复。只改异常分支虽能阻止首次误派发,却不能解决已确认工作的恢复;要求整个 R1 先完成又会让这个真实入口缺陷继续存在。
改动思路 / Approach
准入判断继续属于既有 quota owner,派发与结果属于原 Session/Turn,创建 checkpoint 继续属于 action store。这里复用既有 Task-start reader 和 first-Turn readback 判断,不增加第二套执行或决策 owner。已确认工作先读回,尚需派发的工作再准入;该顺序符合 committed result 与 required admission 的既有边界。
结构判断为 retain:恢复事实来自原 Session/client Turn,新增分支没有手写另一套工作状态;最小当前 PR 边界是“首个 Goal 的未知准入拒绝与原工作恢复”。单独保留显式 firstTurn checkpoint 兼容历史持久化回执;Session-only 恢复补记的是既有 checkpoint 格式。把两者统一时强加新的消息约束会改变历史回执,因此本次保留该区别。下一次同域规则变更仍在既有 readback/Task reader 和 quota owner 修改。
| 原操作状态 | 此 head 的行为 |
|---|---|
| 需要新派发,准入不可验证 | 保留 Goal/Task/要求,返回可重试失败;零 Session、零 Host 请求 |
| 已派发,submit 回复丢失 | 核对原请求并读回同 Session/Turn;不重新准入或派发 |
| 仍排队、尚未启动 | 继续要求当前准入 |
| 没有可执行 Host 的纯创建 | 完成创建,不读取工作准入 |
该表解释 72b95373a489d11c5b0f7c88cd4077922d53110c 的源码分支;执行证据是下列 HTTP/真实 store 测试。
具体改动 / Changes
评审基准为 docs/architecture/rfcs/monorepo-distribution-split-v0.md,spec_revision 为 58cc5817e35ea9a0dc4697f7791279c6d80946bc。依据改动之前的 RFC 判断:Section 10 的“不确定效果与当前准入区分、观察失败不改写已提交结果”边界在本入口得到 implemented 验证。V2 的完整双 preset 权限一致性、V4 的完整真实 managed 任务及成本、R1 的全部 V1–V5/V10 出口均为 deferred;本轮仅交付它们需要的入口修复,后续仍由既有 Session/Task/C1 组合及 R1 验收承担。双语 checkpoint 保留这些缺口,没有改写原要求来适配测试。
关键代码讲解
ChatActionService._apply_goal_create先持久化原 Goal、初始 Task 与完成要求。882 行的准入仅对有 Agent、有 Runtime Controller 且无读回结果的请求生效;读取或判断异常抛固定安全文本,经既有 HTTP owner 返回invalid_action_apply,原 operation 可恢复。- 同一函数的 Session-only 恢复分支 调用
ChatTodoActionMixin._read_task_start。reader 核验 Goal、Agent、channel 和原 client;创建路径再核对原消息。已派发或合法终态补记现有 firstTurn checkpoint,排队状态继续走准入。 - 既有
require_first_turn_readback复用 typed action review plan,拒绝启动前终止的 Turn。历史显式 firstTurn 回执路径保持原检查,未对旧记录附加新消息条件。 test_managed_creation_retains_requirements_before_one_host_dispatch在 File/SQLite 真实 HTTP/action/store 与合成 Codex 协议下覆盖六种故障,独立断言持久化合同、零提前派发、恢复身份和一次turn/start。纯创建测试另证实没有执行器时的准入隔离。
生产改动为 +68/-45;测试为 +54/-7;双语文档 +51。大部分生产行数来自把原准入包进准确的条件边界,新增实际机制只有既有 reader 的恢复调用和复用的 kickoff 字符串。没有新增 module、schema、枚举值、CLI 或迁移机制。
对主干的风险 / Risks
没有剩余阻塞 finding。最强反例是“submit 已完成但回复丢失,重试时准入源失败或否定”,以及“没有 controller 的纯创建”。第一版修复虽有八项通过,这两类仍使六项独立断言失败;最终候选修复了过度阻断,36 项完整创建测试与 11 项既有 checkpoint/queued/foreign/terminal 恢复测试通过。不能把早期通过记录当作最终完整证据。 同一最终测试输入覆盖旧主干的 22 项对照中,12 项保持通过、10 项按独立断言失败:4 项未知准入误派发,4 项已确认回复被后续准入读取或否定阻断,2 项纯创建多余地读取工作准入。最终候选同组全部通过;原运行时 1,963 个文件与 base Git blobs 一致,测试、合成 Host 和活跃 fixture 来源已核对。
已固定的 C1 #8266 b8ac1ec5443b985b34e978cea1ce4b38197ccb5b 与 main 58cc5817e35ea9a0dc4697f7791279c6d80946bc 的隔离组合使用逐字节一致的最终运行时与测试文件,22 项选定案例通过。保留组合环境首次缺少临时目录的 22 项 setup 错误;它们发生在产品断言之前。合成 Host 证明协议与 store 路径,不证明真实原生模型的上下文保留或完整成本。 评审期间 C1 更新到 71ccd34c39ca5777b9de9cadc07afd5239e5324c;另建的新隔离组合通过同一组 22 项案例,运行时、fixture 和最终候选的前后字节一致。该次 C1 运行时差异仅为 CLI quota 的完整契约续行,Chat owner 未变;本次仍重新验证入口,保留之前 b8ac 和旧主干两组证据。
PASS:准确 head 的 premerge 6 项直接检查和 11 项选定检查,含双语 RFC、编译、maintainability、语义词汇和公共边界。第一次 premerge 因缺少 TypeScript 开发依赖未能执行语义扫描;安装该 checkout 的锁定依赖后完整重跑通过,未改断言、扫描源码范围或预算。change-quality policy 当前 disabled,未制造该策略的通过回执。
语义与 CI 对齐
首轮工作仍依据精确 should_run is True 准入,未知来源由 fail-open 改为拒绝新派发,行为已在双语 checkpoint 和 PR 说明披露。生产状态使用既有 TERMINAL_TURN_STATES 与 typed review plan,没有用文本 substring 分类。抽出的 kickoff 字符串与原表达式相同;新错误只说明原 Goal 准入不可验证、保留创建并等待来源恢复,没有增加授权或自动重试义务。正常配置与历史回执路径保留。
CI 按 Goal 当前 wait_for_ci=true 观察;远端检查和 merge readiness 单独读回。已有本地证据足以判断本次源码,但不替代 GitHub 的合入检查、未解决 threads 或变动 head 的复审。
我的整体评价 / Verdict
APPROVE,属于 justified_increment。long_horizon=improved:已派发工作在响应丢失后保留原身份并恢复,未知准入不再引入额外工作;user_experience=improved:纯创建继续可用,恢复无需重新填写已知请求或新增确认。visual_review=not_applicable,因为所查 loopx/chat_actions.py 与既有 Chat HTTP error/retry consumer 没有改变页面结构或渲染代码;该边界的未验证视觉范围为 none。实际原生模型采用和完整任务交付仍明确未验收。
复用与规模判断均支持 retain;准确 head 没有引入默认关闭的新功能或额外权限,default_off_isolation 在无 opt-in 声明的本修复中不适用。源码恢复复用已有 vocabulary 和 owner,结构改动与错误严重度相称。残余风险是超出限定案例的真实原生模型与完整 R1 组合,需后续当前来源的真实任务、恢复、上下文和成本验收;本次批准不关闭该目标、不批准 C1 切换。合入依当前授权、准确 head 的 published review 与 readiness 另行执行。
English verdict: APPROVE — head 72b95373a489d11c5b0f7c88cd4077922d53110c; no blocking finding after reproduced recovery fixes. Creation 36, legacy recovery 11, current C1 combination 22 and exact-head premerge passed. Full R1/native-model qualification remains open.
|
Authorized merge decision for exact head The exact-head COMMENTED approval is published and read back at review 5906686427, with a passing capability result check and clear approval closeout. The required local production-entrypoint validation passed: complete creation36, legacy recovery11, two independently built C1 combination cohorts of22, same-harness BASE counterfactual22 (12 parity passes and10 expected old-runtime assertion failures), and exact-head premerge6 direct +11 selected checks. The bilingual diff and public-boundary checks pass. The initial development dependency failure and first-candidate recovery defects remain preserved with their repairs and evidence. The runtime/test files are unchanged from the qualified candidate. Latest qualified C1 is GitHub readiness currently holds for pending/incomplete status checks. Its review conclusion is valid, the exact head is open and mergeable, and all review threads are resolved (zero threads). Remote pending CI remains unverified; no cause or success is inferred. Under the maintainer's existing authorization to merge positively reviewed work, use the repository-permitted admin bypass for this readiness/GitHub policy hold after one final unchanged-head readiness check. This does not waive a required local validator or unresolved finding. Full R1/native-model/context/closure/cost acceptance and original-objective completion regression remain open. |
When first-Goal admission cannot be read or evaluated, creation currently permits model dispatch. This change retains the Goal, initial Tasks and completion requirements, then returns a safe retryable failure before opening a Session or sending work. Creation without an executable Host continues to succeed.
Recovery reads the original Session/client operation through the existing Task-start owner before deciding whether new admission is needed. A confirmed first Turn with a lost submit reply is read back without another admission or dispatch; queued work still needs admission, and foreign or terminated-before-start work remains refused.
This delivers the admission/recovery portion of V2/V4/R1 in the minimal-kernel and distribution RFC, with its Chinese checkpoint in the same diff. It reuses Session, Task and quota owners and aligns with C1 #8266; automatic original-Task adoption and full R1/C1 acceptance remain open.
Validation:
58cc5817e35e+ C1 C1: unify work sources and validate recovery #8266b8ac1ec5443b: 22 selected cases passed, with source-byte readback.The HTTP/action/store and synthetic Codex protocol exercise the production entrypoint on isolated File/SQLite state. These checks do not qualify native-model context retention, installation closure, full costs or the final original-goal regression.