fix(dashboard): resolve the steward chat runtime from the channel

apps/presentation/dashboard/src/views/dashboard-page.tsx

@@ -1486,8 +1486,20 @@ function PersonalGoalHome({
   const defaultAgentId = discoveredAgents.find((agent) => agent.label === "Codex" && agent.available)?.agentId
     ?? discoveredAgents.find((agent) => agent.available)?.agentId
     ?? "status-only";
+  // The manager channel answers on the executor this machine declares for the
+  // steward, and the client is expected to send no endpoint until the operator
+  // picks one. The picker therefore has to show that same resolution: an
+  // executor merely discovered on this machine is not a reason to present
+  // itself as the steward's runtime, or the header chip, the composer and the
+  // answer would tell three different stories.
+  const declaredStewardEndpoint = managerChannelBinding?.executor_endpoint?.trim() ?? "";
+  const stewardExecutorAgentId = declaredStewardEndpoint
+    ? discoveredAgents.find((agent) => agent.agentId === declaredStewardEndpoint)?.agentId
+    : undefined;
+  const agentDefaultForContext = (targetContextId: string) =>
+    targetContextId === "manager" ? stewardExecutorAgentId ?? defaultAgentId : defaultAgentId;
   const [selectedAgents, setSelectedAgents] = useState<Record<string, string>>(readPersonalAgentSelections);
-  const selectedAgentId = selectedAgents[contextId] ?? defaultAgentId;
+  const selectedAgentId = selectedAgents[contextId] ?? agentDefaultForContext(contextId);
   const selectedAgent = selectAvailableChatAgent(agentOptions, selectedAgentId, defaultAgentId);
   const [agentMenuOpen, setAgentMenuOpen] = useState(false);
   const [detailsOpen, setDetailsOpen] = useState(false);

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

@@ -474,6 +474,42 @@ Validation: `tests/capabilities/test_steward_executor_machine_defaults.py`,
 `tests/capabilities/test_capability_configuration_ui.py`, and
 `examples/loopx-steward-channel-binding-smoke.py`.
 
+### Steward Answer Identity and Runtime Selection (2026-09-16)
+
+Once a machine could declare its steward executor, the Dashboard was still
+answering two different questions with one value: *who answered me* and *which
+runtime served the turn*. The transcript titled a steward answer with whatever
+the chat-runtime picker happened to hold, and the picker itself preferred a
+`Codex` adapter whenever one was discovered on the machine -- so a channel
+resolving to the managed host could still present itself as an individual CLI
+login, and the header chip, the composer and the answer could each name a
+different executor.
+
+Two rules now hold on the manager channel:
+
+* **Answer identity names the speaker.** The transcript labels a steward answer
+  `LoopX Manager` / `LoopX 管家` on every path that can produce one -- return
+  receipts, resumed history, recovery streaming, the streaming placeholder, the
+  completion fallback, interruption and failure. The executor and its model
+  stay in the machine-capability chip, which is the surface that reports them.
+  Goal channels keep naming the Goal's own Agent.
+* **Runtime selection resolves the way the channel resolves.** The chat-runtime
+  picker follows the channel owner's precedence for the manager context: the
+  declared steward executor first, the shipped default only when the machine
+  declares nothing. A discovered adapter is never presented as the steward.
+
+An explicit operator pick still wins for that context, and the client still
+sends no endpoint when the operator made no pick, so the create-session contract
+in `loopx/chat_server.py` -- "each channel resolves its own default through its
+own owner" -- is unchanged. An explicit pick is also the only path that moves
+this channel off the declared executor, which keeps a discovered CLI from
+silently rewriting a machine decision.
+
+Evidence: `examples/personal-workspace-browser-smoke.mjs` (`execution-chip`
+scenario for the picker and composer resolution, `chat-recovery` scenario for
+the answer identity), plus a live readback on the installed Dashboard where the
+header chip resolved `dsh` while the picker previously reported `Codex`.
+
 ## Steward Team Intake (2026-09-16)
 
 The steward answers questions. Since `2026-09-16` its shipped guidance also

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

@@ -373,6 +373,32 @@ journal 与配额语义；B 作为上游接口出现时的低成本替代；只
 `tests/capabilities/test_capability_configuration_ui.py`，以及
 `examples/loopx-steward-channel-binding-smoke.py`。
 
+### 管家回答身份与 Runtime 选择（2026-09-16）
+
+机器可以声明管家执行器之后，Dashboard 仍在用一个值回答两个不同的问题：**谁在回答我**，
+以及**哪个 runtime 跑了这一回合**。会话记录用聊天 Runtime 选择器当时持有的值给管家回答
+署名；而选择器本身只要在本机发现 `Codex` 适配器就优先选它——于是解析到托管宿主的通道
+仍可能把自己呈现成个人 CLI 登录，页头 chip、输入框与回答可以各说各的执行器。
+
+管家通道上现在有两条规则：
+
+* **回答身份署名说话者。** 会话记录在**所有**能产生管家回答的路径上署名
+  `LoopX 管家` / `LoopX Manager`：交接回执、恢复的历史、恢复流式、流式占位、
+  完成兜底、中断与失败。执行器与模型留在机器能力 chip 上，那才是回报它们的表面。
+  Goal 通道继续署名该 Goal 自己的 Agent。
+* **Runtime 选择按通道的解析方式解析。** 在管家上下文里，聊天 Runtime 选择器遵循通道
+  属主的优先级：先看本机声明的管家执行器，只有本机什么都没声明时才回落到出货默认值。
+  一个被发现的适配器永远不会被呈现成管家。
+
+运维方的显式选择仍然在该上下文里优先；运维方没有选择时，客户端依旧不发送任何端点，
+因此 `loopx/chat_server.py` 的建会话契约——"每个通道通过自己的属主解析自己的默认值"
+——没有变化。显式选择也是这条通道离开已声明执行器的唯一路径，这让一个被发现的 CLI
+不会静默改写一项机器决定。
+
+证据：`examples/personal-workspace-browser-smoke.mjs`（`execution-chip` 场景覆盖选择器与
+输入框的解析，`chat-recovery` 场景覆盖回答身份），以及在已安装 Dashboard 上的一次现场
+回读：页头 chip 解析为 `dsh`，而选择器此前报的是 `Codex`。
+
 ## 管家团队入端口径（2026-09-16）
 
 管家负责回答问题。自 2026-09-16 起，它的出厂指引里还带了一段有界流程，用于另一类请求：

examples/personal-workspace-browser/execution-chip.mjs

@@ -196,6 +196,62 @@ export const executionChipScenario = {
       coverageEntries.push(...await managed.close());
     }
 
+    // The chat-runtime picker has to report the same resolution as the chip: a
+    // managed host declared for this machine is what will answer, so neither
+    // the header trigger nor the composer may keep advertising the CLI the
+    // machine merely happens to have installed.
+    const managedHostAdapters = [{
+      adapter_kind: "deepseek_harness_segment",
+      agent_id: "dsh",
+      available: true,
+      display_name: "DeepSeek Harness (managed)",
+      interrupt: false,
+      resume: true,
+      streaming: false,
+    }];
+    const stewardPicker = await openWorkspacePage(browser, url, {
+      apiOptions: { managerChannelBinding: selectedManagedBinding, runtimeAgents: managedHostAdapters },
+      collectCoverage,
+    });
+    try {
+      const pickerLabel = (await stewardPicker.page
+        .locator("div.personal-agent-select button.personal-select-trigger")
+        .innerText()).replace(/\s+/g, " ").trim();
+      if (!pickerLabel.includes("DeepSeek Harness (managed)")) {
+        throw new Error(`Chat runtime picker ignored the declared steward executor: ${pickerLabel}`);
+      }
+      if (pickerLabel.includes("Codex")) {
+        throw new Error(`Chat runtime picker advertised a discovered CLI as the steward: ${pickerLabel}`);
+      }
+      const composerLabel = (await stewardPicker.page
+        .locator(".personal-channel-composer > span")
+        .first()
+        .innerText()).trim();
+      if (composerLabel !== "DeepSeek Harness (managed)") {
+        throw new Error(`Composer named ${composerLabel} instead of the steward's declared executor`);
+      }
+    } finally {
+      coverageEntries.push(...await stewardPicker.close());
+    }
+
+    // A machine that declares no steward executor keeps the shipped default,
+    // so the parity above is a resolution the control plane asked for and not
+    // a new hard-coded preference.
+    const undeclaredSteward = await openWorkspacePage(browser, url, {
+      apiOptions: { runtimeAgents: managedHostAdapters },
+      collectCoverage,
+    });
+    try {
+      const pickerLabel = (await undeclaredSteward.page
+        .locator("div.personal-agent-select button.personal-select-trigger")
+        .innerText()).replace(/\s+/g, " ").trim();
+      if (pickerLabel !== "Chat Codex") {
+        throw new Error(`An undeclared steward executor no longer used the shipped default: ${pickerLabel}`);
+      }
+    } finally {
+      coverageEntries.push(...await undeclaredSteward.close());
+    }
+
     // A managed host that cannot launch here names the missing fact, and never
     // claims the channel lacks a transport it now has.
     for (const [binding, expected] of [
@@ -306,7 +362,7 @@ export const executionChipScenario = {
     }
     return {
       coverageEntries,
-      note: "execution chip reports the selected executor, the credential it is billed to and the resolved model, names the steward channel's shipped default without claiming a credential branch, keeps a configured credential from moving it, names why a selected managed host cannot launch, and stays absent without a binding",
+      note: "execution chip reports the selected executor, the credential it is billed to and the resolved model, names the steward channel's shipped default without claiming a credential branch, keeps a configured credential from moving it, keeps the chat-runtime picker on the executor the machine declares for the steward, names why a selected managed host cannot launch, and stays absent without a binding",
     };
   },
 };

examples/personal-workspace-browser/fixture.mjs

@@ -288,7 +288,7 @@ function filterStatusFixtureToScope(fixture, statusGeneration, scope) {
   }
 }
 
-export async function installApi(page, { goalSubagentConfigurationEnabled = true, initialActionProposals = [], managerChannelBinding = null } = {}) {
+export async function installApi(page, { goalSubagentConfigurationEnabled = true, initialActionProposals = [], managerChannelBinding = null, runtimeAgents = null } = {}) {
   let turnCounter = 0;
   const runtime = page.__loopxRuntime ??= { actionProposals: new Map(), goalSubagentConfigurations: new Map(), larkConnections: [], messages: new Map(), sessions: new Map(), turnMessages: new Map() };
   const actionProposals = runtime.actionProposals;
@@ -1251,7 +1251,7 @@ export async function installApi(page, { goalSubagentConfigurationEnabled = true
           { agent_id: "codex", display_name: "Codex", adapter_kind: "codex_app_server", available: true, streaming: true, resume: true, interrupt: true },
           { agent_id: "claude-code", display_name: "Claude Code", adapter_kind: "claude_code_cli", available: true, streaming: true, resume: true, interrupt: true },
           { agent_id: "offline-agent", display_name: "Offline Agent", adapter_kind: "acp", available: false, streaming: false, resume: false, interrupt: false },
-        ],
+        ].concat(runtimeAgents ?? []),
       }, status: 200 });
       return;
     }

examples/personal-workspace-browser/typed-actions.mjs

@@ -1611,7 +1611,7 @@ export const typedActionsScenario = {
       await page.keyboard.press("Escape");
       if (await agentListbox.isVisible().catch(() => false)) throw new Error("Agent menu did not close on Escape");
       if (!(await agentSelect.evaluate((element) => element === document.activeElement))) throw new Error("Agent menu did not restore trigger focus");
-      pass(14, "Codex remained the healthy default and the unavailable Agent option was disabled with explanation.");
+      pass(14, "With no declared steward executor the shipped Codex default held, and the unavailable Agent option was disabled with explanation.");
       await agentSelect.click();
       const reopenedAgentListbox = page.getByRole("listbox", { name: "选择聊天 Runtime" });
       await reopenedAgentListbox.getByRole("option", { name: "Claude Code", exact: true }).click();