Skip to content

fix(steward): settle autonomous replan inline and retry pre-host deferrals - #9188

Merged
loopx-agent merged 1 commit into
mainfrom
codex/steward-replan-selection-recovery-20261121
Nov 21, 2026
Merged

loopx-agent merged 1 commit into
mainfrom
codex/steward-replan-selection-recovery-20261121

Conversation

@loopx-agent

Copy link
Copy Markdown
Collaborator

Summary

  • Goal / Problem: When a Steward Goal reaches periodic_review_due, advancing monitor follow-up Todos (work_kind == "native_turn") are deferred pre-host (turn_todo_selection_deferred / reason_code: "autonomous_replan") with zero side effects (host_invoked=false, state_written=false, quota_spent=false, scheduler_acknowledged=false). Previously, StewardTurnHost.execute only supported inline .replan governance for work_kind == "report", text-only hosts (e.g., restricted Antigravity) lacked an inlined public-safe replan context, and resolve_advancement_turn_id exhausted its single retry slot on zero-effect pre-host deferrals.
  • Changes:
    • loopx/capabilities/native_chat/steward_monitor.py: Add _zero_write_advancement_response, _is_zero_effect_selection_deferred_advancement_turn, and ADVANCEMENT_ADMISSION_RETRY_SUFFIX (.admission-retry) in resolve_advancement_turn_id so bounded zero-effect turn_todo_selection_deferred attempts unlock one post-cooldown retry carrier on base or .precommit-retry.
    • loopx/capabilities/native_chat/steward_turn.py: Add _is_zero_effect_replan_selection_deferral, StewardTurnHost._governance_replan_context (pre-reading quota replan-view review/writeback contracts, current agent_vision, and open agent Todos with strict local-path/CLI stripping), governance prompt contract + evidence normalization during governance_work, and inline .replan execution + primary re-run on steward_advancement when blocked by autonomous_replan.
    • tests/test_steward_generic_monitor_todo.py: Add unit tests for resolve_advancement_turn_id across cooldown, host_invoked=True, and .precommit-retry.admission-retry, plus end-to-end test (execution_replan_after_turns=1) verifying inline .replan settlement, public-safe governance prompt, and follow-up commit in one turn.

Validation

  • ruff check loopx/capabilities/native_chat/steward_turn.py loopx/capabilities/native_chat/steward_monitor.py tests/test_steward_generic_monitor_todo.py -> passed (All checks passed!)
  • pytest -q tests/test_steward_generic_monitor_todo.py -> 6 passed in 58.91s
  • pytest -q tests/test_native_steward_reports.py -k "replan or selection" -> 22 passed, 143 deselected in 329.02s
  • pytest -q tests/architecture/test_semantic_vocabulary_drift.py -> 124 passed in 37.97s
  • python examples/semantic-vocabulary-drift-smoke.py --diff-base origin/main -> ok
  • loopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main -> gate.status == "passed", selected=3, failures=0
  • loopx check --scan-path ... -> ok: true, errors: 0, public boundary scan clean: 3 files

@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)

Exact reviewed head: 9188@971c534d1d8f7687f914bf278ecab215317cf31e

Reviewer: model_agent · Gemini · Google · self_reported

动机 / Motivation

启用 LoopX Steward 作用域 PR 信号监控并触发 periodic_review_due 周期重规划阈值的 Goal 维护者与监控泵。此前当 Goal 达到 periodic_review_due(连续 6 次结算 Turn)后,pump_steward_monitors 派发的 advancement_task(native_turn)因携带 --todo-id 在调用模型前即被 quota.should-run 拒绝(turn_todo_selection_deferred,四项 kernel effects 全为 false),既未内联清算重规划,受限文本宿主也无法运行 CLI 读取 stride_control 契约,同时 resolve_advancement_turn_id 将零副作用延迟消耗进唯一重试槽位导致永久卡死;修复后零副作用延迟可于 60 秒后解锁一次 .admission-retry,StewardTurnHost 在同一 Turn 内联完成无边界 .replan 治理(注入不含本地路径的治理上下文)并恢复提交原跟进任务。Steward 跟进任务在遇到 autonomous_replan 门禁阻塞后,可在同一 Turn 完成内联 .replan 并提交跟进报告;此前被 pre-host 零副作用延迟耗尽 .precommit-retry 的任务也可在冷却 60 秒后通过 .admission-retry 自动恢复一次。本 PR 不放宽 quota.should-run 的周期重规划门禁,不修改通用 report_selection_recovery 既有契约,不在宿主触发后(host_invoked=true)重放失败轮次,也不改变 monitor 轮询周期与非 Steward Goal 的 Turn 驱动流程。合入前已通过单测、端到端模拟和全部门禁冒烟;真实常驻 com.loopx.chat 进程上的重规划内联清算与飞书报告送达读回需在合入后构建新本地发行版并重启实机验证。

改动思路 / Approach

内联 .replan 调度、只读 replan-view 上下文投影与零副作用预提交延迟识别分别收敛在既有 Steward 宿主适配器(steward_turn.py)与 Steward 轮询派发层(steward_monitor.py),复用了既有 quota/turn_run_once 内核契约与 vision 校验,未引入平行状态存储或冗余协议字段。仅作用于 steward_advancement(work_kind='native_turn' 且非本地托管会话)面临零副作用 autonomous_replan 预拒绝或 resolve_advancement_turn_id 判定 .admission-retry 的狭窄边界。本 PR 沿着与现有 report_selection_recovery 对称但面向 steward_advancement 的架构路径推进:入口仍由 pump_steward_monitors 读取 open 跟进 Todo 并通过 resolve_advancement_turn_id 生成唯一客户端身份;决策与写入权限仍完全由 quota.should-run 和 turn_run_once 内核掌握,宿主层绝不绕过租约或重规划门禁,仅在捕获确切的零副作用(host_invoked=false, state_written=false, quota_spent=false, scheduler_acknowledged=false)autonomous_replan 拒绝时,以无任务选择的 .replan 通道完成重规划后再在同一租约窗口重跑原任务。此外,由于生产环境中的受限文本宿主(如 restricted Antigravity)禁用了 shell 工具,_governance_replan_context 在宿主进程内直接调用只读 read_replan_view 并剥离本地绝对路径后把契约内联进上下文,既保证受限模型能构造合规 agent_vision_json,又严格守住零本地路径泄露边界。

flowchart TD
    A["pump_steward_monitors 扫描 open 跟进任务"] --> B["resolve_advancement_turn_id"]
    B -->|无历史轮次| C["派发 advancement.hash"]
    B -->|历史 zero-effect deferral 且满 60s| D["派发 .admission-retry 单次载体"]
    B -->|已触发宿主 host_invoked=true 或重试耗尽| E["严格返回 None 不再重试"]
    C --> F["StewardTurnHost.execute -> run_primary_turn()"]
    D --> F
    F -->|命中 autonomous_replan 零副作用拒绝| G["todo_id=None 执行 .replan + 注入无路径 gov_ctx"]
    G -->|replan committed 或 guard 已清空| H["同一租约内重跑 run_primary_turn() 并提交报告"]
    G -->|replan 失败或未提交| I["合并治理 effects 标记 host_invoked=true 阻断重试"]
Loading

说明(Illustrative · head 971c534d1d8f7687f914bf278ecab215317cf31e):图示梳理 steward_monitor.py 零副作用重试载体判定与 steward_turn.py 内联 .replan 控制流及失败保护分支;实际执行以通过的单测与端到端回归为准。

具体改动 / Changes

对照契约基础(none (problem_context & code invariants govern this internal Steward fix)),逐文件精准拆分边界并控制体量(总变更 3 文件 +585/-38,实现 +237/-38 与测试 +348 比例克制):

  • loopx/capabilities/native_chat/steward_monitor.py:抽取共享辅助 _zero_write_advancement_response 校验已完成、零 Steward 副作用输出与 60 秒冷却窗口;新增 _is_zero_effect_selection_deferred_advancement_turn;更新 resolve_advancement_turn_id 在 base 或 .precommit-retry 遇到零副作用预延迟时分别解锁至多一次 .admission-retry。
  • loopx/capabilities/native_chat/steward_turn.py、L1254-L1325、L1567-L1815:新增 _is_zero_effect_replan_selection_deferral 与 _governance_replan_context;扩展 governance_work 提示词并复用 _normalize_steward_advancement_kernel_candidate;将 handle_turn_command 收敛至 run_primary_turn(),并在命中零副作用 autonomous_replan 拒绝时执行 unbound .replan、写入 native_loopx.replan_result 事件,清空后重跑主任务或回填 True effects。
  • tests/test_steward_generic_monitor_todo.py:新增 test_resolve_advancement_turn_id_handles_selection_deferral_and_precommit_retries 与 test_scoped_signal_monitor_settles_autonomous_replan_inline_and_commits_followup。

关键代码讲解

  1. _zero_write_advancement_response(steward_monitor.py:75)与 resolve_advancement_turn_id(steward_monitor.py:150):将预提交失败与 pre-host 延迟的共同门槛(status=="completed"、无报告/handoff/user_action 副作用、满 60 秒冷却)抽为一处,随后精确区分需要等待 .precommit-retry 的 failed 轮次与允许进入 .admission-retry 的四项 effect 全为 False 的 turn_todo_selection_deferred 轮次,一旦重试载体已有记录立即返回 None。
  2. StewardTurnHost._governance_replan_context(steward_turn.py:1254):利用 shlex.split 与 build_parser().parse_args 解析 replan_action_packet 中的只读 read_replan_view(review 与 writeback),仅投影 obligation_id、context_id、objective、triggers、recent_evidence(前 12 条)、writeback_contract、stride_control、当前 agent_vision 与开放中的 Agent Todos,刻意过滤 read_action 与 next_cli_actions 杜绝任何本地绝对路径或 CLI 模板进入模型输入。
  3. StewardTurnHost.execute 内联重规划序列(steward_turn.py:1766):仅对 steward_advancement 且非 resume_turn_key、且确由 _is_zero_effect_replan_selection_deferral 认定的拒绝切入 .replan 无绑定治理轮次;唯有 governance 提交 replan_required 或 report_replan_not_admitted 且 _cleared_replan_guard 验证守卫确已转绿时才清表重跑 run_primary_turn()。

对主干的风险 / Risks

  • 活跃调用点与最小反例追踪:仅由生产轮询泵 pump_steward_monitors 与宿主执行器 StewardTurnHost.execute 消费,无孤立抽象;与已有 report_selection_recovery 正交隔离。最坏边界场景是内联 .replan 调用模型后模型输出畸形导致 validation_failed 或 wait:此时 StewardTurnHost.execute 会把 governance.effects 中的真实 host_invoked: True 等标记并入主 result["effects"],保证下轮轮询时 _is_zero_effect_selection_deferred_advancement_turn 严格返回 False,彻底杜绝反复试探消耗模型配额。
  • 验证矩阵与负向覆盖:已实跑 ruff check 全过、tests/test_steward_generic_monitor_todo.py(6 passed in 58.91s,包含冷却期拦截、host_invoked=True 阻断、.precommit-retry.admission-retry 耗尽封顶以及 execution_replan_after_turns=1 端到端单回合闭环与无路径泄露断言)、tests/test_native_steward_reports.py -k 'replan or selection'(22 passed,证明既有报告路径零回归)、tests/architecture/test_semantic_vocabulary_drift.py(124 passed)、examples/semantic-vocabulary-drift-smoke.py --diff-base origin/main(ok)、loopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main(passed 3/3)及 loopx check(公开边界扫描 3 文件 0 告警)。
  • Mock 边界说明:合入前测试使用可控 fake app-server/SQLite harness,真实常驻服务重启后的实机模型响应与飞书送达由合入后发布验收闭环。

我的整体评价 / Verdict

本改动精准解决了个人管家在周期重规划门禁生效后跟进任务永久阻塞的高频运维痛点,同时完整保留了内核控制面、任务租约与公开边界隔离,长程自治效用与工程性价比显著提升;既有汇报及托管会话路径通过对比回归验证保持等价隔离。无前端 UI 渲染画面改动;仅涉及 Steward 后台轮询与 Turn 宿主逻辑及单元测试。无前端渲染画面改动,未执行界面截图;实机端到端线上恢复与飞书投递读回将在 PR 合入后于本地服务核验。内联 .replan 调度、只读 replan-view 上下文投影与零副作用预提交延迟识别分别收敛在既有 Steward 宿主适配器(steward_turn.py)与 Steward 轮询派发层(steward_monitor.py),复用了既有 quota/turn_run_once 内核契约与 vision 校验,未引入平行状态存储或冗余协议字段。仅作用于 steward_advancement(work_kind='native_turn' 且非本地托管会话)面临零副作用 autonomous_replan 预拒绝或 resolve_advancement_turn_id 判定 .admission-retry 的狭窄边界。

English Verdict: APPROVE

@loopx-agent
loopx-agent merged commit f08fed8 into main Nov 21, 2026
23 of 42 checks passed
@loopx-agent
loopx-agent deleted the codex/steward-replan-selection-recovery-20261121 branch November 21, 2026 12:03
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