diff --git a/apps/desktop/electron/main/services/session-collaboration.ts b/apps/desktop/electron/main/services/session-collaboration.ts
index 44038ebc6..6595c48d7 100644
--- a/apps/desktop/electron/main/services/session-collaboration.ts
+++ b/apps/desktop/electron/main/services/session-collaboration.ts
@@ -287,6 +287,14 @@ export function createSessionCollaborationService(deps: SessionCollaborationDepe
             const chosen = requested ? models.find((item) => item.key === requested)
               : models.find((item) => item.availableForSubagents) ?? models.find((item) => item.isDefault);
             if (!chosen) fail("MODEL_NOT_CONFIGURED", "No matching configured model is available");
+            // A worker the agent starts is AI-driven delegation, so naming a model here
+            // needs that model's own `availableForSubagents` opt-in — the gate `Task.model`
+            // already applies (ADR subagent-model-opt-in; #386). Inheriting stays open, and
+            // naming the default is that inheritance spelled out, the way repeating a
+            // definition's own pin is on the Task path.
+            if (requested && !chosen.availableForSubagents && !chosen.isDefault) {
+              fail("PERMISSION_DENIED", `Model "${requested}" is not enabled for AI delegation. Turn on "Available for AI delegation" for it in model settings, or omit modelKey to inherit.`);
+            }
             model = parsePluginModelKey(chosen.key);
           }
           checkCurrent(host, input.signal);
diff --git a/apps/desktop/test/session-collaboration-service.test.mjs b/apps/desktop/test/session-collaboration-service.test.mjs
index 411f23bbc..7bb640534 100644
--- a/apps/desktop/test/session-collaboration-service.test.mjs
+++ b/apps/desktop/test/session-collaboration-service.test.mjs
@@ -20,10 +20,18 @@ function load(relative, imports) {
   return module.exports;
 }
 
+// The real model helpers, not a stub: the opt-in flag has to survive the whole
+// way from `providers.list` through the plugin model list into the spawn
+// decision, and a hand-written stub would prove only the last step.
+const pluginAgentComplete = load("../electron/main/plugin-agent-complete.ts", {
+  "@pi-desktop/agent-runtime": await import("@pi-desktop/agent-runtime"),
+  "@pi-desktop/shared": await import("@pi-desktop/shared"),
+});
+
 const { createSessionCollaborationService } = load("../electron/main/services/session-collaboration.ts", {
   "node:crypto": crypto,
   "../agent-host-bridge": { DESKTOP_PRINCIPAL: { kind: "desktop" } },
-  "../plugin-agent-complete": {},
+  "../plugin-agent-complete": pluginAgentComplete,
 });
 
 function createHost(name, respond = () => ({})) {
@@ -234,3 +242,92 @@ test("an omitted session message kind defaults to message", async () => {
     assert.equal(send[0].params.kind, kind ?? "message");
   }
 });
+
+// Spawning a worker is the agent choosing a model for work it delegates, so
+// the per-model "Available for AI delegation" opt-in governs it the same way
+// it governs `Task.model` (#386).
+const SPAWN_PROVIDERS = [
+  {
+    id: "provider-a",
+    name: "Provider A",
+    enabled: true,
+    authKind: "none",
+    models: [
+      { id: "default-model" },
+      { id: "private-model" },
+      { id: "delegable-model", availableForSubagents: true },
+    ],
+  },
+];
+
+function spawnHost() {
+  const stored = sessionMessage({ id: "spawn-message", status: "completed", kind: "task" });
+  return createHost("host-1", (method) => {
+    if (method === "providers.list") return { providers: SPAWN_PROVIDERS };
+    if (method === "settings.get") {
+      return { defaultProviderId: "provider-a", defaultModelId: "default-model" };
+    }
+    if (method === "session.collaboration.spawn") return { message: stored };
+    if (method === "session.collaboration.message") return { message: stored };
+    return { messages: [] };
+  });
+}
+
+function invokeSpawn(service, args) {
+  return service.invoke({
+    operation: "session/collaboration/spawn",
+    source: "plugin",
+    args: [args],
+    pluginContext: {
+      pluginId: "demo-plugin",
+      sessionId: "source-session",
+      turnId: "turn-1",
+      invocationId: "invocation-1",
+    },
+  });
+}
+
+function spawnService(host) {
+  return createService({ getHost: () => host, activeTurns: { "source-session": "turn-1" } }).service;
+}
+
+test("spawn refuses a model the user did not enable for AI delegation (#386)", async () => {
+  const host = spawnHost();
+
+  await assert.rejects(
+    invokeSpawn(spawnService(host), { task: "Review the diff", modelKey: "provider-a/private-model" }),
+    { code: "PERMISSION_DENIED", message: /not enabled for AI delegation/ },
+  );
+  assert.deepEqual(methodCalls(host, "session.collaboration.spawn"), []);
+});
+
+test("spawn accepts a model enabled for AI delegation (#386)", async () => {
+  const host = spawnHost();
+
+  await invokeSpawn(spawnService(host), { task: "Review the diff", modelKey: "provider-a/delegable-model" });
+
+  const spawns = methodCalls(host, "session.collaboration.spawn");
+  assert.equal(spawns.length, 1);
+  assert.equal(spawns[0].params.providerId, "provider-a");
+  assert.equal(spawns[0].params.modelId, "delegable-model");
+});
+
+test("spawn treats the default model's own key as inheritance, not a selection (#386)", async () => {
+  const host = spawnHost();
+
+  await invokeSpawn(spawnService(host), { task: "Review the diff", modelKey: "provider-a/default-model" });
+
+  const spawns = methodCalls(host, "session.collaboration.spawn");
+  assert.equal(spawns.length, 1);
+  assert.equal(spawns[0].params.modelId, "default-model");
+});
+
+test("spawn without a model key still prefers an enabled model over the default (#386)", async () => {
+  const host = spawnHost();
+
+  await invokeSpawn(spawnService(host), { task: "Review the diff" });
+
+  const spawns = methodCalls(host, "session.collaboration.spawn");
+  assert.equal(spawns.length, 1);
+  assert.equal(spawns[0].params.modelId, "delegable-model");
+});
diff --git a/docs/adr/subagent-model-opt-in.md b/docs/adr/subagent-model-opt-in.md
index 8c255779e..397065dc3 100644
--- a/docs/adr/subagent-model-opt-in.md
+++ b/docs/adr/subagent-model-opt-in.md
@@ -2,7 +2,7 @@
 
 - Status: Accepted for implementation
 - Date: 2026-09-13
-- Related: D278, ADR 0062, ADR 0089, E2E-166, issue #286
+- Related: D278, ADR 0062, ADR 0089, ADR 0237, E2E-166, issues #286, #386
 
 ## Context
 
@@ -34,6 +34,18 @@ name rule as pin resolution; ambiguous vendor aliases fail closed unless the
 caller uses the exact provider id. A changed launch list retires an idle
 runtime on the next prompt, including after opt-in is revoked.
 
+The opt-in belongs to the decision, not to the Task tool: it governs every
+entry point through which the AI picks a model for work it delegates.
+`session/collaboration/spawn` is the second such entry point (ADR 0237), and
+its `modelKey` is that same selection written by a plugin on the agent's
+behalf, so an unopted model is refused with `PERMISSION_DENIED` before a worker
+exists. The inheritance half is unchanged: omitting `modelKey` still takes the
+first opted-in model and otherwise the default, and naming the default model's
+own key is that inheritance spelled out, exactly as repeating a definition's
+own pin is on the Task path. `models.list` keeps reporting every ready model
+with its `availableForSubagents` flag — the flag is advice to the caller and
+authority only in main, which is the side a plugin cannot rewrite (#386).
+
 D278's priority remains Task.model → definition pin → session model. The
 existing exact-session-model exception remains unchanged. Repeating the
 target definition's own pin key is treated as omitting `model`, so catalog
diff --git a/docs/spec/03-runtime/11-provider-model-system.md b/docs/spec/03-runtime/11-provider-model-system.md
index d9158e6d8..fd7158c2b 100644
--- a/docs/spec/03-runtime/11-provider-model-system.md
+++ b/docs/spec/03-runtime/11-provider-model-system.md
@@ -351,7 +351,9 @@ normal pin resolution, including when `Task.model` repeats that definition's
 own pin key. On-demand matching uses unique provider id/vendor/name lookup and
 must not overwrite a pin with another account's credentials. If vendor/model aliases collide across accounts, the
 opted-in account uses its exact provider ID as the override key. Selection priority remains Task.model → definition pin
-→ session model (D278; ADR subagent-model-opt-in).
+→ session model (D278; ADR subagent-model-opt-in). The opt-in governs every entry point that lets the AI pick a model
+for delegated work, not only `Task.model`: a `session/collaboration/spawn` `modelKey` naming a model without it is
+refused with `PERMISSION_DENIED`, while omitting the key, or naming the default model's own key, still inherits.
 
 ## 8. Secrets
 
diff --git a/docs/spec/06-delivery/04-e2e-test-plan.md b/docs/spec/06-delivery/04-e2e-test-plan.md
index 4894f933a..486c5ad67 100644
--- a/docs/spec/06-delivery/04-e2e-test-plan.md
+++ b/docs/spec/06-delivery/04-e2e-test-plan.md
@@ -10922,7 +10922,9 @@ are withdrawn with ADR 0165.
   using the exact returned `sessionId`, and stop another worker. 10) Restart
   the host/plugin with a queued delivery and confirm it remains held, while an
   interrupted turn is not replayed. Repeat the parallel creation step with a
-  large existing session list while keeping the parent visible.
+  large existing session list while keeping the parent visible. 11) With one
+  model's «Available for AI delegation» left off, ask the parent to spawn a
+  worker on it by `modelKey`, then spawn again with no `modelKey`.
 - **Expected**: Each worker is a real durable session with the parent's
   project/model/thinking/permission ceiling and an independent empty
   transcript at creation. The host ledger binds every delivery to the actual
@@ -10937,7 +10939,10 @@ are withdrawn with ADR 0165.
   interrupts only the selected delivery/turn without deleting the session.
   Sends without an active plugin tool invocation, forged source ids, targets
   above the source permission ceiling, worker fan-out overflow, inbox overflow,
-  and autonomous callback loops fail closed. Unrelated sessions and the
+  and autonomous callback loops fail closed. A `spawn` naming a model the user
+  has not enabled for AI delegation is refused with `PERMISSION_DENIED` before
+  a worker exists, while omitting `modelKey` — or naming the default model's
+  own key — still inherits. Unrelated sessions and the
   existing Task family are unchanged, and no localhost MCP call or token
   access occurs. Bursts of worker notifications serialize and coalesce
   session-list refreshes while preserving the final worker list and the
@@ -10945,7 +10950,8 @@ are withdrawn with ADR 0165.
 - **Specs linked**: `07-plugins/03-plugin-api.md`,
   `07-plugins/04-plugin-security.md`, `07-plugins/11-plugin-storage-isolation.md`,
   `03-runtime/01-ipc-protocol.md`, `03-runtime/06-host-rpc-protocol.md`,
-  `03-runtime/04-data-storage.md`, ADR 0237, ADR 0239
+  `03-runtime/04-data-storage.md`, `03-runtime/11-provider-model-system.md`,
+  ADR 0237, ADR 0239, ADR subagent-model-opt-in
 - **Acceptance**: C (parallel durable sessions), D (plugin security), Quality
 - **Milestone**: M6+
 - **Status**: host ledger coverage is automated by
diff --git a/docs/spec/07-plugins/03-plugin-api.md b/docs/spec/07-plugins/03-plugin-api.md
index e16954186..c2601f8a0 100644
--- a/docs/spec/07-plugins/03-plugin-api.md
+++ b/docs/spec/07-plugins/03-plugin-api.md
@@ -464,6 +464,13 @@ and `result` are bounded projections and do not load a full transcript.
 `cancel` interrupts only the exact queued delivery or bound turn and retains
 the target session and history.
 
+A named `spawn` `modelKey` is an AI-driven delegation choice and needs that
+model's own `ModelBinding.availableForSubagents` opt-in; the host answers
+`PERMISSION_DENIED` for a model the user has not enabled, before creating a
+worker. Omitting `modelKey` still inherits — the first enabled model, else the
+default — and naming the default model's own key is that same inheritance
+rather than a selection (ADR subagent-model-opt-in).
+
 `list` returns at most 100 non-deleted Agent sessions that can receive a
 message, including sessions created independently of Session Orchestrator. Each
 entry contains only its Session ID, title, status, updated time, readable
diff --git a/docs/spec/08-meta/decisions-log.md b/docs/spec/08-meta/decisions-log.md
index dfd59e70c..ad20ac794 100644
--- a/docs/spec/08-meta/decisions-log.md
+++ b/docs/spec/08-meta/decisions-log.md
@@ -3072,6 +3072,13 @@ D193, and D194.
   different provider id, and do not retire an idle runtime. Repeating a
   definition's own pin key is omit. On-demand matching uses unique provider
   lookup. The Task catalog discloses each definition's default model.
+- Scope clarification (2026-09-15, issue #386): the opt-in governs every entry
+  point through which the AI picks a model for delegated work, not only
+  `Task.model`. A `session/collaboration/spawn` `modelKey` naming a model
+  without the opt-in is refused with `PERMISSION_DENIED` before a worker is
+  created; omitting the key, or naming the default model's own key, remains
+  inheritance. `models.list` still reports every ready model with its flag —
+  the flag advises the caller and is authoritative only in Electron main.
 
 - Models not pre-resolved at sidecar launch are resolved on-demand via the
   `provider.resolveSubagentModel` RPC to Electron main, where credentials and
diff --git a/docs/zh-CN/spec/03-runtime/11-provider-model-system.md b/docs/zh-CN/spec/03-runtime/11-provider-model-system.md
index 6f65423f6..f6037ad97 100644
--- a/docs/zh-CN/spec/03-runtime/11-provider-model-system.md
+++ b/docs/zh-CN/spec/03-runtime/11-provider-model-system.md
@@ -318,7 +318,9 @@ agent 系统提示的委托目录中。父 agent 随后就能通过 Task 工具
 正常的固定模型解析生效，包括 `Task.model` 重复该定义自己的固定键。按需匹配使用唯一
 provider id/vendor/name 查找，不得用另一账号凭据覆盖固定模型。多个账号的 vendor/model 别名冲突时，已勾选账号改用
 确切的提供商 ID 作为覆盖键。优先级保持 Task.model → 定义固定模型 → 会话模型
-（D278；ADR subagent-model-opt-in）。
+（D278；ADR subagent-model-opt-in）。该许可约束所有让 AI 为委派工作挑选模型的入口，
+而不只是 `Task.model`：`session/collaboration/spawn` 的 `modelKey` 指向未勾选的模型时
+以 `PERMISSION_DENIED` 拒绝，省略该键或写出默认模型自己的键仍按继承处理。
 
 ## 8. 秘密
 
diff --git a/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md b/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
index 7981d1ab4..9b96c4199 100644
--- a/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
+++ b/docs/zh-CN/spec/06-delivery/04-e2e-test-plan.md
@@ -6956,12 +6956,15 @@ runner 会在运行时的隔离临时目录中生成六个插件形态 fixture
   上下文，不创建替代会话。状态和面板刷新使用有界的轻量轮询；`wait` 在其等待上限内返回
   `timedOut`，不占满宿主工具超时时间。`cancel` 中止但不删除，无关会话和现有 Task 系列
   保持不变，且不发生 localhost MCP 调用或 token 访问。Worker 不能再创建 Worker，
-  不属于调用方的 Session ID 必须被拒绝，并发上限超出时必须安全失败。Worker 创建和
+  不属于调用方的 Session ID 必须被拒绝，并发上限超出时必须安全失败。`spawn` 指定用户
+  未勾选「可供 AI 自动调度」的模型时，在创建 Worker 之前以 `PERMISSION_DENIED` 拒绝；
+  省略 `modelKey`，或写出默认模型自己的键，仍按继承处理。Worker 创建和
   prompt 通知突发时，会话列表刷新串行执行并合并，同时保留最终 Worker 列表和前台会话。
   后到达的通知等待后续读取，不会因复用 Worker 创建之前已开始的读取而丢失。
 - **链接规格**：`07-plugins/03-plugin-api.md`、`07-plugins/04-plugin-security.md`、
   `07-plugins/11-plugin-storage-isolation.md`、`03-runtime/01-ipc-protocol.md`、
-  `03-runtime/06-host-rpc-protocol.md`、ADR 0237
+  `03-runtime/06-host-rpc-protocol.md`、`03-runtime/11-provider-model-system.md`、
+  ADR 0237、ADR subagent-model-opt-in
 - **接受**：C（并行持久化会话）、D（插件安全性）、品质
 - **里程碑**：M6+
 - **状态**：host ledger 覆盖由 `pnpm test:e2e:collaboration` 自动化；marketplace 插件测试覆盖插件运行时，host-core 和 desktop 单元测试覆盖新增的宿主原子能力。完整真实 provider/Electron 旅程仍需在具备条件的 runner 中验证，遵循无本地 E2E 策略
diff --git a/docs/zh-CN/spec/07-plugins/03-plugin-api.md b/docs/zh-CN/spec/07-plugins/03-plugin-api.md
index 013d40654..35b6fc4aa 100644
--- a/docs/zh-CN/spec/07-plugins/03-plugin-api.md
+++ b/docs/zh-CN/spec/07-plugins/03-plugin-api.md
@@ -381,6 +381,12 @@ Session ID，并复用该会话的项目、模型、上下文和权限配置；`
 不是 worker 身份。`status` 和 `result` 是有界投影，不会加载完整转录本。`cancel` 只中断
 精确的排队投递或绑定回合，并保留目标会话及其历史。
 
+`spawn` 中显式指定的 `modelKey` 属于 AI 自动调度的模型选择，需要该模型自身的
+`ModelBinding.availableForSubagents` 许可；对用户未勾选的模型，宿主在创建 worker 之前
+返回 `PERMISSION_DENIED`。省略 `modelKey` 仍然是继承——先取已勾选的模型，否则取默认
+模型——显式写出默认模型自己的键同样按继承处理，而不是一次选择
+（ADR subagent-model-opt-in）。
+
 `spawn` 和 `send` 仅在插件当前 Agent 工具调用期间有效。broker 注入 `pluginId`、来源
 `sessionId`、来源 `turnId` 和调用身份；插件参数不能提供或覆盖这些字段。面向用户的插件
 面板可使用自有插件身份调用 `cancel`，但不能用该路径发送或创建工作。宿主执行来源权限
diff --git a/docs/zh-CN/spec/08-meta/decisions-log.md b/docs/zh-CN/spec/08-meta/decisions-log.md
index 9f762ec02..2f9cd518b 100644
--- a/docs/zh-CN/spec/08-meta/decisions-log.md
+++ b/docs/zh-CN/spec/08-meta/decisions-log.md
@@ -2684,6 +2684,11 @@ D193 和 D194。
   只有启动许可键进入复用快照。按需授权写入独立缓存，不得用不同 provider id
   覆盖固定模型，也不会替换空闲运行时。重复定义自己的固定键视为省略 `model`。
   按需匹配使用唯一提供商查找。Task 目录同时展示各定义的默认模型。
+- 范围澄清（2026-09-15，issue #386）：该许可约束所有让 AI 为委派工作挑选模型
+  的入口，不只是 `Task.model`。`session/collaboration/spawn` 的 `modelKey`
+  指向未勾选的模型时，在创建 worker 之前以 `PERMISSION_DENIED` 拒绝；省略该键
+  或写出默认模型自己的键仍按继承处理。`models.list` 依旧返回全部就绪模型及其
+  标志——该标志只是给调用方的提示，权威判断仅在 Electron main。
 - sidecar 启动时未预解析的模型通过发往 Electron main 的
   `provider.resolveSubagentModel` RPC 按需解析，凭据和 models.dev 快照都在
   那里。