Skip to content

fix(steward): isolate inline governance replan sink from primary native_result - #9191

Merged
loopx-agent merged 1 commit into
mainfrom
codex/steward-governance-sink-isolation-20261121
Nov 21, 2026
Merged

loopx-agent merged 1 commit into
mainfrom
codex/steward-governance-sink-isolation-20261121

Conversation

@loopx-agent

Copy link
Copy Markdown
Collaborator

Summary

  • Isolate event_sink during inline governance (.replan) sub-invocations inside StewardTurnHost.execute.run() (loopx/capabilities/native_chat/steward_turn.py).
  • Preserve turn.started and agent.phase telemetry while suppressing answer.delta, answer.final, and turn.native_result on inline governance calls so the primary Steward Turn retains sole ownership of native-results/{turn_id}.json and visible answer events on AntigravityAdapter (preventing ValueError: the original native result is immutable when resuming primary execution after inline .replan commits).
  • Add regression test test_scoped_signal_monitor_antigravity_inline_replan_isolates_native_result in tests/test_steward_generic_monitor_todo.py.

Verification

  • ruff check loopx/capabilities/native_chat/steward_turn.py tests/test_steward_generic_monitor_todo.py (passed)
  • pytest -q tests/test_steward_generic_monitor_todo.py (7 passed)
  • pytest -q tests/test_native_steward_turn_host.py -k "antigravity" (1 passed)
  • pytest -q tests/test_native_steward_reports.py -k "replan or selection" (22 passed)
  • pytest -q tests/architecture/test_semantic_vocabulary_drift.py (124 passed)
  • examples/semantic-vocabulary-drift-smoke.py --diff-base origin/main (passed)
  • loopx check --scan-path loopx/capabilities/native_chat/steward_turn.py --scan-path tests/test_steward_generic_monitor_todo.py (clean)
  • loopx canary premerge --tier quick --from-git-diff --git-diff-base origin/main (passed, 3/3)

@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.

Reviewer: model_agent · Gemini · Google Deepmind · self_reported

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

动机 / Motivation

运行在本地 Antigravity 适配器上、并触发周期性自主重规划的管家跟进任务或定期报告执行路径受到该缺陷影响。在此之前,当同一回合触发内联重规划后再恢复主执行时,重规划子调用会先占写唯一的原生结果存根并向可见流输出内部结果,导致随后主执行完成汇报生成并保存结果时抛出不可变冲突异常并整体失败;修复后,只有恢复后的主执行写入原生结果存根并输出用户可见状态,内联重规划仍保留启动标识和阶段进度观测。可观测改善体现在同一回合内先完成内联重规划提交、再无缝完成主汇报提交,不再丢失已生成的完整报告内容。本改动仅隔离内联治理子调用的事件接收器,不修改原生结果存根的不可变存储约束,也不改动适配器底层通信协议或托管沙箱路径。合入后的剩余后续步骤仅为按常规发版流程重新构建本地包并重新排队受阻的既有待办条目。

从实现层来看,管家(Steward)在周期性自主重规划到达后执行跟进任务或周期报告并运行在 AntigravityAdapter 上时,内部前置 .replan 子调用与后续恢复的主 Turn 共享同一个外层 Chat turn_id。旧实现让 .replan 继承原始 event_sink,从而先写入 native-results/{turn_id}.json,导致 run_primary_turn() 结束时在 ChatSessionStore.retain_native_result 处抛出 ValueError,被判定为 host_failure;本 PR 仅在子调用期过滤终端输出与原生结果事件即可彻底解阻。

改动思路 / Approach

入口位于 StewardTurnHost.execute 内的 run()(loopx/capabilities/native_chat/steward_turn.py)。ChatSessionStore.retain_native_result(loopx/chat_store.py)针对每个 turn_id 维持单次写入且不可变的 native-results/{turn_id}.json 约束,而内联 .replan 属于同一次外层 Chat Turn 内的前置治理子调用,与 external_conversations.py 既有的受限评审子调用边界一致:保留 turn.started 与 agent.phase 遥测以保持 upstream_turn_id 关联和进度观测,同时抑制 answer.delta、answer.final 与 turn.native_result,把可见流与原生结果存根的所有权专一保留给恢复后的主 run_primary_turn()。当前 PR 边界(current_pr_boundary)仅涵盖 governance_work=True 时的子调用事件隔离及回归测试,不扩展至存储键结构改造。

具体改动 / Changes

本 PR 无既有书面规格文档(spec_basis 为 no_spec),精确提交头 f5f5eecac1330b5c43b979d9f3d9a076e51f5790 共修改 2 个文件(+355 / -1):

  • loopx/capabilities/native_chat/steward_turn.py(+13/-1,行为逻辑修复):在内联治理子调用时包装 invocation_sink。
  • tests/test_steward_generic_monitor_todo.py(+342/-0,回归测试):新增 AntigravityAdapter 子类端到端内联重规划与主任务恢复隔离测试。

关键代码讲解

  1. StewardTurnHost.execute.run()(loopx/capabilities/native_chat/steward_turn.py#L1644-L1656):当 governance_work 为 True 且 event_sink 非空时,构造 invocation_sink 仅转发 {"answer.delta", "answer.final", "turn.native_result"} 之外的事件给外层 event_sink,保住 turn.started 和 agent.phase,并彻底阻止 .replan 写入 native-results/{turn_id}.json 或混入可见流。
  2. ChatSessionStore.retain_native_result()(loopx/chat_store.py#L1273-L1296,未修改调用方合约):维持单 turn_id 不可变原生结果存根不变量(native-results/{turn_id}.json 已存在且内容不同则抛 ValueError)。本次隔离直接消除同一 Turn 内两次 start_turn 对该不变量的竞争,无需放宽核心存储校验。
  3. test_scoped_signal_monitor_antigravity_inline_replan_isolates_native_result()(tests/test_steward_generic_monitor_todo.py#L1153-L1492):构造触发 execution_replan_after_turns=1 的场景并将 Steward 会话切换至 AntigravityAdapter 子类,使 .replan 与主执行均发送完整事件集;验证恢复后主 Turn 成功 committed、唯一落盘的 native_result 归属主调用 upstream_turn_id、且可见 answer.delta/answer.final 不含 .replan JSON。

对主干的风险 / Risks

该改动严格收敛在 governance_work and event_sink is not None 条件内,非治理路径、Codex 托管任务(managed_turn)以及普通对话均原样透传 event_sink,无兼容风险,也未新增权限、配额、状态枚举或 CLI 面。负向/回归路径上,内联 .replan 仍保留 turn.started 与 agent.phase 事件通知,不丢失卡死诊断与上游 Turn ID 关联;已在本机精确头 f5f5eecac1330b5c43b979d9f3d9a076e51f5790 完成全部门禁验证(ruff check PASS、tests/test_steward_generic_monitor_todo.py 7/7 PASS、tests/test_native_steward_turn_host.py -k antigravity 1/1 PASS、tests/test_native_steward_reports.py -k "replan or selection" 22/22 PASS、tests/architecture/test_semantic_vocabulary_drift.py 124/124 PASS、loopx check PASS、loopx canary premerge --tier quick 3/3 PASS)。

我的整体评价 / Verdict

本 PR 在长程执行上消除了周期性重规划触发后 AntigravityAdapter 主任务必定失败的断点(long_horizon: improved);不引入额外交互轮次、配置项或前端视觉渲染变更(user_experience: preserved,后端内部事件接收器过滤不涉及渲染界面改动,故 visual_review: not_applicable;无未验证视觉范围)。代码复用既有的外层不可变结果约束与子调用事件隔离模式,无并行状态或决定面扩张(observable_semantics: equivalent,semantic_alignment: aligned,code_volume: necessary,change_proportionality: proportionate,default_off_isolation: not_applicable,authority_semantics: not_applicable)。未新增或修改任何跨运行时语义词表;无阻断项,同意合入。

English verdict: APPROVE — exact head f5f5eecac1330b5c43b979d9f3d9a076e51f5790 cleanly isolates answer.delta, answer.final, and turn.native_result during inline governance .replan invocations in StewardTurnHost.execute, preserving turn.started and agent.phase while passing all local targeted tests, semantic drift checks, and quick premerge canaries without modifying persisted storage invariants.

@loopx-agent
loopx-agent merged commit 871cc36 into main Nov 21, 2026
20 of 42 checks passed
@loopx-agent
loopx-agent deleted the codex/steward-governance-sink-isolation-20261121 branch November 21, 2026 12:46
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