feat(core): Durable Managed Session journal and failover

.github/workflows/.size-baseline

@@ -55,7 +55,7 @@
 35955 release.yml
 44285 repo-hygiene.yml
 1079 scorecard-monthly.yml
-12011 sdk-java.yml
+16542 sdk-java.yml
 3886 sdk-python.yml
 5595 security-checks.yml
 6777 serve-ab-publish.yml

.github/workflows/sdk-java.yml

@@ -8,6 +8,7 @@ on:
     paths:
       - 'packages/sdk-java/**'
       - 'packages/core/src/config/approval-modes.json'
+      - 'packages/core/src/managed-runtime/contracts/managed-session-store-v1.fixtures.json'
       - 'packages/cli/src/commands/serve.ts'
       - 'packages/cli/src/serve/**'
       - 'packages/cli/src/acp-integration/**'
@@ -27,6 +28,7 @@ on:
     paths:
       - 'packages/sdk-java/**'
       - 'packages/core/src/config/approval-modes.json'
+      - 'packages/core/src/managed-runtime/contracts/managed-session-store-v1.fixtures.json'
       - 'packages/cli/src/commands/serve.ts'
       - 'packages/cli/src/serve/**'
       - 'packages/cli/src/acp-integration/**'
@@ -185,7 +187,7 @@ jobs:
         run: 'mvn --batch-mode --no-transfer-progress -DskipTests -Dgpg.skip=true package'
 
   mysql-integration:
-    name: 'Runtime Broker MariaDB / Java 21'
+    name: 'Runtime Broker and Managed Agent MariaDB / Java 21'
     runs-on: 'ubuntu-latest'
     timeout-minutes: 15
     services:
@@ -211,7 +213,10 @@ jobs:
           distribution: 'temurin'
           java-version: '21'
           cache: 'maven'
-          cache-dependency-path: 'packages/sdk-java/runtime-broker/pom.xml'
+          cache-dependency-path: |-
+            packages/sdk-java/qwencode/pom.xml
+            packages/sdk-java/runtime-broker/pom.xml
+            packages/sdk-java/managed-agent-server/pom.xml
           settings-path: '${{ runner.temp }}/setup-java-m2'
 
       - name: 'Run Runtime Broker MySQL integration tests'
@@ -223,6 +228,22 @@ jobs:
           -Dmysql.url='jdbc:mysql://127.0.0.1:3306/runtime_broker_test?allowPublicKeyRetrieval=true&useSSL=false'
           -Dmysql.user=root -Dmysql.password=runtime-broker clean verify
 
+      - name: 'Install Managed Agent dependencies'
+        env:
+          MAVEN_ARGS: '--settings ${{ runner.temp }}/setup-java-m2/settings.xml --toolchains ${{ runner.temp }}/setup-java-m2/toolchains.xml'
+        run: |-
+          mvn --batch-mode --no-transfer-progress -f packages/sdk-java/qwencode/pom.xml -DskipTests -Dgpg.skip=true -Dmaven.javadoc.skip=true install
+          mvn --batch-mode --no-transfer-progress -f packages/sdk-java/runtime-broker/pom.xml -DskipTests install
+
+      - name: 'Run Managed Agent tests, Checkstyle, and MySQL integration'
+        working-directory: 'packages/sdk-java/managed-agent-server'
+        env:
+          MAVEN_ARGS: '--settings ${{ runner.temp }}/setup-java-m2/settings.xml --toolchains ${{ runner.temp }}/setup-java-m2/toolchains.xml'
+        run: >-
+          mvn --batch-mode --no-transfer-progress -Pmysql-integration
+          -Dmysql.url='jdbc:mysql://127.0.0.1:3306/managed_agent_test?createDatabaseIfNotExist=true&allowPublicKeyRetrieval=true&useSSL=false'
+          -Dmysql.user=root -Dmysql.password=runtime-broker clean verify checkstyle:check
+
   daemon-e2e:
     name: 'Real daemon E2E / Java 11'
     runs-on: '${{ (github.repository == ''QwenLM/qwen-code'' && vars.MAINTAINER_ECS_RUNNER_DISABLED != ''true'' && (github.event_name == ''push'' || github.event.pull_request.head.repo.full_name == github.repository || contains(fromJSON(''["OWNER","MEMBER","COLLABORATOR"]''), github.event.pull_request.author_association))) && fromJSON(''["self-hosted", "linux", "x64", "ecs-qwen"]'') || ''ubuntu-latest'' }}'

docs/design/2026-09-15-task-session-startup-config.md

@@ -0,0 +1,92 @@
+# Model and reasoning selection at session creation
+
+[English](2026-09-15-task-session-startup-config.md) | [简体中文](2026-09-15-task-session-startup-config.zh-CN.md)
+
+Tracking: [#11944](https://github.com/QwenLM/qwen-code/issues/11944), independent PR 4.
+Status: implemented on this branch; not yet merged. Baseline: `39b22efb6f` (2026-09-20).
+
+## Goal
+
+An external caller can start a new session with the same model and reasoning
+selection as a user selecting them in the composer before sending a prompt.
+Creation applies the selection only to that session, without saving shared
+defaults. Subsequent prompts, manual changes, reload, fallback, skills and
+restoration retain ordinary session behavior. The selection is not automatically
+cleared after the first turn.
+
+## Contract
+
+Ordinary and standalone creation accept optional
+`startupConfig: { modelServiceId, reasoningEffort? }`. `modelServiceId` is required;
+`reasoningEffort` is optional. Omission selects only the model, without calling the
+reasoning setter or substituting `default`, so models without reasoning controls
+are supported. Explicit reasoning uses the existing `ReasoningSelection` values, including
+`default` and `none`, with the composer's existing validation and application
+semantics. Unknown properties, a simultaneous legacy `modelServiceId`, and explicit
+`sessionScope: 'single'` are rejected. Ordinary startup configuration implies
+`sessionScope: 'thread'`. Omission preserves the legacy API unchanged.
+
+The daemon advertises `session_startup_config`. The TypeScript SDK checks this
+capability and supports REST and daemon ACP HTTP/WS mappings. Direct child ACP,
+load/resume and the `create_sub_session` tool schema are outside this PR.
+
+Success returns `modelApplied: true` and `startupConfigApplied`, describing the
+canonical current model selector, requested reasoning selection and effective
+reasoning state when reasoning was explicitly requested. Model-only requests
+return only the confirmed model selector in `startupConfigApplied`; they do not
+claim to have applied a reasoning choice. This acknowledges preparation, not an
+immutable lifetime policy.
+Malformed requests are rejected before workspace mutations. Invalid selections
+fail creation rather than running the initial prompt with a different selection.
+Existing authentication, ownership, timeout and uncertain-outcome errors retain
+their semantics.
+
+## Implementation
+
+The current UI prepares an ordinary session by creating and attaching it, setting
+the model, setting reasoning, and then submitting the prompt. Reuse these existing
+setters with model persistence disabled and reasoning `persist: false`.
+
+Ordinary bridge creation applies and confirms both selections before returning.
+Standalone preparation happens after managed binding commit (which authenticates
+the deferred workspace) and before release or initial-prompt admission. A failed
+preparation follows existing owned-session cleanup without killing siblings in
+the shared child. A definite standalone selection rejection removes the owned
+recording so the same id can be retried, with best-effort empty-directory cleanup.
+Timeouts and channel failures retain the existing uncertain-outcome recovery;
+cleanup whose outcome cannot be confirmed enters existing containment. SDK
+confirmation mismatch after a readable success response is reported directly,
+without recovery/adoption.
+Startup selection emits session state changes, not a shared
+`settings_changed` event.
+
+Use the existing model route parser and config-option result for confirmation.
+Do not add a second provider reasoning model. All daemon routes continue to use
+their resolved workspace runtime and owner.
+
+## Scope decisions
+
+This supersedes the earlier lifetime-isolation proposal attached to issue #11944.
+There is no `startupPinned`, persistent policy record, recording schema change,
+strict restore protocol, fallback restriction, mutation gate or UI modification.
+There is no promise to select the route before `Config.initialize` or initial
+authentication: preparation follows the existing UI ordering. Shared defaults
+remain unchanged by this startup operation; later manual operations retain their
+existing persistence semantics. The other three PRs are not dependencies.
+
+## Acceptance
+
+- Compare actual provider requests from composer-equivalent preparation and the
+  new creation API, including tool-loop continuation requests.
+- Cover model-only creation on models without reasoning controls, explicit tiers,
+  `default`, `none`, unsupported choices and shared `none`
+  with requested `high`.
+- Confirm standalone commit → selection → release → prompt ordering; preparation
+  failures admit zero prompts.
+- Verify unchanged user/workspace settings and siblings, no default-change event,
+  ordinary second-turn/manual-change behavior and legacy creation compatibility.
+- Cover unsupported daemons, invalid requests, workspace ownership, disconnects
+  and cleanup failures.
+- Run build, typecheck, bundle, focused package tests and E2E; distinguish
+  controlled-provider evidence from real external-provider validation. Finish two
+  consecutive clean full-diff self-audit passes.

docs/design/2026-09-15-task-session-startup-config.zh-CN.md

@@ -0,0 +1,60 @@
+# 新建会话时选择模型与思考强度
+
+[English](2026-09-15-task-session-startup-config.md) | [简体中文](2026-09-15-task-session-startup-config.zh-CN.md)
+
+跟踪：[#11944](https://github.com/QwenLM/qwen-code/issues/11944)，独立 PR 4。
+状态：已在本分支实现，尚未合并。基线：`39b22efb6f`（2026-09-20）。
+
+## 目标
+
+外部调用方可以指定模型和思考强度发起新会话，效果等同于用户在输入框中选择后发送提问。
+创建操作只设置本会话，不保存共享默认值。后续提问、人工切换、reload、fallback、skill 和恢复沿用普通会话行为。
+首轮结束后不会自动清除选择。
+
+## 契约
+
+普通及 standalone 创建接受可选的 `startupConfig: { modelServiceId, reasoningEffort? }`。
+`modelServiceId` 必填，`reasoningEffort` 可选。省略时只选择模型，不调用思考强度 setter，也不自动补 `default`，因此支持没有思考强度控件的模型。显式思考强度复用现有 `ReasoningSelection`，包括 `default`、`none`，
+校验和应用语义与输入框一致。拒绝未知字段、同时提供旧的顶层 `modelServiceId`、显式
+`sessionScope: 'single'`。普通会话的启动配置隐含 `sessionScope: 'thread'`。省略对象完全保留旧 API。
+
+daemon 声明 `session_startup_config` 能力。TypeScript SDK 检查能力并支持 REST 与 daemon ACP HTTP/WS 映射。
+直接子进程 ACP、load/resume 和 `create_sub_session` 工具 schema 不在本 PR 范围内。
+
+成功返回 `modelApplied: true` 和 `startupConfigApplied`，描述当前规范模型选择符；仅显式请求思考强度时才包含请求档位及实际思考状态。
+仅指定模型时，`startupConfigApplied` 只返回已确认的模型选择符，不声称应用过思考强度选择。
+这表示启动准备完成，不代表整个生命周期内不可变。结构错误在任何工作区变更之前拒绝。
+不合法的选择使创建失败，不允许初始提问使用不同配置执行。
+认证、归属、超时与结果不确定错误保留既有语义。
+
+## 实现
+
+当前 UI 先创建并 attach 普通会话，再设置模型、设置思考强度，最后发送 prompt。
+复用这些已有 setter，关闭模型持久化并使用 reasoning `persist: false`。
+
+普通 bridge 创建在返回之前应用并确认组合。standalone 在 managed binding commit 完成延迟认证后、
+release 和首条 prompt 放行前准备组合。准备失败沿用已有的本次会话清理，不关闭共享子进程里的其他会话。
+确定的 standalone 选择拒绝会删除本次持久记录，使同一 ID 可重试，并尽力清理空输出目录。
+超时和断连保持既有的结果不确定恢复语义；清理无法确认时进入原有隔离流程。
+SDK 收到可读的成功响应但确认不匹配时直接报错，不触发恢复或采用该会话。
+启动选择只发送会话状态变更，不发送共享 `settings_changed` 事件。
+
+复用模型路由解析和 config-option 返回值进行确认，不引入第二套 provider reasoning 模型。
+所有 daemon 路由继续使用已解析的 workspace runtime 和 owner。
+
+## 范围取舍
+
+本文替代 issue #11944 中先前的生命周期隔离提案。
+不引入 `startupPinned`、持久策略记录、recording schema 变更、严格恢复协议、fallback 限制、变更门禁或 UI 修改。
+不承诺在 `Config.initialize` 或初始认证前选择路由；准备顺序与现有 UI 相同。
+本次启动操作不修改共享默认设置；之后人工操作保持现有持久化语义。其他三个 PR 不是依赖。
+
+## 验收
+
+- 比较输入框等效准备流程与新创建 API 的实际 provider 请求，包含工具循环的后续请求。
+- 覆盖不支持思考强度控件的模型仅指定模型创建、显式档位、`default`、`none`、不支持的选择，以及共享为 `none`、请求为 `high`。
+- 验证 standalone commit → 配置 → release → prompt 顺序；准备失败时 prompt 次数为零。
+- 验证用户/工作区设置和兄弟会话不变、不产生默认值变更事件、第二轮和人工修改沿用现状、legacy 创建兼容。
+- 覆盖旧 daemon、非法请求、工作区归属、断连和清理失败。
+- 运行 build、typecheck、bundle、聚焦包测试及 E2E；区分受控 provider 与真实外部 provider 验证证据。
+  完成两轮连续干净的完整 diff 自审。