fix: retain session ownership until compaction cancellation completes

.agents/notes/implemented/bug-fix/2026-09-12-native-compaction-cancellation.md

@@ -0,0 +1,106 @@
+# Cancel compaction through the native turn
+
+Status: implemented
+Translation: current
+
+[中文](2026-09-12-native-compaction-cancellation.zh.md)
+
+## Abstract
+
+Manual Codex `/compact` was treated as a command without a native turn, so Stop
+could return an ACP cancellation while Codex kept compacting. The adapter now
+captures the native turn and interrupts it through the ordinary cancellation
+path, retaining ownership until terminal confirmation. Lody sends provider cancel
+without interrupting an in-flight prompt's owner fiber, so normal cancellation
+finalizes history only after ACP returns. A shared five-second drain terminates
+an unresponsive provider; failed termination keeps ownership. PR #618 removes its
+automatic historical reconciliation protocol and fixes these execution boundaries.
+Existing stale histories are not migrated by opening a conversation.
+
+## Decision
+
+Codex 0.153.4 emits standard turn and item notifications for manual compaction.
+`thread/compact/start` acknowledges submission with an empty result; the turn id
+arrives through `turn/started`. The previous assumption that compaction had no
+interruptible turn confused an empty acknowledgement with an absent lifecycle.
+
+The app-server client registers an owner before submission and resolves only on
+the matching `turn/completed`. Start failure and connection closure reject the
+owner. The command routes the native id and terminal result through the same ACP
+command lifecycle used by other native turns. Stop and request cancellation both
+interrupt the captured turn. A cancellation before its id arrives stays pending
+until the turn starts; the ACP prompt remains occupied while cancellation drains.
+
+Lody records `runtime.cancelRequested` and sends provider cancel while the prompt
+is in flight. It retains the existing owner fiber and runtime until ACP returns;
+the scope then runs cancellation finalization and releases ownership. A new user
+message remains pending and cannot reach ACP during that interval. Steer admission
+also checks the existing cancellation flag after asynchronous preparation and after
+the provider's acceptance ACK. A successful ACK after Stop cannot replace the source
+invocation, force its settlement to handled, or transfer ownership. It marks that exact
+steer user turn `canceled` without changing dispatch pointers, then returns `stale-turn`
+and releases the application lease. The accepted steer cannot remain `pending_apply`
+or be requeued; the existing owner continues to its cancelled terminal outcome. Neither the
+daemon nor the client treats that disposition as permission to replay the steer.
+Cancellation before prompt submission and finalization teardown retain their existing paths.
+
+The earlier CLI code already retained the runtime through `pendingPromptCompletion`
+inside its scope finalizer; it did not unconditionally release ownership at Stop.
+It nevertheless interrupted the owner fiber and finalized history before draining
+the provider. Normal Stop now waits in the prompt itself and starts the existing
+five-second raw-request drain independently of the cancel acknowledgement. If the
+prompt remains pending at the deadline, Lody terminates that session; connection
+closure rejects the prompt and lets the owner enter its finalizer naturally.
+Failed termination waits for raw ACP completion without releasing ownership.
+Stop and external-interruption finalization share one drain promise on the runtime,
+so repeated cancellation neither resets the deadline nor terminates twice.
+This restores #571's recovery policy, which the earlier in-flight Stop change had
+unintentionally restricted to external interruption.
+
+After ACP completion, the existing CLI finalization from
+[provider-failure settlement](2026-09-10-context-compaction-terminal-state.md)
+then persists unresolved compaction as failed before releasing execution ownership.
+Explicit provider terminal item updates remain authoritative.
+
+The Session view does not initiate data repair. The branch's new reconciliation
+capability, both transport methods, renderer retry hook, and daemon history repair
+are removed. Historical unresolved records are a separate maintenance concern;
+neither hiding progress nor rewriting history can interrupt native execution.
+
+## Evidence and verification
+
+- [Pinned native protocol](https://github.com/openai/codex/blob/rust-v0.153.4/codex-rs/app-server/README.md#example-trigger-thread-compaction).
+- Adapter owner: `packages/acp-extension-codex/src/CodexAppServerClient.ts`;
+  command and cancel routing: `CodexCommands.ts` and `CodexAcpServer.ts` in that directory.
+- Deterministic tests cover Stop and request abort before/after native start,
+  rejection of a second prompt while draining, successful continuation after
+  interruption, failure, terminal events before the start ACK, start rejection,
+  unrelated turn completion, and process exit before/after native start.
+- The Lody execution suite uses the real `AgentClient` with a controlled ACP
+  transport: cancel ACK leaves the prompt signal live and history unfinished,
+  a second dispatch stays pending, and native terminal evidence enables the next
+  prompt. Fake timers cover normal Stop's five-second termination, a withheld cancel
+  acknowledgement, repeated Stop without deadline reset, and failed termination
+  retaining unfinished history and ownership until ACP ends. External-interruption
+  coverage retains raw completion, process termination, and failed-termination cases.
+- Steer coverage retains undelivered history and its dispatch pointer when Stop
+  precedes the request or arrives while prompt blocks are being built. A controlled
+  acceptance ACK arriving after Stop preserves the source invocation, user-turn and
+  dispatch owner, leaves source history unfinished, and never reports handled or
+  replays the steer. The accepted steer is `canceled` both before and after provider
+  terminal, which settles the source as cancelled and releases ownership. The terminal
+  steer assertion failed on `ae24a723` with the entry stranded in `pending_apply`.
+- Contract: [Session history writes](../../../../specs/session-history-writes.md).
+- Pull request: [Lody #618](https://github.com/LodyAI/Lody/pull/618).
+- Adapter implementation: [Codex adapter #41](https://github.com/LodyAI/acp-extension-codex/pull/41).
+
+Adapter typechecks and all 617 enabled adapter tests passed; 27 tests are skipped
+by their existing environment gates. Real Codex smoke checks completed manual
+compaction normally and confirmed that Stop produces native `interrupted` before
+the ACP prompt returns `cancelled`.
+
+Root typechecks, lint, formatting, i18n, documentation and boundary checks passed.
+The targeted Lody execution, dispatch-watcher and AgentClient suites pass 207 tests.
+The full `pnpm check` reached Electron tests: 103 passed, while the relay suite
+could not load because this checkout lacks the installed Electron binary. All
+preceding workspace test suites passed.

.agents/notes/implemented/bug-fix/2026-09-12-native-compaction-cancellation.zh.md

@@ -0,0 +1,78 @@
+# 通过 native turn 取消压缩
+
+Status: implemented
+Translation: current
+
+[English](2026-09-12-native-compaction-cancellation.md)
+
+## 摘要
+
+手动 Codex `/compact` 原本被当作没有 native turn 的命令，Stop 可以返回 ACP 取消，
+但 Codex 仍在压缩。适配器现在捕获 native turn，通过常规取消路径中断它，并持有 ownership
+直到 terminal 确认。Lody 发送 provider cancel，不中断 in-flight prompt 的 owner fiber，
+因此正常取消在 ACP 返回后才完成历史收尾；共用的五秒 drain 会终止无响应的 provider，
+终止失败则继续持有 owner。PR #618 删除自动历史 reconciliation 协议，改为修复这些执行边界；
+打开会话不会迁移已有的陈旧历史。
+
+## 决策
+
+Codex 0.153.4 的手动压缩会发出标准 turn 和 item 通知。`thread/compact/start` 用空结果确认
+提交，turn id 随 `turn/started` 到达。此前认为压缩没有可中断 turn，是把空 ACK 与不存在
+生命周期混为一谈。
+
+app-server client 在提交前注册 owner，只在匹配的 `turn/completed` 到来时 resolve。
+启动失败和连接关闭会 reject owner。命令通过普通 native turn 的 ACP 命令生命周期传递
+native id 和 terminal 结果。Stop 与请求取消均中断捕获的 turn；若取消早于 id 到达，
+则等待 turn 启动后中断。在取消 drain 期间，ACP prompt 一直被占用。
+
+Lody 在 prompt in-flight 时记录 `runtime.cancelRequested` 并发送 provider cancel，
+保留原有 owner fiber 和 runtime，直到 ACP 返回，再由 scope 执行取消收尾并释放 ownership。
+期间新用户消息保持 pending，不能到达 ACP。steer 在异步准备后及 provider 接受 ACK 后都
+检查取消标记。Stop 后到达的成功 ACK 不得替换 source invocation、将 source 强制结算为
+handled 或转移 ownership。它将该 exact steer 用户轮次标为 `canceled`，不改变 dispatch pointer，
+然后返回 `stale-turn` 并释放 application lease。已接受的 steer 不能停留在 `pending_apply`，
+也不能重新排队；现有 owner 继续等待 cancelled terminal。daemon 与客户端均不会把这个
+disposition 当作重放 steer 的许可。提交 prompt 前的取消及 finalization teardown 保留原路径。
+
+此前 CLI 的 scope finalizer 已通过 `pendingPromptCompletion` 保留 runtime，并非在 Stop 时
+无条件释放 ownership；但它会先中断 owner fiber、完成历史收尾，然后才 drain provider。
+正常 Stop 现在等待 prompt 本身，同时启动原有五秒 raw-request drain，不等待 cancel ACK。
+到期时 prompt 若仍 pending，则终止该 session；连接关闭 reject prompt，让 owner 自然进入
+finalizer。终止失败时继续等待 raw ACP 完成，不释放 ownership。Stop 与外部中断后的
+finalizer 共用 runtime 上的一个 drain promise，重复取消不会重置期限或重复终止。
+这恢复了 #571 的恢复策略；此前保留 in-flight owner 的修改误将该策略限制在外部中断路径。
+
+ACP 完成后，既有的 [provider 失败收尾](2026-09-10-context-compaction-terminal-state.md)
+会在释放执行 ownership 前，将尚未结束的压缩持久化为 failed。provider 明确发出的 terminal
+item 更新仍是权威结果。
+
+Session 视图不主动修复数据。本分支新增的 reconciliation capability、两种传输方法、
+renderer retry hook 和 daemon 历史修复均已移除。历史上的未完成记录是独立维护问题；
+隐藏进度或改写历史都不能中断 native 执行。
+
+## 证据与验证
+
+- [固定版本的 native 协议](https://github.com/openai/codex/blob/rust-v0.153.4/codex-rs/app-server/README.md#example-trigger-thread-compaction)。
+- 适配器 owner 位于 `packages/acp-extension-codex/src/CodexAppServerClient.ts`；
+  同目录的 `CodexCommands.ts` 和 `CodexAcpServer.ts` 负责命令及取消路由。
+- 确定性测试覆盖 native start 前后的 Stop/abort、drain 期间拒绝第二个 prompt、
+  中断后继续执行、失败、start ACK 前的 terminal、启动拒绝、无关 turn 完成及进程退出。
+- Lody execution 测试使用真实 `AgentClient` 与受控 ACP transport：cancel ACK 后 signal
+  仍存活、历史未完成、第二次 dispatch 保持 pending，native terminal 后下一次 prompt 才能运行。
+  fake timer 覆盖五秒终止、cancel ACK 不返回、重复 Stop 不重置期限，以及终止失败后继续
+  保留未完成历史和 ownership。外部中断测试仍覆盖 raw 完成、进程终止及终止失败。
+- steer 测试覆盖请求前和异步准备期间的 Stop，验证未投递历史及其 dispatch pointer 保留。
+  受控的接受 ACK 在 Stop 后到达时，source invocation、用户轮次和 dispatch owner 均不改变，
+  source 历史不提前完成、不报告 handled，也不重放 steer。已接受的 steer 在 provider terminal
+  前后均为 `canceled`；terminal 将 source 结算为 cancelled 并释放 owner。steer 终态断言在
+  `ae24a723` 上失败，当时条目会停留在 `pending_apply`。
+- 契约：[会话历史写入](../../../../specs/session-history-writes.zh.md)。
+- PR：[Lody #618](https://github.com/LodyAI/Lody/pull/618)。
+- 适配器实现：[Codex adapter #41](https://github.com/LodyAI/acp-extension-codex/pull/41)。
+
+适配器类型检查及 617 个启用的测试通过；27 个测试由原有环境条件跳过。真实 Codex smoke
+验证手动压缩正常完成，并确认 Stop 后 native `interrupted` 先于 ACP prompt 的 `cancelled`。
+
+根仓库类型、lint、格式、i18n、文档和边界检查通过。Lody execution、dispatch-watcher 和
+AgentClient 三组相关测试共 207 个通过。完整 `pnpm check` 到达 Electron 测试时，103 个通过，
+relay suite 因本地缺少已安装的 Electron 二进制无法加载；此前的 workspace 测试均通过。

apps/cli/src/session/AGENTS.md

@@ -37,8 +37,8 @@ Contract: specs/session-orchestration.md.
 - Never re-dispatch a late-arriving history entry; recovery is a fresh send.
 - `hasPendingUserTurnActivation` is the ONLY pending-turn predicate; never compare those two
   pointers in a consumer.
-- Session metadata is the activation index: never inspect historical Session documents to infer
-  work, and never publish or clear active presence here (`../lib/loro/session-active-presence.ts`).
+- Never inspect historical Session documents to infer work, or publish or clear active presence
+  here (`../lib/loro/session-active-presence.ts`).
 - Keep bootstrap and live reconciliation bounded as README describes; add no per-trigger scan or
   extra throttle.
 
@@ -47,26 +47,27 @@ Contract: specs/session-orchestration.md.
 - Gate turn-scoped history LIST writes on user-entry sync (`turn-history-gate.ts`, 20s);
   never gate status or meta writes.
 - Goals obey [this contract](../../../../specs/session-goal-control.md).
-- Keep `TurnRuntimeState` until raw ACP completion or confirmed termination after cancel; no
-  second visible turn. Assistant ids use `userTurnId`. `invocation` atomically
-  owns source Turn, requester and config; steer replaces it before tools.
+- In-flight Stop cancels ACP, never its owner fiber. Keep `TurnRuntimeState` until raw ACP
+  completion or confirmed termination; no second turn. Assistant ids use `userTurnId`.
+  `invocation` owns source Turn, requester and config atomically; steer replaces it before tools.
 - Publish `latestUserMsgId` in the SAME write as the history append (`appendUserTurn`). Only
   dispatch producers publish it. Renderer sends and queue promotion retain the missing-history
   tombstone; CLI dispatch producers keep their own marker policy.
 - Ordinary turn execution writes only `processingUserMsgId` and `lastHandledUserMsgId`; no start
   or terminal path may read-await-rewrite the other slots.
-- INVARIANT: a steer the agent never accepted must not stay parked in `pending_apply`. Requeue it
-  through the pointer, not the entry status, only for pre-submission rejections or
-  `AgentSteerNotDeliveredError`; skip active or already-handled entries.
+- Never submit steer after Stop. A late accepted ACK cancels that exact steer entry without
+  transferring ownership, changing dispatch pointers or requeueing it.
+  Requeue unaccepted steer via its pointer, not entry status, only before submission or on
+  `AgentSteerNotDeliveredError`; skip active or handled entries.
 - Resume must REOPEN the in-progress assistant entry, clearing
   `finished`/`endedAt`/`permissionWaitMs` there only; never write `finished=false` from teardown.
 - Keep JSON-RPC/transport matching in `acp-error-classification.ts`: disposed/stale `-32603` is
   `agent_disconnected`, Harness compression mismatch is `acp_session_storage_incompatible`.
 - Continue-session recovery may restore the ACP session and retry the same prompt once, only
   while that turn has no ACP output.
-- INVARIANT: a resolved prompt is not proof of success. A turn that emitted no ACP update takes
-  `recordSilentTurnFailure`, not `setDispatchHandled` (read `turnProducedVisibleOutput` before
-  `finalizeTurn` clears it); it still finalizes, still ADVANCES the pointer, and fails open.
+- A turn with no ACP updates takes `recordSilentTurnFailure`, not `setDispatchHandled`.
+  Read `turnProducedVisibleOutput` before `finalizeTurn` clears it; still finalize, advance
+  the pointer, and fail open.
 - Diff content comes only from the CLI-local ACP evidence store; GitHub `diffStats` use PR compare
   semantics, and `session-diff-stats-target.ts` skips rather than overwrites a good total.