Skip to content

fix(steward): bound advancement evidence items and allow one pre-commit retry - #9187

Merged
loopx-agent merged 1 commit into
mainfrom
codex/steward-advancement-evidence-retry-20261121
Nov 21, 2026
Merged

loopx-agent merged 1 commit into
mainfrom
codex/steward-advancement-evidence-retry-20261121

Conversation

@loopx-agent

Copy link
Copy Markdown
Collaborator

Summary

  • Normalize over-long Steward advancement evidence: Add _normalize_steward_advancement_kernel_candidate and _split_bounded_evidence_items in loopx/capabilities/native_chat/steward_turn.py (for steward_advancement turns) to cleanly split work_evidence / validation_evidence items exceeding the kernel per-item <=240 char limit along natural clause/whitespace boundaries and compact within kernel limits (<=8 items, <=1200 total chars) while preserving malformed shapes for kernel validation and leaving generic turn_driver clipping refusal untouched. Also append a <=200 char prompt reminder.
  • Prevent permanent wedges on pre-commit advancement.<hash> failures: Add ADVANCEMENT_PRECOMMIT_RETRY_COOLDOWN_SECONDS = 60, _is_precommit_retryable_advancement_turn, and resolve_advancement_turn_id in loopx/capabilities/native_chat/steward_monitor.py. When an advancement turn failed at validation_failed or host_failure with zero kernel writes (writeback=False / state_written=False, quota_spent=False, scheduler_acknowledged=False) and zero Steward side effects (steward_owner_report_queued, context_handoff, steward_user_action_receipts), pump_steward_monitors allows at most one .precommit-retry turn after a 60s cooldown instead of skipping the open follow-up Todo forever.
  • Tests: Add focused regression coverage in tests/test_steward_generic_monitor_todo.py for lossless over-long evidence splitting, overflow packing, malformed evidence preservation, 60s retry cooldown, single .precommit-retry dispatch/commit, and post-commit non-retry guarantees.

Verification

  • ruff check loopx/capabilities/native_chat/steward_turn.py loopx/capabilities/native_chat/steward_monitor.py tests/test_steward_generic_monitor_todo.py: passed
  • pytest -q tests/test_steward_generic_monitor_todo.py (final diff): 4 passed
  • pytest -n 4 -q tests/test_steward_generic_monitor_todo.py tests/test_native_steward_turn_host.py tests/test_native_steward_monitor.py tests/test_native_steward_reports.py tests/test_lark_steward_reports.py tests/test_steward_report_recovery_and_validator.py tests/test_native_steward_managed_mode.py: 491 passed
  • loopx check on changed paths: ok (errors=0, warnings=0, public boundary clean)
  • loopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main: passed (selected=3, failures=0, including control-plane-maintainability-ratchet-smoke.py, semantic-vocabulary-drift-smoke.py, and monitor-scheduler-contract-smoke.py)

@loopx-agent loopx-agent left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

Reviewer: model_agent · Gemini · Google · self_reported

动机 / Motivation

部署并启用 LoopX Steward 作用域 PR 信号监控的维护者在触发聚合汇报跟进任务时。此前模型生成聚合报告与超过 240 字符的证据条目时,内核校验会在回写前整轮失败,且同一修订版本号下的后续任务将被永久跳过;本次重构在发布适配器中拆短证据并在零写入预提交失败后允许一次冷却重试。超长但有效的证据字符串在 Steward 适配器中无损拆分后可通过内核长度限制并送达聚合报告,未产生任何写入的预提交失败任务在 60 秒后仅重试一次。本次不放宽通用内核的 240 字符限制与裁剪拒绝策略,不重试已写状态或已排队的轮次,也不更改普通报告或周期任务路径。合入前仅通过单元测试与契约冒烟验证;真实线上 Steward 会话的二次恢复与飞书投递读回需在合入并更新本地发行版后核验。

改动思路 / Approach

本 PR 在进入通用内核结果校验与轮询去重之前于 Steward 适配层分别消除两个阻塞:StewardTurnHost.execute 仅在 steward_advancement 路径归一化 work_evidence/validation_evidence,保持 host_candidate 对通用动作裁剪的严格拒绝;pump_steward_monitors 通过 resolve_advancement_turn_id 判定零副作用预提交失败后再决定是否派生 .precommit-retry。将证据长度适配限定在 Steward 宿主响应归一化层并将一次预提交重试限定在轮询泵,保留了通用 Turn 内核校验与已提交幂等性,未扩大故障影响面。仅覆盖 Steward 作用域跟进轮次的证据条目上限拆分及零写入预提交失败的单次冷却重试;通用 CLI 或已写入任何副作用的轮次不在重试范围内。

flowchart TD
    A["pump_steward_monitors 读取 open 跟进任务"] --> B["resolve_advancement_turn_id 计算 advancement.&lt;hash&gt;"]
    B -->|无历史轮次| C["首次派发原生 Turn"]
    B -->|历史 validation_failed/host_failure 且零写入/零报告/满60秒| D["仅派生 1 次 .precommit-retry"]
    B -->|已提交写入/已有报告/已重试/未满60秒| E["保持跳过 禁止重复执行"]
    C --> F["StewardTurnHost.execute"]
    D --> F
    F --> G["_normalize_steward_advancement_kernel_candidate 按语义分句拆短 &lt;=240 字符"]
    G --> H["通用 executor 内核校验 -&gt; 提交并写 pending 投递文件"]
Loading

说明(Illustrative · head 17e7fef4500b2953c11d9614471207795276cf1a):图示描述 steward_monitor.py 的单次重入判定与 steward_turn.py 的适配器归一化边界;真实行为以单测与源码校验为准。

具体改动 / Changes

关键代码讲解

  1. loopx/capabilities/native_chat/steward_turn.py — _split_bounded_evidence_items 与 _normalize_steward_advancement_kernel_candidate:
    • 在保留非列表、空列表或非字符串等非法输入以触发内核明确报错的前提下,对合法非空字符串列表按分号、句号、逗号和空格等自然边界拆分为单条不超过 240 字符的片段,并在超出 8 条或总长 1200 字符时进行有界压缩与省略号截断;同时兼容把 multi_surface_or_implementation 规范化为 multi_surface。
    • 于 StewardTurnHost.execute 的 steward_advancement 分支提取 steward_owner_report 等副作用字段之后调用该归一化,并在提示词补充单条不超过 200 字符、总长不超过 1200 字符的约束提示。
  2. loopx/capabilities/native_chat/steward_monitor.py — _is_precommit_retryable_advancement_turn 与 resolve_advancement_turn_id:
    • 仅当基础 advancement.<hash> 轮次已完成、内核状态为 failed(validation_failed 或 host_failure)、commit_eligibility.writeback / effects 证明无任何状态写入/额度扣除/调度确认、且响应未包含 steward_owner_report_queued/context_handoff/steward_user_action_receipts、且距完成已满 60 秒时,才允许返回一次 .precommit-retry token;若 .precommit-retry 记录已存在则直接跳过。
    • 在 pump_steward_monitors 调用 resolve_advancement_turn_id 取代原有的裸 turn_for_client is not None 永久跳过判断。
  3. tests/test_steward_generic_monitor_todo.py:
    • 新增 _normalize_steward_advancement_kernel_candidate 单测与作用域监控端到端恢复回归测试,覆盖无损拆分、溢出打包、畸形输入保留报错、60 秒冷却窗口、单次 .precommit-retry 成功提交及拒绝第三次重试和拒绝已入队报告重试。

对主干的风险 / Risks

  • 范围与兼容性:修改仅限 Steward 原生推进路径及对应测试文件,未改动通用 turn_driver/executor.py 或 host_candidate.py,因此非 Steward 任务与既有通用裁剪拒绝契约完全不变(已通过 test_native_steward_turn_host.py、test_native_steward_managed_mode.py 及 canary premerge 核验)。
  • 幂等与防重放风险:若极少数情况下历史轮次遗漏了 effects/receipt 字典(类型为 None 或缺字段),_is_precommit_retryable_advancement_turn 默认拒绝重试(fail-closed),避免误重放副作用;对于确认为零写入的失败,.precommit-retry 最多触发一次。
  • 验证覆盖与已知主干现状:7 组 Steward 相关测试共 491 项通过,tests/test_steward_generic_monitor_todo.py 最终 diff 4 项通过,ruff check、loopx check 及 canary premerge --tier quick(3 项 catalog smoke 含词表漂移检查)全部通过;另外独立的 tests/test_loopx_turn_driver.py 209 项通过、有 2 项 initial_before 用例失败,已于未改动 main 基线复现确认属主干既有问题,与本次 Steward 修改无因果关联。

我的整体评价 / Verdict

本次改动以最小、可逆的边界修复了 Steward 作用域聚合巡检因 work_evidence 超过 240 字符而被内核拒绝且永久卡死的问题,对长程自治可用性与用户体验均为正向改善(PASS);无 UI 渲染改动(本次仅修改 Steward 后端适配器与监控轮询逻辑(loopx/capabilities/native_chat/steward_turn.py、steward_monitor.py)及单元测试,未改动任何前端 UI 渲染画面。),未验证实机送达范围为:无前端渲染画面改动,未执行额外截图;实机端到端送达将在合入后于本地服务核验。。所有核心受影响不变式均具备当前 17e7fef4500b2953c11d9614471207795276cf1a 头的独立验证证据。

English Verdict: APPROVE

@loopx-agent
loopx-agent merged commit 5e2c896 into main Nov 21, 2026
25 of 43 checks passed
@loopx-agent
loopx-agent deleted the codex/steward-advancement-evidence-retry-20261121 branch November 21, 2026 11:22
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant