feat: enable durable local-led and nested mixed-agent teams

.github/workflows/ark-turn.yml

@@ -0,0 +1,52 @@
+name: Optional Ark Turn
+
+on:
+  pull_request:
+    paths:
+      - ".github/workflows/ark-turn.yml"
+      - "packages/loopx-ark-turn/**"
+      - "loopx/control_plane/turn_driver/**"
+      - "loopx/control_plane/collaboration/**"
+      - "loopx/collaboration_mcp.py"
+      - "loopx/dsh_goal_mode/**"
+      - "loopx/cli_commands/turn*.py"
+      - "loopx/control_plane/goals/acceptance*.ts"
+      - "loopx/control_plane/goals/acceptance.py"
+      - "examples/managed-research-team/**"
+      - "pyproject.toml"
+      - "tests/test_local_delegation.py"
+      - "tests/test_managed_research_scenario.py"
+      - "tests/test_managed_research_team.py"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  adapter-contract:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    strategy:
+      matrix:
+        python: ["3.11", "3.13"]
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-node@v6
+        with:
+          node-version: "24.21.0"
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ matrix.python }}
+      - run: python -m pip install -e ".[test]" -e packages/loopx-ark-turn
+      - name: Public SDK, real stdio MCP, negative cases and DSH parity
+        run: >-
+          python -m pytest -q packages/loopx-ark-turn/tests
+          tests/test_dsh_goal_mode.py tests/test_turn_managed_executor_binding.py
+          tests/test_ark_managed_agent_host.py
+          tests/test_loopx_turn_driver.py tests/test_collaboration_mcp.py
+          tests/test_auto_research_artifact_receipt.py
+          tests/test_worker_command_validation.py tests/test_workspace_story_demo.py
+          tests/test_local_delegation.py tests/test_managed_research_scenario.py
+          tests/test_managed_research_team.py
+      - name: Lint optional package and example
+        run: python -m ruff check packages/loopx-ark-turn examples/managed-research-team

docs/architecture/rfcs/agent-session-execution-modes-v0.md

@@ -269,6 +269,47 @@ This is a delivery priority, not a default migration. The
 Reuse the existing profile editor and session projections; do not add a
 manager-only creation service, task ledger or scheduling loop.
 
+### Model selection within Agent creation and attachment
+
+This is a proposed refinement of the reusable Agent operations above, not a
+shipped model-catalog API. Model selection belongs to an Agent's execution
+profile and binding, independently of whether that Agent coordinates others.
+Reuse the existing managed execution profile, subagent launch preferences and
+profile editor; the steward's machine defaults are one caller's defaults, not
+the universal configuration owner. A model change does not create a new logical
+Agent or grant permission to launch one.
+
+| Step | Owner and required observation |
+| --- | --- |
+| Discover choices | The selected executor/provider adapter reports model IDs, supported parameters, capability limits, discovery scope and freshness. Keep provider catalog presence, managed-host compatibility and account authorization separate; unknown or failed discovery is not an empty supported list. |
+| Request and resolve a profile | The shared typed TS boundary validates caller scope, allowed profiles, budget constraints and explicit configuration precedence. Retain the requested model and parameters separately from resolved values and their sources. SDK/network discovery remains in the adapter; do not copy selection or admission rules into each Python launcher. |
+| Create or attach | Creation uses the resolved profile through the selected adapter. Attachment observes the existing host's actual profile; it cannot silently change its model, start a replacement executor or claim that a requested preference already took effect. Repeated creation reuses the existing identity/binding contract. |
+| Start and read back | Bind the profile revision to the execution generation and read back the provider-reported model and effective parameters. A visible catalog row or successful Agent-definition creation does not establish that an inference session can run. A mismatch or unsupported option produces an actionable failure, never an implicit model fallback. |
+
+Parameter support is provider-specific: the same reasoning-effort label need
+not have the same meaning across hosts, and speed, thinking mode, context limits
+and tool support are not universal knobs. Use a small common selection contract
+with validated provider-owned options rather than a core list of vendor models
+or one global parameter enum. Credentials remain in the selected provider's
+credential scope and never enter an Agent profile or public projection.
+
+Mutable model aliases require explicit readback. Record a resolved version only
+when the provider exposes it; otherwise record that the backing version is
+unknown rather than treating the alias as a reproducible snapshot. Profile
+changes use the existing binding revision/generation and rebind boundary;
+running work retains its admitted profile until a qualified transition occurs.
+Changing the parent profile does not silently change existing children. An
+authorized child coordinator may choose only within its inherited profile and
+budget scope, using the same operation as the lead.
+
+The next implementation slice must connect discovery, selection, creation or
+attachment, launch and readback for both a local and a cloud executor. Qualify
+unsupported parameters, stale discovery, unavailable authorization, retry,
+mutable aliases, and profile changes during active work. Reuse the existing
+CLI, frontend and Lark configuration owners/projections where affected; a
+backend field alone does not complete that user journey. Existing defaults and
+explicit host choices remain unchanged until a disclosed implementation lands.
+
 ### State model and schema
 
 The binding is the unit of mode ownership. Its canonical fields:
@@ -517,6 +558,16 @@ Preserve feature-off behavior for every existing profile and entrypoint.
 
 DSH steward Chat is currently single-segment, read-only and without cross-turn host sessions; `turn run-once` is a separate bounded execution path. The next slice proves successor wake, cancellation/stop, crash recovery and returning stale-executor fences with packaged frontend/CLI/Lark readback. An executor name, one segment or multiple registrations cannot establish continuous managed execution. Disconnection never switches attached hosts to managed, and unqualified hosts retain their existing boundary.
 
+The opt-in [local delegation interface](../../reference/local-delegation.md)
+now provides durable operations around bounded Turns, including member-to-member
+launch grants and TS task acceptance. Its Ark process-loss drill resumes the
+original Session/input after cloud tool waiting; it does not resend acknowledged
+effects or reset the deadline. This qualifies local execution recovery, not
+successor wake, attached-host takeover or full fleet cancellation. Provider file
+profiles preserve the existing model/tool configuration boundary; the general
+Agent creation/model discovery proposal above remains separate.
+
+
 ## 12. Normative delivery plan
 
 | Milestone | Shipped behavior | Entry gate | Exit evidence | Rollback |

docs/architecture/rfcs/agent-session-execution-modes-v0.zh-CN.md

@@ -216,6 +216,37 @@ claim 与完成回执，以及"只有经过验证的回写才推进工作"这一
 [harness 选型 RFC](harness-selection-dsh-pi-v0.zh-CN.md)负责。复用现有 profile editor
 和会话投影，不新建管家专属创建服务、任务账本或调度循环。
 
+### Agent 创建与接入中的模型选择
+
+这是对上述可复用 Agent 操作的设计细化，尚未交付通用模型目录 API。模型选择属于
+Agent 的 execution profile 与绑定，不取决于它是否担任协调员。复用现有 managed
+execution profile、子 Agent 启动偏好和 profile editor；管家的机器默认值只是一个
+调用入口的默认配置，不是通用配置 owner。换模型不产生新的逻辑 Agent，也不授予
+创建或启动 Agent 的权限。
+
+| 步骤 | Owner 与必须读回的事实 |
+| --- | --- |
+| 发现候选 | 所选 executor/provider adapter 返回模型 ID、支持参数、能力限制、发现范围与新鲜度。分别记录 provider 目录存在、managed host 兼容和账号授权；未知或发现失败不能冒充空的支持列表。 |
+| 请求与解析配置 | 共享 typed TS 边界校验调用者范围、允许的 profile、预算约束和显式配置优先级。保留请求模型/参数与解析值及来源的区别。SDK/网络发现留在 adapter，不让每个 Python launcher 复制选择或准入规则。 |
+| 创建或接入 | 创建经所选 adapter 使用已解析配置；接入只观察既有宿主实际配置，不能静默改模型、启动替代执行器，或把请求偏好显示为已生效。重复创建复用现有身份与绑定合同。 |
+| 启动并读回 | 配置 revision 绑定执行 generation，读回 provider 报告的模型与实际参数。目录可见或 Agent 定义创建成功，都不能证明推理会话可运行。配置不符或参数不支持必须给出可操作错误，不隐式换模型。 |
+
+参数支持由 provider 决定：相同 reasoning-effort 标签在不同宿主中未必同义，speed、
+thinking mode、上下文限制和工具支持也不是通用旋钮。采用小型公共选择合同与经校验的
+provider 参数，不在核心维护厂商模型名单或一个全局参数枚举。凭据留在所选 provider
+的凭据作用域中，不进入 Agent profile 或公开投影。
+
+动态模型别名需要显式读回。只有 provider 暴露解析版本时才记录具体版本；否则标明
+底层版本未知，不能把别名当作可复现快照。配置变更沿现有 binding revision/generation
+与 rebind 边界生效；运行中工作保留已准入的配置，直到经过已验证的迁移。父 Agent
+配置变更不静默修改已有子 Agent。获授权的子协调员仅能在继承的 profile 与预算范围
+内选择，使用和主 Agent 相同的操作。
+
+下一实现切片须为一个本地和一个云端执行器打通发现、选择、创建或接入、启动及读回。
+验证不支持的参数、过期目录、授权不可用、重试、动态别名和运行中配置变更。受影响的
+CLI、前端和 Lark 复用现有配置 owner/投影；单有后端字段不代表用户路径完成。在明确
+披露的实现交付之前，既有默认值和显式宿主选择保持不变。
+
 ### 状态模型与 schema
 
 绑定是模式归属的单元。其规范字段：
@@ -416,6 +447,13 @@ worker 请求并采用另一 worker 的产物；driver 切换竞态拒绝旧执
 
 目前 DSH 管家 Chat 是单段、只读、无跨 turn 宿主会话；`turn run-once` 是另一条有界执行路径。下一切片要证明 successor wake、取消/停止、崩溃恢复及旧执行器返回 fence，经 packaged frontend/CLI/Lark 回读真实状态。不能仅增加一个 executor 名称、启动一个片段或绑定若干 Agent 就声称持续 managed 模式完成。attached host 不因掉线而改为 managed，未验收宿主保持原资格边界。
 
+显式启用的[本地委派接口](../../reference/local-delegation.md)已为有界 Turn 提供持久
+操作，涵盖成员继续委派的授权及 TS 任务验收。Ark 进程中断实验在云端等待本地工具后，
+以原 Session/输入接回，不重发已确认副作用、不重置期限。这验证本地执行恢复，不代表
+successor wake、attached 接管或完整团队取消。Provider 文件配置保留现有模型/工具
+边界；上文通用 Agent 创建与模型发现提案仍是独立后续范围。
+
+
 ## 12. 规范性交付计划
 
 | 里程碑 | 交付行为 | 进入门槛 | 退出证据 | 回滚 |

docs/architecture/rfcs/harness-selection-dsh-pi-v0.md

@@ -70,6 +70,35 @@ dated 2026-09-15 and is written to land with the managed stack:
 | L1 event source and session-owning runtime candidate | DSH | opt-in, not promoted; the bounded Turn host role is the default row above | the C0, C1, overhead, retention and Mode B rows in this document being run and reviewed |
 | Optional visible host loop | Pi | not a managed runtime | declare a per-binding session mode with readback, prove single-executor behavior under restart, "conversation is not a receipt", non-authoritative host-local state, and one real-host restart row |
 
+### Optional Ark governed Turn profile
+
+[`loopx-ark-turn`](../../../packages/loopx-ark-turn/README.md) is a separately
+installed provider selected explicitly through `--host generic-cli` with fresh
+iteration context. It shares DSH's signed request/candidate conversion; LoopX
+still owns admission, independent validation, work writeback and quota. A
+per-Turn stdio MCP process exposes only operator-selected tools with bound work
+identity. Provider model usage and resource-cleanup receipts are observations,
+not accepted-work quota or a second task lifecycle.
+
+This profile does not change the default host or the native Ark `goal_once`
+profile. Native Goal continuation and outer LoopX Turn continuation must not
+drive the same binding. The [research composition example](../../../examples/managed-research-team/README.md)
+exercises a managed coordinator delegating to local workers; it does not promote
+a persistent steward Chat transport, recursive fleet supervision, full live
+steering, or a shared authority service. Use its explicit setup/readback/cleanup
+instructions and preserve failed versus untested qualification boundaries.
+
+The example's integrated acceptance path uses five preauthorized canonical
+tasks and startup-only owner configuration. Both hosts select exact work via
+`turn --todo-id`; fresh TS Todo completion precedes accepted result return.
+Synthesis checks current child completion, binding and artifact hashes.
+Provider cleanup, Turn progress and canonical completion remain separate.
+This does not supply dynamic work derivation or another Python lifecycle owner.
+Its default profile uses a local DSH lead with two DSH and two Ark members;
+the cloud reviewer consumes a completed local analysis before returning its own
+result. A secondary cloud-led profile tests the inverse delegation direction.
+Both reuse the same Turn host adapters; neither changes the steward default.
+
 ### Managed host binding and live qualification (2026-09-15)
 
 A managed host binding names four things: the host adapter, the provider, the

docs/architecture/rfcs/harness-selection-dsh-pi-v0.zh-CN.md

@@ -58,6 +58,27 @@ C1、开销、保留与 Mode B 各行。
 | L1 事件源与会话归属 runtime 候选 | DSH | opt-in，未晋级；有界 Turn 宿主角色见上一行默认值 | 本文 C0、C1、开销、保留与 Mode B 各行被真实执行并通过评审 |
 | 可选的可见宿主循环 | Pi | 不是 managed runtime | 先声明按绑定持久化且可回读的会话模式，证明重启下的单执行器行为、"对话不是回执"、宿主本地状态非权威，并提供一条真实宿主重启行 |
 
+### 可选 Ark 受控 Turn 档位
+
+[`loopx-ark-turn`](../../../packages/loopx-ark-turn/README.md)单独安装，通过
+`--host generic-cli` 显式选择并使用 fresh iteration context。它复用 dsh 的签名
+请求/候选转换；admission、独立验收、工作写回及 quota 仍由 LoopX 拥有。每次 Turn
+绑定一个 stdio MCP 进程，只暴露 operator 选择的工具和绑定的工作身份。Provider
+模型用量、资源清理回执是观测，不是已验收工作 quota 或第二份任务生命周期。
+
+本档位不改变默认宿主，也不改变原有 Ark `goal_once` 档位。原生 Goal 自驱和外层
+LoopX Turn 驱动不能同时驱动同一绑定。[投研组合示例](../../../examples/managed-research-team/README.md)
+验证 managed 协调员委派本地 worker，不据此晋升持久管家 Chat、递归团队监督、完整
+实时 steer 或共享权威服务。按示例显式配置、回读和清理，保留失败与未验证的区别。
+
+示例已集成五个预授权 canonical 任务和启动时一次性 owner 配置。两类宿主通过
+`turn --todo-id` 选择精确工作；TS Todo 新鲜验收完成后才返回 accepted 结果。
+综合任务检查子任务当前完成状态、绑定和产物哈希。Provider 清理、Turn 进展和
+canonical 完成仍是不同事实；本切片不提供动态派生授权或另一份 Python 生命周期。
+默认示例由本地 DSH 协调员组织两个 DSH 与两个 Ark 成员；云端核验员消费已完成的
+本地分析，再返回自己的产物。辅助云端协调档位验证反向委派。两者复用相同 Turn
+适配器，不改变管家默认执行器。
+
 ### 托管宿主绑定与真实环境验证（2026-09-15）
 
 一个托管宿主绑定要说明四件事：宿主适配器、provider、模型，以及凭据来自哪里。

docs/architecture/rfcs/loopx-overall-roadmap-v0.md

@@ -234,6 +234,34 @@ These priorities do not change live Goal quota or authorize experiments/cloud re
 - **Exit:** 2–3 workers, one dependency, one failure and one direction correction; Agents select and revise delegation without manual phase input or result forwarding. Inspect through packaged frontend and independent CLI readback. An authorized Lark entry reads the corresponding audience-visible feedback. Untested Lark remains explicitly unqualified.
 - **Rollback:** stop new admission, drain accepted work and retain bindings/receipts; attached fallback cannot be used to simulate availability.
 
+**Optional mixed-team Turn slice.** The [Ark adapter](../../../packages/loopx-ark-turn/README.md)
+uses the existing generic-cli Turn boundary alongside DSH. The
+[shared local delegation interface](../../reference/local-delegation.md) now
+composes semantic peer requests/adoption/return, explicit operator execution
+bindings and the merged TS acceptance owner. It replaces demo-owned delegation;
+coordinators and ordinary members use the same grant contract. Disabled stdio
+servers retain their original five non-executing tools. Configuration files
+compact provider launch arguments without changing default executor selection.
+
+The [synthetic research example](../../../examples/managed-research-team/README.md)
+uses a local lead, two DSH members and two Ark members. One cloud reviewer adopts
+local analysis; another Ark member delegates to DSH before returning to local
+synthesis. Five stable preauthorized tasks bind exact criteria once. Turn
+validation and ordinary Todo completion independently execute current pinned
+checks; accepted returns read canonical completion and exact artifacts. All
+business questions/order remain model decisions; the Goal stays active.
+
+Durable operation ids and existing Turn journals recover results after a source
+conversation disappears. A real process-group interruption after Ark input ACK
+has been resumed on the original Session/input to canonical completion; cloud
+waiting for a local tool was observed, and owned resources were cleaned.
+Uncertain creation/input acknowledgements or tool effects remain reconciliation
+cases. The adapter retains the original deadline and does not resend work.
+This is a local trusted-host foundation, not G1/G3 completion: attached persistent
+sessions, generic Agent creation, dynamic governed work derivation, complete
+inbox/queue/steer, authenticated remote authority and packaged frontend/Lark
+companion work remain R2/R3/R4/R6 boundaries. Existing Goals are not promoted.
+
 ### R3: Semantic Requests and Automatic Return
 
 - **Owner:** manager RFC M2/M3; migrate existing `manager_context` request/tracking/return into one typed collaboration transaction, incorporating the #4094 adapter.