diff --git a/apps/desktop/src/components/settings/ModelConfigPage.tsx b/apps/desktop/src/components/settings/ModelConfigPage.tsx
index ac41c47bb2..c8a670be9f 100644
--- a/apps/desktop/src/components/settings/ModelConfigPage.tsx
+++ b/apps/desktop/src/components/settings/ModelConfigPage.tsx
@@ -32,6 +32,13 @@ import {
 } from "./image-generation-default";
 import { copyProviderConfiguration, type ProviderCopyDraft } from "./provider-copy";
 import { ImageGenerationModelRow } from "./ImageGenerationModelRow";
+import {
+  subagentModelCandidates,
+  subagentModelOptInDraft,
+  type SubagentModelCandidate,
+} from "./subagent-delegation";
+import { SubagentModelRow } from "./SubagentModelRow";
+import { SubagentModelsDialog } from "./SubagentModelsDialog";
 import { OAuthLoginDialog } from "./OAuthLoginDialog";
 import { ProviderSetupDialog } from "./ProviderSetupDialog";
 import { ServiceList } from "./ServiceList";
@@ -110,6 +117,8 @@ export function ModelConfigPage() {
   const rowBusy = useRowBusy();
   const testingRows = useRowBusy();
   const [changingImageModel, setChangingImageModel] = useState(false);
+  const [changingSubagentModel, setChangingSubagentModel] = useState(false);
+  const [subagentEditorOpen, setSubagentEditorOpen] = useState(false);
   const [refreshingCatalog, setRefreshingCatalog] = useState(false);
   const [catalogStatus, setCatalogStatus] = useState<CatalogStatus | null>(null);
   const [importOpen, setImportOpen] = useState(false);
@@ -148,6 +157,9 @@ export function ModelConfigPage() {
     ], null),
     [settings?.imageGenerationModels, settings?.imageGeneration, providers],
   );
+
+  // Every configured model the page can offer for delegation, in service order.
+  const subagentModels = useMemo(() => subagentModelCandidates(providers), [providers]);
   const providerReady = (provider: ProviderPublic) =>
     providerServesChatModels(provider, imageGenerationCandidates);
 
@@ -289,9 +301,53 @@ export function ModelConfigPage() {
         variant: "error",
       });
     } finally {
+
       setChangingImageModel(false);
     }
   };
+  /**
+   * Persist one delegation opt-in.
+   *
+   * The flag lives on the service's own model binding, so this is the same write
+   * the provider dialog performs. It must start from the stored service rather
+   * than this page's snapshot: `providers.update` replaces the whole model list,
+   * and a stale list would drop a model another window added in the meantime.
+   */
+  const setSubagentDelegation = async (
+    candidate: SubagentModelCandidate,
+    selected: boolean,
+  ) => {
+    setChangingSubagentModel(true);
+    try {
+      const { providers: stored } = await api.listProviders();
+      const provider = stored.find((entry) => entry.id === candidate.providerId);
+      // A plugin-owned row is refused by the RPC, and a model that disappeared
+      // from the stored bindings has no flag left to write.
+      const writable =
+        !!provider &&
+        !provider.ownerPluginId &&
+        provider.models.some((model) => model.id === candidate.modelId);
+      if (!provider || !writable) {
+        showToast(t("settings.subagentModelSaveFailed"), { variant: "error" });
+        return;
+      }
+      await api.updateProvider({
+        id: provider.id,
+        models: subagentModelOptInDraft(provider.models, candidate.modelId, selected),
+      });
+      await refreshProviders();
+      showToast(
+        t(selected ? "settings.subagentModelSelected" : "settings.subagentModelUnselected"),
+        { variant: "success" },
+      );
+    } catch (error) {
+      showToast(error instanceof Error ? error.message : String(error), {
+        variant: "error",
+      });
+    } finally {
+      setChangingSubagentModel(false);
+    }
+  };
 
   const toggleEnabled = async (provider: ProviderPublic) => {
     if (!rowBusy.begin(provider.id)) return;
@@ -559,6 +615,32 @@ export function ModelConfigPage() {
         </section>
       ) : null}
 
+      {subagentModels.length > 0 ? (
+        <section className="settings-card-block">
+          <div className="settings-panel model-default-panel">
+            <SubagentModelRow
+              candidates={subagentModels}
+              onEdit={() => setSubagentEditorOpen(true)}
+            />
+          </div>
+        </section>
+      ) : null}
+
+      {subagentEditorOpen ? (
+        <SubagentModelsDialog
+          candidates={subagentModels}
+          busy={changingSubagentModel}
+          onChange={(candidate, selected) =>
+            void setSubagentDelegation(candidate, selected)
+          }
+          onEdit={(providerId) => {
+            setSubagentEditorOpen(false);
+            setSetupFor(providerId);
+          }}
+          onClose={() => setSubagentEditorOpen(false)}
+        />
+      ) : null}
+
       {setupFor !== null ? (
         <ProviderSetupDialog
           key={setupFor}
diff --git a/apps/desktop/src/components/settings/SubagentModelRow.tsx b/apps/desktop/src/components/settings/SubagentModelRow.tsx
new file mode 100644
index 0000000000..6fb1e92959
--- /dev/null
+++ b/apps/desktop/src/components/settings/SubagentModelRow.tsx
@@ -0,0 +1,120 @@
+import { useLayoutEffect, useRef, useState } from "react";
+import { useTranslation } from "react-i18next";
+import { Button } from "../ui";
+import {
+  fitSubagentChips,
+  selectedSubagentModels,
+  type ChipFit,
+  type SubagentModelCandidate,
+} from "./subagent-delegation";
+
+/** Chip spacing, mirrored in `model-config.css`. */
+const CHIP_GAP = 5;
+/** The chip line never takes more than this share of the row. */
+const CHIP_SHARE = 0.6;
+/** Keep a few pixels clear so a rounding error cannot clip the last chip. */
+const CHIP_SAFETY = 12;
+
+/**
+ * One quiet line on the Models page: which models the AI may pick when it
+ * delegates, and a way into the editor.
+ *
+ * The row stays short by design — a configured install can delegate to a dozen
+ * models — so the chips take at most 60% of the row and the rest collapse into
+ * `+N`. Widths are measured from the rendered chips, because only the browser
+ * knows how wide a model id is at the current font and zoom.
+ */
+export function SubagentModelRow({
+  candidates,
+  onEdit,
+}: {
+  candidates: SubagentModelCandidate[];
+  onEdit: () => void;
+}) {
+  const { t } = useTranslation();
+  const selected = selectedSubagentModels(candidates);
+  const names = selected.map((candidate) => candidate.label);
+  const identity = names.join("\u0000");
+  const measureRef = useRef<HTMLSpanElement | null>(null);
+  const [fit, setFit] = useState<ChipFit>({ visible: names.length, hidden: 0 });
+
+  useLayoutEffect(() => {
+    if (typeof ResizeObserver === "undefined") return;
+    const element = measureRef.current;
+    if (!element) return;
+    const compute = () => {
+      const chips = Array.from(element.children) as HTMLElement[];
+      if (chips.length < 2) return;
+      const moreChip = chips[chips.length - 1];
+      const widths = chips.slice(0, -1).map((chip) => chip.getBoundingClientRect().width);
+      const rowWidth = element.parentElement?.getBoundingClientRect().width ?? 0;
+      setFit(
+        fitSubagentChips({
+          widths,
+          gap: CHIP_GAP,
+          available: Math.max(0, rowWidth * CHIP_SHARE - CHIP_SAFETY),
+          moreWidth: moreChip.getBoundingClientRect().width,
+        }),
+      );
+    };
+    compute();
+    const observer = new ResizeObserver(compute);
+    if (element.parentElement) observer.observe(element.parentElement);
+    return () => observer.disconnect();
+  }, [identity]);
+
+  // Nothing configured yet: the page hides this row rather than offering an
+  // editor with no models in it.
+  if (candidates.length === 0) return null;
+
+  return (
+    <div className="settings-row model-default-row model-subagent-row">
+      <div className="settings-row-copy model-default-copy">
+        <div className="settings-row-title model-default-label">
+          {t("settings.subagentModels")}
+        </div>
+        <div className="settings-row-detail model-default-value">
+          {selected.length === 0 ? (
+            <span className="model-default-empty" role="status">
+              {t("settings.subagentModelsUnset")}
+            </span>
+          ) : (
+            <span
+              className="model-subagent-chips"
+              aria-label={names.join(", ")}
+              title={names.join(", ")}
+            >
+              {names.slice(0, fit.visible).map((name) => (
+                <span className="model-subagent-chip" key={name}>
+                  {name}
+                </span>
+              ))}
+              {fit.hidden > 0 ? (
+                <span
+                  className="model-subagent-chip model-subagent-chip-more"
+                  aria-label={t("settings.subagentModelsMore", { count: fit.hidden })}
+                >
+                  {`+${fit.hidden}`}
+                </span>
+              ) : null}
+            </span>
+          )}
+          {/* Off-screen twin: the only way to know how wide each chip is. */}
+          <span ref={measureRef} className="model-subagent-measure" aria-hidden="true">
+            {names.map((name) => (
+              <span className="model-subagent-chip" key={name}>
+                {name}
+              </span>
+            ))}
+            <span className="model-subagent-chip">{`+${names.length}`}</span>
+          </span>
+        </div>
+      </div>
+      <div className="settings-row-control">
+        <Button size="sm" variant="secondary" onClick={onEdit}>
+          {t("settings.subagentModelEdit")}
+        </Button>
+      </div>
+    </div>
+  );
+}
diff --git a/apps/desktop/src/components/settings/SubagentModelsDialog.tsx b/apps/desktop/src/components/settings/SubagentModelsDialog.tsx
new file mode 100644
index 0000000000..92d4b8a604
--- /dev/null
+++ b/apps/desktop/src/components/settings/SubagentModelsDialog.tsx
@@ -0,0 +1,171 @@
+import { useState } from "react";
+import { useTranslation } from "react-i18next";
+import { formatTokenCount } from "@pi-desktop/shared";
+import { useBlockingOverlay } from "../../lib/blocking-overlay";
+import { portalToBody } from "../../lib/portal-visibility";
+import { IconClose, IconSearch } from "../icons";
+import { Button, OverlayBackdrop, TooltipButton } from "../ui";
+import { selectedSubagentModels, type SubagentModelCandidate } from "./subagent-delegation";
+
+/**
+ * The subagent delegation editor: the models the AI may pick when it delegates,
+ * listed the way a service lists its added models — same search field, same row,
+ * same Advanced button, same remove control.
+ *
+ * Adding is not this dialog's job: a model joins the list by ticking **Available
+ * for AI delegation** in that model's own editor, which is exactly where Advanced
+ * goes. Both surfaces read and write `availableForSubagents` on the binding, so
+ * neither can drift from the other.
+ */
+export function SubagentModelsDialog({
+  candidates,
+  busy = false,
+  onChange,
+  onEdit,
+  onClose,
+}: {
+  candidates: SubagentModelCandidate[];
+  busy?: boolean;
+  onChange: (candidate: SubagentModelCandidate, selected: boolean) => void;
+  /** Opens the owning service's model editor, where the opt-in is ticked. */
+  onEdit: (providerId: string) => void;
+  onClose: () => void;
+}) {
+  useBlockingOverlay();
+  const { t } = useTranslation();
+  const [query, setQuery] = useState("");
+  const selected = selectedSubagentModels(candidates);
+  const needle = query.trim().toLowerCase();
+  const visible = needle
+    ? selected.filter((candidate) =>
+        `${candidate.providerName} ${candidate.label}`.toLowerCase().includes(needle),
+      )
+    : selected;
+  const name = (candidate: SubagentModelCandidate) =>
+    `${candidate.providerName} · ${candidate.label}`;
+
+  const dialog = (
+    <OverlayBackdrop
+      className="overlay project-create-dialog-overlay"
+      role="presentation"
+      onClick={(event) => {
+        if (event.target === event.currentTarget) onClose();
+      }}
+    >
+      <div
+        className="dialog model-subagent-dialog"
+        role="dialog"
+        aria-modal="true"
+        aria-labelledby="model-subagent-dialog-title"
+        onClick={(event) => event.stopPropagation()}
+      >
+        <div className="project-create-dialog-head">
+          <div className="project-create-dialog-heading">
+            <h2 id="model-subagent-dialog-title" className="project-create-dialog-title">
+              {t("settings.subagentModels")}
+            </h2>
+            <p className="project-edit-dialog-description">
+              {t("settings.subagentModelAdd")}
+            </p>
+          </div>
+          <TooltipButton
+            type="button"
+            className="project-create-dialog-close"
+            tooltip={t("common.close")}
+            ariaLabel={t("common.close")}
+            onClick={onClose}
+          >
+            <IconClose size={17} />
+          </TooltipButton>
+        </div>
+
+        <div className="provider-chosen model-subagent-panel">
+          <div className="provider-chosen-head">
+            <h4 className="provider-chosen-title">{t("settings.subagentModels")}</h4>
+            <span className="provider-chosen-count">{selected.length}</span>
+            <div className="provider-chosen-search-wrap">
+              <IconSearch size={13} aria-hidden />
+              <input
+                className="provider-chosen-search"
+                value={query}
+                placeholder={t("settings.searchChosenModels")}
+                aria-label={t("settings.searchChosenModels")}
+                disabled={busy || selected.length === 0}
+                spellCheck={false}
+                autoCorrect="off"
+                autoCapitalize="off"
+                autoComplete="off"
+                onChange={(event) => setQuery(event.target.value)}
+              />
+            </div>
+          </div>
+          {selected.length === 0 ? (
+            <div className="provider-chosen-empty">{t("settings.subagentModelsUnset")}</div>
+          ) : visible.length === 0 ? (
+            <div className="provider-chosen-empty">{t("settings.noChosenModelMatches")}</div>
+          ) : (
+            <ul className="provider-chosen-list">
+              {visible.map((candidate) => (
+                <li
+                  className="provider-chosen-row"
+                  key={`${candidate.providerId}:${candidate.modelId}`}
+                >
+                  <div className="provider-chosen-row-head">
+                    <span className="provider-chosen-row-id font-mono selectable">
+                      {candidate.label}
+                    </span>
+                    <span className="model-subagent-provider">{candidate.providerName}</span>
+                    <span className="provider-chosen-row-limits">
+                      {formatTokenCount(candidate.contextWindow)} ·{" "}
+                      {formatTokenCount(candidate.maxTokens)}
+                    </span>
+                    {candidate.blocked ? (
+                      <span className="model-subagent-note">
+                        {t(
+                          candidate.blocked === "provider-disabled"
+                            ? "settings.subagentModelProviderDisabled"
+                            : "settings.subagentModelCredentialMissing",
+                        )}
+                      </span>
+                    ) : null}
+                    <button
+                      type="button"
+                      className="provider-chosen-advanced-toggle"
+                      onClick={() => onEdit(candidate.providerId)}
+                    >
+                      {t("settings.advanced")}
+                    </button>
+                    {candidate.editable ? (
+                      <TooltipButton
+                        type="button"
+                        className="provider-chosen-remove"
+                        ariaLabel={t("settings.subagentModelRemove", { model: name(candidate) })}
+                        tooltip={t("settings.subagentModelRemove", { model: name(candidate) })}
+                        disabled={busy}
+                        onClick={() => onChange(candidate, false)}
+                      >
+                        <IconClose size={12} />
+                      </TooltipButton>
+                    ) : (
+                      <span className="model-subagent-note">
+                        {t("settings.subagentModelPluginManaged")}
+                      </span>
+                    )}
+                  </div>
+                </li>
+              ))}
+            </ul>
+          )}
+        </div>
+
+        <div className="model-subagent-dialog-actions">
+          <Button variant="primary" onClick={onClose}>
+            {t("common.close")}
+          </Button>
+        </div>
+      </div>
+    </OverlayBackdrop>
+  );
+
+  return typeof document === "undefined" ? dialog : portalToBody(dialog);
+}
diff --git a/apps/desktop/src/components/settings/subagent-delegation.ts b/apps/desktop/src/components/settings/subagent-delegation.ts
new file mode 100644
index 0000000000..0885346088
--- /dev/null
+++ b/apps/desktop/src/components/settings/subagent-delegation.ts
@@ -0,0 +1,198 @@
+/**
+ * The Models page's subagent-delegation list.
+ *
+ * `ModelBinding.availableForSubagents` is the only source of truth for the
+ * delegation catalog: the runtime builds that catalog from it at launch
+ * (`packages/host-runtime/src/launch-resolver.ts`) and authorizes every
+ * `Task.model` override against the same catalog. This list therefore reads and
+ * clears that flag on the binding itself — the same field a service's model
+ * editor writes — so the page and the catalog cannot disagree, and a change made
+ * in either place is visible in both.
+ *
+ * The helpers stay free of runtime imports so the catalog rules can be tested
+ * without the application bundle.
+ */
+import type { ModelBinding, ProviderPublic } from "@pi-desktop/shared";
+
+/** Why an opted-in model cannot reach the catalog today. */
+export type SubagentModelBlockedReason = "provider-disabled" | "credential-missing";
+
+/** One configured model, with the delegation state the page renders. */
+export type SubagentModelCandidate = {
+  providerId: string;
+  providerName: string;
+  modelId: string;
+  /** The alias the rest of Settings shows, else the wire id. */
+  label: string;
+  /** The binding's limits, shown the way a service's model list shows them. */
+  contextWindow: number;
+  maxTokens: number;
+  /** True when this binding carries the delegation opt-in. */
+  selected: boolean;
+  /**
+   * False only when a plugin owns the service row: the plugin refreshes it on
+   * every load and `providers.update` refuses it. A service that is merely off
+   * stays editable so its opt-in can be prepared before it is switched on.
+   */
+  editable: boolean;
+  /** Set when the opt-in exists but the service cannot serve it yet. */
+  blocked?: SubagentModelBlockedReason;
+};
+
+/** The provider facts the delegation rules read; anything else is ignored. */
+export type SubagentProviderFacts = {
+  enabled: boolean;
+  hasSecret?: boolean;
+  hasOauth?: boolean;
+  authKind?: string;
+};
+
+/**
+ * Whether this service can serve a delegated run at all.
+ *
+ * Same rule the subagent editor's own model picker applies to the Composer's
+ * runnable services: the row must be enabled and hold a usable credential (an
+ * API key, a vendor-account login, or a keyless local endpoint).
+ */
+export function providerServesSubagentDelegation(provider: SubagentProviderFacts): boolean {
+  return (
+    provider.enabled &&
+    (provider.hasSecret === true ||
+      provider.hasOauth === true ||
+      provider.authKind === "none")
+  );
+}
+
+/**
+ * Why an opt-in cannot reach the catalog yet.
+ *
+ * The rule is credential *presence*, not auth-kind identity: `hasSecret` is the
+ * host's own "this row holds a usable credential" flag (an API key or a vendor
+ * account), so an auth flow added later still counts here without teaching this
+ * page about it. Only the keyless local endpoint is named explicitly, because
+ * that is the one case where no credential is expected — which is why a row that
+ * reports no credential at all really is unusable, rather than a guess at why.
+ */
+function blockedReason(provider: SubagentProviderFacts): SubagentModelBlockedReason | undefined {
+  if (!provider.enabled) return "provider-disabled";
+  return providerServesSubagentDelegation(provider) ? undefined : "credential-missing";
+}
+
+function candidateKey(providerId: string, modelId: string): string {
+  // Wire ids travel the request path, so identity is case-insensitive here the
+  // same way `sameWireId` treats them elsewhere in Settings.
+  return `${providerId}\u0000${modelId.trim().toLowerCase()}`;
+}
+
+function modelLabel(binding: ModelBinding): string {
+  const alias = binding.alias?.trim();
+  return alias && alias.length > 0 ? alias : binding.id;
+}
+
+/**
+ * Every configured model, in service order, with its delegation state.
+ *
+ * A model is listed even when its service is off or missing a credential: the
+ * flag is a durable choice, and hiding it would leave an opt-in the user cannot
+ * see or withdraw.
+ */
+export function subagentModelCandidates(
+  providers: readonly ProviderPublic[],
+): SubagentModelCandidate[] {
+  const candidates: SubagentModelCandidate[] = [];
+  const seen = new Set<string>();
+  for (const provider of providers) {
+    const blocked = blockedReason(provider);
+    for (const binding of provider.models ?? []) {
+      const modelId = binding.id.trim();
+      if (!modelId) continue;
+      const key = candidateKey(provider.id, modelId);
+      if (seen.has(key)) continue;
+      seen.add(key);
+      candidates.push({
+        providerId: provider.id,
+        providerName: provider.name,
+        modelId,
+        label: modelLabel(binding),
+        contextWindow: binding.contextWindow,
+        maxTokens: binding.maxTokens,
+        selected: binding.availableForSubagents === true,
+        editable: !provider.ownerPluginId,
+        ...(blocked ? { blocked } : {}),
+      });
+    }
+  }
+  return candidates;
+}
+
+/** The opted-in rows, which are the delegation catalog's own membership. */
+export function selectedSubagentModels(
+  candidates: readonly SubagentModelCandidate[],
+): SubagentModelCandidate[] {
+  return candidates.filter((candidate) => candidate.selected);
+}
+
+/**
+ * The service's bindings with one delegation opt-in changed.
+ *
+ * A cleared opt-in drops the key instead of storing `false`, matching the
+ * provider dialog: the field is absent on every binding that never wanted
+ * delegation, and `availableForSubagents: false` never accumulates on records.
+ * Every other field is carried over untouched, because the RPC replaces the
+ * whole model list.
+ */
+export function subagentModelOptInDraft(
+  models: readonly ModelBinding[],
+  modelId: string,
+  selected: boolean,
+): ModelBinding[] {
+  const target = modelId.trim().toLowerCase();
+  return models.map((model) =>
+    model.id.trim().toLowerCase() === target
+      ? { ...model, availableForSubagents: selected ? true : undefined }
+      : model,
+  );
+}
+
+/** How many chips fit on one line, and how many collapse into `+N`. */
+export type ChipFit = { visible: number; hidden: number };
+
+/**
+ * Fit chips into a bounded line, reserving room for the `+N` chip when not
+ * everything fits.
+ *
+ * Widths are measured by the caller (fonts and zoom decide them), so the rule
+ * itself stays pure and testable: fill in order, then give back trailing chips
+ * until `+N` fits beside them. One chip always survives, so a very narrow row
+ * still says what it is about instead of showing only `+6`.
+ */
+export function fitSubagentChips(input: {
+  widths: readonly number[];
+  gap: number;
+  available: number;
+  moreWidth: number;
+}): ChipFit {
+  const { widths, gap, available, moreWidth } = input;
+  if (widths.length === 0) return { visible: 0, hidden: 0 };
+
+  let used = 0;
+  let visible = 0;
+  for (let index = 0; index < widths.length; index += 1) {
+    const next = used + (index > 0 ? gap : 0) + widths[index];
+    if (next > available) break;
+    used = next;
+    visible += 1;
+  }
+  if (visible === widths.length) return { visible, hidden: 0 };
+
+  let hidden = widths.length - visible;
+  while (visible > 1 && used + gap + moreWidth > available) {
+    visible -= 1;
+    used -= widths[visible] + gap;
+    hidden += 1;
+  }
+  // A very narrow row still names one model; the container clips the overflow
+  // rather than showing a row that only says `+6`.
+  if (visible === 0) return { visible: 1, hidden: widths.length - 1 };
+  return { visible, hidden };
+}
diff --git a/apps/desktop/src/components/settings/subagent-models.ts b/apps/desktop/src/components/settings/subagent-models.ts
index c27515aa68..495da4e616 100644
--- a/apps/desktop/src/components/settings/subagent-models.ts
+++ b/apps/desktop/src/components/settings/subagent-models.ts
@@ -7,6 +7,7 @@
  */
 import { modelWireIdsEqual as sameComposerModelId, type ProviderPublic } from "@pi-desktop/shared";
 import { defaultModelOptions } from "./default-model";
+import { providerServesSubagentDelegation } from "./subagent-delegation";
 
 export type SubagentModelChoice = {
   /** Frontmatter `model:` value. */
@@ -25,10 +26,7 @@ export type SubagentModelChoiceGroup = {
 
 /** True when the Composer would offer this provider's configured models. */
 export function isSubagentModelProvider(provider: ProviderPublic): boolean {
-  return (
-    provider.enabled &&
-    (provider.hasSecret || Boolean(provider.hasOauth) || provider.authKind === "none")
-  );
+  return providerServesSubagentDelegation(provider);
 }
 
 /**
diff --git a/apps/desktop/src/styles/model-config.css b/apps/desktop/src/styles/model-config.css
index d98a2962cf..bab1fc9b99 100644
--- a/apps/desktop/src/styles/model-config.css
+++ b/apps/desktop/src/styles/model-config.css
@@ -1921,3 +1921,88 @@
 @media (prefers-reduced-motion: reduce) {
   .model-provider-row { transition: none; }
 }
+
+/* --------------------------------------- subagent delegation model list */
+
+/*
+  The Models page carries one quiet line: the delegated model names and an Edit
+  action. The full list — search, limits, Advanced, remove — lives in the dialog
+  that Edit opens, reusing the service model list's own parts (`.provider-chosen-*`).
+*/
+.model-subagent-chips {
+  display: inline-flex;
+  align-items: center;
+  gap: 5px;
+  max-width: 60%;
+  flex-wrap: nowrap;
+  overflow: hidden;
+}
+
+.model-subagent-chip {
+  display: inline-flex;
+  align-items: center;
+  padding: 2px 8px;
+  border-radius: var(--radius-lg);
+  background: var(--ds-bg-chip);
+  color: var(--ds-text-secondary);
+  font-size: var(--text-2xs);
+  line-height: var(--leading-normal);
+  white-space: nowrap;
+}
+
+.model-subagent-chip-more {
+  color: var(--ds-text-muted);
+}
+
+/* Off-screen twin used to measure chips at the current font and zoom. */
+.model-subagent-measure {
+  position: absolute;
+  display: inline-flex;
+  gap: 5px;
+  visibility: hidden;
+  pointer-events: none;
+  white-space: nowrap;
+}
+
+.model-subagent-dialog {
+  display: flex;
+  flex-direction: column;
+  gap: 12px;
+  width: min(720px, calc(100vw - 48px));
+  max-height: min(720px, calc(100vh - 64px));
+  padding: 16px;
+}
+
+.model-subagent-panel {
+  min-height: 0;
+  overflow: hidden;
+  border-radius: var(--radius-md);
+  background: var(--ds-tile);
+  padding: 4px 0 6px;
+}
+
+.model-subagent-dialog .provider-chosen-list {
+  max-height: 420px;
+  overflow-y: auto;
+  overscroll-behavior: contain;
+}
+
+.model-subagent-dialog-actions {
+  display: flex;
+  justify-content: flex-end;
+  gap: 8px;
+}
+
+.model-subagent-provider {
+  color: var(--ds-text-muted);
+  font-size: var(--text-2xs);
+  letter-spacing: var(--tracking-wide);
+  line-height: var(--leading-normal);
+}
+
+.model-subagent-note {
+  color: var(--ds-text-muted);
+  font-size: var(--text-2xs);
+  letter-spacing: var(--tracking-wide);
+  line-height: var(--leading-normal);
+}
diff --git a/apps/desktop/test/subagent-delegation.test.mjs b/apps/desktop/test/subagent-delegation.test.mjs
new file mode 100644
index 0000000000..acd1697ad8
--- /dev/null
+++ b/apps/desktop/test/subagent-delegation.test.mjs
@@ -0,0 +1,272 @@
+/**
+ * The Models page states which models the AI may pick when it delegates, and
+ * edits the same `availableForSubagents` opt-in the runtime reads. These tests
+ * pin that mapping: the list must never invent a delegation candidate, hide an
+ * opt-in the user cannot withdraw, or write a model list that loses another
+ * window's edit.
+ */
+import assert from "node:assert/strict";
+import { readFile } from "node:fs/promises";
+import { register } from "node:module";
+import { dirname, join } from "node:path";
+import test from "node:test";
+import { fileURLToPath, pathToFileURL } from "node:url";
+
+const here = dirname(fileURLToPath(import.meta.url));
+register(pathToFileURL(join(here, "helpers/ts-import-hooks.mjs")));
+const {
+  fitSubagentChips,
+  providerServesSubagentDelegation,
+  selectedSubagentModels,
+  subagentModelCandidates,
+  subagentModelOptInDraft,
+} = await import("../src/components/settings/subagent-delegation.ts");
+
+const binding = (id, over = {}) => ({
+  id,
+  contextWindow: 128_000,
+  maxTokens: 8_192,
+  thinkingLevels: [],
+  defaultThinkingLevel: null,
+  ...over,
+});
+
+const provider = (over = {}) => ({
+  id: "p1",
+  name: "Anthropic",
+  vendorKey: "anthropic",
+  enabled: true,
+  hasSecret: true,
+  hasOauth: false,
+  authKind: "api_key_and_base_url",
+  models: [binding("claude-haiku-4-5"), binding("claude-sonnet-4-6")],
+  ...over,
+});
+
+test("a service serves delegation exactly when it is enabled with a usable credential", () => {
+  assert.equal(providerServesSubagentDelegation(provider()), true);
+  assert.equal(providerServesSubagentDelegation(provider({ enabled: false })), false);
+  assert.equal(
+    providerServesSubagentDelegation(provider({ hasSecret: false, hasOauth: false })),
+    false,
+  );
+  assert.equal(
+    providerServesSubagentDelegation(provider({ hasSecret: false, hasOauth: true })),
+    true,
+  );
+  assert.equal(
+    providerServesSubagentDelegation(
+      provider({ hasSecret: false, hasOauth: false, authKind: "none", baseUrl: "http://127.0.0.1:11434" }),
+    ),
+    true,
+  );
+});
+
+test("every configured model is listed in service order with its opt-in state", () => {
+  const providers = [
+    provider({ models: [binding("a", { availableForSubagents: true }), binding("b")] }),
+    provider({ id: "p2", name: "Local", models: [binding("c", { availableForSubagents: true })] }),
+  ];
+  assert.deepEqual(subagentModelCandidates(providers), [
+    { providerId: "p1", providerName: "Anthropic", modelId: "a", label: "a", contextWindow: 128_000, maxTokens: 8_192, selected: true, editable: true },
+    { providerId: "p1", providerName: "Anthropic", modelId: "b", label: "b", contextWindow: 128_000, maxTokens: 8_192, selected: false, editable: true },
+    { providerId: "p2", providerName: "Local", modelId: "c", label: "c", contextWindow: 128_000, maxTokens: 8_192, selected: true, editable: true },
+  ]);
+});
+
+test("the alias is the label, and the wire id stays the identity", () => {
+  const [candidate] = subagentModelCandidates([
+    provider({ models: [binding("claude-haiku-4-5", { alias: "  Quick review " })] }),
+  ]);
+  assert.equal(candidate.label, "Quick review");
+  assert.equal(candidate.modelId, "claude-haiku-4-5");
+  // A blank alias falls back to the id rather than rendering an empty row.
+  const [blank] = subagentModelCandidates([
+    provider({ models: [binding("model-x", { alias: "   " })] }),
+  ]);
+  assert.equal(blank.label, "model-x");
+});
+
+test("an opt-in on a service that is off or has no credential stays visible and withdrawable", () => {
+  const disabled = subagentModelCandidates([
+    provider({ enabled: false, models: [binding("a", { availableForSubagents: true })] }),
+  ]);
+  assert.equal(disabled[0].selected, true);
+  assert.equal(disabled[0].editable, true);
+  assert.equal(disabled[0].blocked, "provider-disabled");
+
+  const keyless = subagentModelCandidates([
+    provider({ hasSecret: false, models: [binding("a", { availableForSubagents: true })] }),
+  ]);
+  assert.equal(keyless[0].editable, true);
+  assert.equal(keyless[0].blocked, "credential-missing");
+
+  // A service that is ready carries no reason at all, so the editor adds no note.
+  const [ready] = subagentModelCandidates([provider()]);
+  assert.equal("blocked" in ready, false);
+});
+
+test("an unfamiliar auth kind is judged by the credential it holds, not by its name", () => {
+  // Review note on #1758: a future auth flow must not be labelled "no
+  // credential" merely because this page does not know its name. The rule reads
+  // credential presence, so a stored grant still counts.
+  const stored = subagentModelCandidates([
+    provider({
+      authKind: "device_code",
+      hasSecret: true,
+      hasOauth: false,
+      models: [binding("a")],
+    }),
+  ]);
+  assert.equal("blocked" in stored[0], false);
+
+  // Nothing stored: the runtime cannot obtain a key for it either, so the label
+  // states the missing credential instead of guessing at the flow.
+  const empty = subagentModelCandidates([
+    provider({
+      authKind: "device_code",
+      hasSecret: false,
+      hasOauth: false,
+      models: [binding("a")],
+    }),
+  ]);
+  assert.equal(empty[0].blocked, "credential-missing");
+  assert.equal(empty[0].editable, true);
+});
+
+test("a plugin-owned service is read-only because the RPC refuses it", () => {
+  const [candidate] = subagentModelCandidates([
+    provider({ ownerPluginId: "acme.catalogue", models: [binding("a", { availableForSubagents: true })] }),
+  ]);
+  assert.equal(candidate.selected, true);
+  assert.equal(candidate.editable, false);
+});
+
+test("duplicate and blank ids never become two rows", () => {
+  const candidates = subagentModelCandidates([
+    provider({
+      models: [binding("a", { availableForSubagents: true }), binding("A"), binding("  ")],
+    }),
+  ]);
+  assert.equal(candidates[0].selected, true);
+});
+
+test("the list carries the limits a service's own model row shows", () => {
+  const [candidate] = subagentModelCandidates([
+    provider({
+      models: [
+        binding("a", {
+          contextWindow: 1_000_000,
+          maxTokens: 128_000,
+          availableForSubagents: true,
+        }),
+      ],
+    }),
+  ]);
+  assert.equal(candidate.contextWindow, 1_000_000);
+  assert.equal(candidate.maxTokens, 128_000);
+  assert.equal(candidate.selected, true);
+  assert.deepEqual(
+    selectedSubagentModels([candidate]).map((entry) => entry.modelId),
+    ["a"],
+  );
+});
+
+
+test("the write sets the opt-in and clears it by dropping the key", () => {
+  const stored = [
+    binding("a", { alias: "Fast", contextWindowSource: "user", supportsImages: false }),
+    binding("b", { availableForSubagents: false }),
+  ];
+  const optedIn = subagentModelOptInDraft(stored, "a", true);
+  assert.deepEqual(optedIn[0], {
+    id: "a",
+    contextWindow: 128_000,
+    maxTokens: 8_192,
+    thinkingLevels: [],
+    defaultThinkingLevel: null,
+    alias: "Fast",
+    contextWindowSource: "user",
+    supportsImages: false,
+    availableForSubagents: true,
+  });
+  // The untouched binding keeps its identity, so the RPC does not rewrite it.
+  assert.equal(optedIn[1], stored[1]);
+
+  const cleared = subagentModelOptInDraft(optedIn, "a", false);
+  // Cleared means absent on the wire, which is what keeps `false` from piling up
+  // on records that never wanted delegation.
+  assert.equal(cleared[0].availableForSubagents, undefined);
+  assert.equal(JSON.stringify(cleared[0]).includes("availableForSubagents"), false);
+  assert.equal(cleared[0].alias, "Fast");
+});
+
+test("the write matches the wire id case-insensitively and leaves an absent id alone", () => {
+  const stored = [binding("Claude-Haiku-4-5")];
+  assert.equal(subagentModelOptInDraft(stored, "claude-haiku-4-5", true)[0].availableForSubagents, true);
+  const untouched = subagentModelOptInDraft(stored, "not-configured", true);
+  assert.equal(untouched[0], stored[0]);
+});
+
+test("the page mounts the row and writes the flag through the provider, not a second list", async () => {
+  const page = await readFile(
+    join(here, "../src/components/settings/ModelConfigPage.tsx"),
+    "utf8",
+  );
+  assert.match(page, /<SubagentModelRow/);
+  assert.match(page, /subagentModelCandidates\(providers\)/);
+  // The opt-in is the service's own binding field, so the save path must be the
+  // provider update that already owns it.
+  assert.match(page, /models: subagentModelOptInDraft\(provider\.models/);
+  // A plugin-owned row and a vanished model are refused instead of overwritten.
+  assert.match(page, /!provider\.ownerPluginId/);
+  assert.match(page, /settings\.subagentModelSaveFailed/);
+
+  const row = await readFile(
+    join(here, "../src/components/settings/SubagentModelRow.tsx"),
+    "utf8",
+  );
+  // The page keeps the row quiet: delegated names plus an entry point.
+  assert.match(row, /selectedSubagentModels\(candidates\)/);
+  assert.match(row, /settings\.subagentModelEdit/);
+  assert.doesNotMatch(row, /provider-chosen-list/);
+
+  const dialog = await readFile(
+    join(here, "../src/components/settings/SubagentModelsDialog.tsx"),
+    "utf8",
+  );
+  // The editor reuses the service model list's own parts: same row, same search
+  // field, same Advanced button (which opens the editor that holds the opt-in),
+  // same remove control. Adding never happens here.
+  assert.match(dialog, /className="provider-chosen-list"/);
+  assert.match(dialog, /provider-chosen-row-head/);
+  assert.match(dialog, /settings\.searchChosenModels/);
+  assert.match(dialog, /settings\.advanced/);
+  assert.match(dialog, /onEdit\(candidate\.providerId\)/);
+  assert.match(dialog, /settings\.subagentModelRemove/);
+  assert.match(dialog, /settings\.subagentModelPluginManaged/);
+  assert.match(dialog, /settings\.subagentModelProviderDisabled/);
+});
+
+test("chips fill the line and reserve room for the +N chip", () => {
+  // All five fit: 60+5+60+5+60+5+60+5+60 = 320 within 400.
+  assert.deepEqual(
+    fitSubagentChips({ widths: [60, 60, 60, 60, 60], gap: 5, available: 400, moreWidth: 30 }),
+    { visible: 5, hidden: 0 },
+  );
+  // Three fit (190), but `+N` needs 30 more than the 210 available, so the
+  // last two collapse: two chips plus `+3`.
+  assert.deepEqual(
+    fitSubagentChips({ widths: [60, 60, 60, 60, 60], gap: 5, available: 210, moreWidth: 30 }),
+    { visible: 2, hidden: 3 },
+  );
+  // A row too narrow for even one chip still names one model.
+  assert.deepEqual(
+    fitSubagentChips({ widths: [60, 60], gap: 5, available: 40, moreWidth: 30 }),
+    { visible: 1, hidden: 1 },
+  );
+  assert.deepEqual(
+    fitSubagentChips({ widths: [], gap: 5, available: 100, moreWidth: 30 }),
+    { visible: 0, hidden: 0 },
+  );
+});
diff --git a/docs/spec/03-runtime/02-agent-runtime.md b/docs/spec/03-runtime/02-agent-runtime.md
index aaa0bc144a..e00c5ad32d 100644
--- a/docs/spec/03-runtime/02-agent-runtime.md
+++ b/docs/spec/03-runtime/02-agent-runtime.md
@@ -947,8 +947,9 @@ core set rather than the on-demand catalog of §7.1:
   session model. The parent agent sees a model summary in the system prompt
   listing all models marked `availableForSubagents` in provider settings. The
   empty-catalog system summary, Task description and rejected-override error
-  point to Settings → Models → edit service/account → model Advanced →
-  "Available for AI delegation" → save, and require an exact catalog key
+  point to the **Subagent models** row on Settings → Models — or to the same flag
+  under Settings → Models → edit service/account → model Advanced → "Available
+  for AI delegation" → save — and require an exact catalog key
   rather than guessed provider/model keys. If
   the delegation catalog is empty, the prompt tells the model to omit `model`
   and use the definition pin, or inherit the session model when unpinned; an
diff --git a/docs/spec/03-runtime/11-provider-model-system.md b/docs/spec/03-runtime/11-provider-model-system.md
index 8f4402699a..f97e64ed0e 100644
--- a/docs/spec/03-runtime/11-provider-model-system.md
+++ b/docs/spec/03-runtime/11-provider-model-system.md
@@ -422,6 +422,15 @@ opted-in account uses its exact provider ID as the override key. Selection prior
 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.
 
+The flag's editing surface is not tied to the provider dialog. Settings → Models
+lists the opted-in bindings in one **Subagent models** row — each row's Advanced
+action opens the service editor where the opt-in is ticked — and reads the same
+field and rewrites only the touched binding through `providers.update`
+(`apps/desktop/src/components/settings/subagent-delegation.ts`). That row stores
+no second selection, so the delegation catalog and the page cannot disagree, and
+a service that is off, credential-less or plugin-owned is listed with its reason
+instead of disappearing.
+
 ## 8. Secrets
 
 - API keys stored via secure storage (`SECRET_*` APIs)
diff --git a/docs/spec/04-ux/06-settings-ia.md b/docs/spec/04-ux/06-settings-ia.md
index 5c31eb36aa..94296e1d52 100644
--- a/docs/spec/04-ux/06-settings-ia.md
+++ b/docs/spec/04-ux/06-settings-ia.md
@@ -445,9 +445,18 @@ a usage tab.
 
 ### Model configuration (`agent` tab)
 - The model configuration page groups its sections in this order: **AI
-  services** (including model-catalog status and refresh), **Jev**, then the
-  image-generation model selector. The sections use one consistent vertical
-  spacing rhythm.
+  services** (including model-catalog status and refresh), **Jev**, the
+- **Subagent models** is one settings row, below the image-generation selector.
+  It lists only the models that carry the `availableForSubagents` opt-in — the
+  delegation catalog the parent agent sees — as `provider · model` entries with a
+  remove action each, or an empty-state line when none does. Adding is the row's
+  own control: a menu beside the list offers exactly the configured models that
+  as `provider · model` entries with the same Advanced action and remove control
+  a service's own model list has, or an empty-state line when none does.
+  Advanced opens that service's model editor — where the opt-in is ticked — so
+  adding a model happens in the model editor, never from this row. It reads and
+  writes `ModelBinding.availableForSubagents`, never a second
+  The row is hidden while no service has a configured model.
 - **Defaults** card: a compact settings row shows the provider name and exact
   model ID beneath the Default model label. A quiet Change action opens the
   picker without duplicating the current value. The picker groups model-level
diff --git a/docs/spec/06-delivery/04-e2e-test-plan.md b/docs/spec/06-delivery/04-e2e-test-plan.md
index 3de443c3fc..e9ec0a87fb 100644
--- a/docs/spec/06-delivery/04-e2e-test-plan.md
+++ b/docs/spec/06-delivery/04-e2e-test-plan.md
@@ -10639,6 +10639,7 @@ must keep splitting are covered by `markdown-blocks.test.mjs`.
 
 | Acceptance | Scenarios |
 |---|---|
+| C / E / Quality — Subagent delegation catalog on the Models page | [E2E-268](#e2e-268-models-page-states-and-edits-the-subagent-delegation-catalog) |
 | C / E / Security / Quality — Gemini work setup | E2E-LIVE-VOICE-gemini-work-schema |
 | C / Quality — Continue after malformed tool output | E2E-RUNTIME-malformed-tool-call-continue-recovery |
 | B / Security / Quality — Optional provider client certificate requests | E2E-PROVIDER-optional-client-certificate |
@@ -18951,6 +18952,47 @@ host-created files. The full app's file-preview viewer is covered separately.
   `apps/desktop/test/main-wiring-contract.test.mjs`; the interactive Electron
   scenario remains a manual validation run.
 
+### E2E-268: Models page states and edits the subagent delegation catalog
+
+- **Preconditions**: Two AI services are configured and enabled with
+  credentials, the first with two models and the second with one; exactly one
+  model carries "Available for AI delegation". A parent Agent-mode session can
+  delegate a task.
+- **Steps**:
+  1. Open Settings → Models and read the **Subagent models** row at the bottom.
+  2. Open the second service's model editor from the row's Advanced action, tick
+     "Available for AI delegation", save, and re-read the list.
+  3. Start a parent turn and inspect the delegation model summary the Task tool
+     receives.
+  4. Return to the row, remove that model with its × action, and re-read the
+     list.
+  5. Turn the second service off and re-read the list.
+  6. Toggle "Available for AI delegation" in the first service's own model
+     dialog and return to the Models page.
+- **Expected**: The row lists exactly the opted-in models as `provider · model`
+  and shows its empty-state line when none is opted in; a model joins the list
+  only through that editor's tick. Adding writes
+  `availableForSubagents` on that service's binding and the next parent turn
+  lists the model in the delegation catalog; removing it drops the model from the
+  catalog and leaves no `false` behind on the stored binding. A service that is
+  off or holds no credential keeps its listed models visible and removable,
+  annotated with the reason they are idle, while its models stay out of the
+  catalog. The provider dialog and the row show the same flag in both
+  directions, so neither surface can drift. A plugin-owned service row renders
+  read-only. The row is hidden while no service has a configured model, and a
+  refused save reports the error instead of silently keeping the old value.
+- **Specs linked**: `03-runtime/02-agent-runtime.md` §5f,
+  `03-runtime/11-provider-model-system.md` §7, `04-ux/06-settings-ia.md`
+  (Model configuration), D278, ADR subagent-model-opt-in, issue #1736.
+- **Acceptance**: C / E / Quality — the catalog the parent agent may pick from is
+  reviewable and editable in one place on the Models page without adding a
+  second stored selection, and every row states why an opted-in model is not in
+  effect.
+- **Milestone**: Post-MVP model configuration.
+- **Status**: Unit coverage in `apps/desktop/test/subagent-delegation.test.mjs`
+  (candidate construction, grouping, read-only and blocked states, and the exact
+  provider write). The interactive save → next-turn catalog path remains Draft
+  and is run only when requested in a capable environment.
 ### E2E-FAST-MODE: `/fast` asks for the priority tier only on an allowed model
 
 - **Preconditions**: A ChatGPT subscription (Codex) account and a Responses
diff --git a/docs/zh-CN/spec/03-runtime/02-agent-runtime.md b/docs/zh-CN/spec/03-runtime/02-agent-runtime.md
index a8694c9f99..27539006d8 100644
--- a/docs/zh-CN/spec/03-runtime/02-agent-runtime.md
+++ b/docs/zh-CN/spec/03-runtime/02-agent-runtime.md
@@ -684,8 +684,9 @@ Frontmatter 新增 `permission: inherit | ask | accept-edits | auto`（默认
   运行中覆盖该委托的模型。解析优先级：Task.model 参数 → 定义 frontmatter 的
   引脚 → 会话模型。父 agent 会在系统提示中看到一份模型摘要，列出提供商设置里
   所有标记为 `availableForSubagents` 的模型。空目录下的系统摘要、Task 说明和
-  覆盖请求错误均指向「设置 → 模型 → 编辑服务或账号 → 模型高级设置 →
-  可供 AI 自动调度 → 保存」，要求使用目录中的准确键，避免猜测 provider/model。
+  覆盖请求错误均指向设置 → 模型页的 **子代理模型** 行（或同一开关的
+  「设置 → 模型 → 编辑服务或账号 → 模型高级设置 → 可供 AI 自动调度 → 保存」），
+  要求使用目录中的准确键，避免猜测 provider/model。
   若委托目录为空，提示会告诉模型
   省略 `model`，使用定义的固定模型，无固定模型时继承会话模型；显式给出的键如果正好就是当前会话的
   provider/model，同样按继承处理。其他显式模型键必须已配置并已为委托启用。
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 d7001b5f24..37168a4fc2 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
@@ -351,6 +351,13 @@ provider id/vendor/name 查找，不得用另一账号凭据覆盖固定模型
 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。
 
+该标志的编辑入口不限于提供商对话框。设置 → 模型在同一处用一列 **子代理模型**
+列出已加入的绑定（每行的“高级”打开勾选该选项的服务模型编辑器），读取同一个字段，并只通过
+`providers.update` 重写被改动的
+那一条绑定（`apps/desktop/src/components/settings/subagent-delegation.ts`）。
+该行不保存第二份选择，因此委托目录与页面不会互相矛盾；服务已关闭、缺少凭据或由
+插件拥有时，该行会一并说明原因，而不是直接消失。
+
 ## 8. 秘密
 
 - 通过安全存储存储的 API 密钥（`SECRET_*` API）
diff --git a/docs/zh-CN/spec/04-ux/06-settings-ia.md b/docs/zh-CN/spec/04-ux/06-settings-ia.md
index a2fa800a03..1b24b16240 100644
--- a/docs/zh-CN/spec/04-ux/06-settings-ia.md
+++ b/docs/zh-CN/spec/04-ux/06-settings-ia.md
@@ -195,11 +195,19 @@ Token 用量**不是设置目的地**（D335 / ADR 0173）。已完成回合历
     `Cmd/Ctrl + Shift + W` 呼出组合键已移除 —— 并且避开 `Cmd/Ctrl + W`，因为
     macOS 把它用于自己的关闭窗口命令；读取配置映射时，已存储的
     `closeWindow` / `summonWindow` 覆盖项会并入它（D438、D439）
-
-### 模型配置（`agent` 选项卡）
-- **Studio Hero**：提供商计数、就绪计数和当前默认 provider/model 摘要
-- **默认**卡：默认提供商/模型选择器。全局运行模式、命令 Shell
-  和回车发送由全局 AI 目的地负责。
+- **子代理模型**行（位于生图模型选择器之后）只列出带 `availableForSubagents` 选项的
+  模型 —— 即父 agent 看到的委托目录 —— 每条以 `provider · model` 呈现并带删除操作；
+  模型 —— 即父 agent 看到的委托目录 —— 每条以 `provider · model` 呈现，并带与服务
+  自身模型列表相同的“高级”和删除操作；一个都没有时显示空状态说明。“高级”打开该服务
+  的模型编辑器（勾选项就在那里），因此添加始终发生在模型编辑器里，而不是这一行。它读写
+  `ModelBinding.availableForSubagents`，不另存第二份选择；保存前重新读取该服务，只
+  没有任何服务配置模型时整行隐藏。
+  `availableForSubagents` 选项的模型 —— 即父 agent 看到的委托目录；一个都没有时显示
+  空状态说明。其折叠区按服务分组列出全部已配置绑定，每个模型一个复选框，因此不必逐个
+  打开服务的模型对话框。它读写 `ModelBinding.availableForSubagents`，不另存第二份选择；
+  保存前重新读取该服务，只通过 `providers.update` 重写被改动的那条绑定，并发编辑不会被
+  覆盖。服务已关闭或缺少凭据时，已勾选的模型仍然可见可改，并注明为何暂不生效；插件拥有
+  的服务行只读（RPC 会拒绝）。没有任何服务配置模型时整行隐藏。
 - **厂商账户**卡片（D237），位于默认值与提供商之间：
   - 卡片列出的是账户而非厂商：已登录的厂商各占一行，一个也没有时用一句
     说明取代列表。运行时目录不提供任何 OAuth 厂商时整个分区隐藏
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 93bbba382f..c639a5f231 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
@@ -5920,6 +5920,7 @@ eleven-tool-round desktop paths are verified by
 ## 8. 可追溯性矩阵
 | 验收 | 应用场景 |
 |---|---|
+| C / E / Quality — 模型页的子代理委托目录 | [E2E-268](#e2e-268模型页显示并编辑子代理委托目录) |
 | C / E / Security / Quality — Gemini work setup | E2E-LIVE-VOICE-gemini-work-schema |
 | C / Quality — Continue after malformed tool output | E2E-RUNTIME-malformed-tool-call-continue-recovery |
 | B / Security / Quality — Optional provider client certificate requests | E2E-PROVIDER-optional-client-certificate |
@@ -10032,6 +10033,32 @@ host-created files. The full app's file-preview viewer is covered separately.
 - **验收**：项目折叠时仍能回答“这里是否有正在运行的工作”，且两种提示可从持久化设置独立控制。
 - **状态**：草稿。现有 `apps/desktop/test/sidebar-project-runs.test.mjs` 覆盖数量分类、总开关与子开关门控、两种渲染入口，以及主题令牌和减少动态效果保证；它不覆盖设置 API 缺省值规范化或完整交互，相关场景仍需在可用环境手动验证。
 
+### E2E-268：模型页显示并编辑子代理委托目录
+
+- **前提条件**：两个 AI 服务已配置、已启用并持有凭据，第一个有两个模型、第二个有一个；
+  仅其中一个模型勾选了“可供 AI 自动调度”。可委派任务的父 Agent 会话可用。
+- **步骤**：
+  1. 打开设置 → 模型，阅读底部的 **子代理模型** 行。
+  2. 用该行的“高级”打开第二个服务的模型编辑器，勾选“可供 AI 自动调度”并保存，再看这一行。
+  3. 发起一次父会话回合，检查 Task 工具收到的委托模型摘要。
+  4. 回到该行用 × 移除该模型，再看这一行。
+  5. 关闭第二个服务，再看这一行。
+  6. 在第一个服务自己的模型对话框中切换“可供 AI 自动调度”，返回模型页。
+- **预期**：该行列出的正是已加入的模型（`provider · model`），一个都没有时显示空状态说明；
+  模型只能通过在该模型编辑器里勾选来加入。加入会把 `availableForSubagents` 写入该服务的
+  绑定，下一次父会话回合即可在委托目录中看到该模型；移除会把它从目录移出，且不在已存绑定上
+  留下 `false`。服务已关闭或缺少凭据时，已加入的模型仍然列出并可移除，并注明为何暂不生效，
+  同时其模型不进入目录。服务对话框与这一行双向显示同一开关，两个界面不会各说各话。插件拥有的
+  行只读。没有任何服务配置模型时整行隐藏；保存被拒绝时报告错误，而不是静默保留旧值。
+- **关联规格**：`03-runtime/02-agent-runtime.md` §5f、
+  `03-runtime/11-provider-model-system.md` §7、`04-ux/06-settings-ia.md`（模型配置）、
+  D278、ADR subagent-model-opt-in、issue #1736。
+- **验收**：C / E / Quality —— 父 agent 可挑选的委托目录能在模型页一处查看并修改，且不新增
+  第二份保存的选择；每一行都说明某个已勾选模型为何暂不生效。
+- **里程碑**：Post-MVP 模型配置。
+- **状态**：`apps/desktop/test/subagent-delegation.test.mjs` 有单元覆盖（候选构造、分组、
+  只读与受阻状态、以及精确的提供商写入）。交互式的“保存 → 下一回合目录”路径仍为草稿，
+  仅在明确请求且环境可用时运行。
 ### E2E-FAST-MODE：`/fast` 只在允许的模型上请求优先档位
 
 - **前置条件**：一个 ChatGPT 订阅（Codex）账号和一个 Responses API key 提供商，各配置两个模型。
diff --git a/packages/i18n/src/locales/de/index.ts b/packages/i18n/src/locales/de/index.ts
index 8ecf62369c..f2f9b0db36 100644
--- a/packages/i18n/src/locales/de/index.ts
+++ b/packages/i18n/src/locales/de/index.ts
@@ -1796,6 +1796,18 @@ sklm: {
     "contextWindowCatalogHint": "Folgt models.dev; eine Änderung fixiert Ihren Wert.",
     "availableForSubagents": "Verfügbar für AI-Delegation",
     "availableForSubagentsHint": "Ermöglichen Sie AI, dieses Modell zu verwenden, wenn Aufgaben an Subagenten delegiert werden.",
+    "subagentModels": "Subagenten-Modelle",
+    "subagentModelsUnset": "Noch kein Modell für Subagenten freigegeben",
+    "subagentModelAdd": "Modelle in den Modelleinstellungen des Dienstes hinzufügen — Erweitert → Verfügbar für AI-Delegation",
+    "subagentModelEdit": "Bearbeiten",
+    "subagentModelsMore": "+{{count}} weitere",
+    "subagentModelRemove": "{{model}} aus den Subagenten-Modellen entfernen",
+    "subagentModelSelected": "Modell für Delegation verfügbar",
+    "subagentModelUnselected": "Modell aus der Delegation entfernt",
+    "subagentModelSaveFailed": "Auswahl für die Delegation konnte nicht gespeichert werden",
+    "subagentModelPluginManaged": "Von einem Plugin verwaltet",
+    "subagentModelProviderDisabled": "Dienst ist ausgeschaltet",
+    "subagentModelCredentialMissing": "Kein Zugangsschlüssel",
     "nativeWebSearch": "Native Websuche",
     "nativeWebSearchHint": "Hängt das vom Anbieter gehostete Websuche-Tool an Anfragen für dieses Modell an. Erfordert einen Endpunkt, der dies unterstützt; die Abrechnung erfolgt pro Suche beim Anbieter.",
     nativeWebSearchUnsupported: "Diese Verbindung unterstützt in der App noch keine native Suche. Die Website oder eine andere API kann Suche anbieten.",
diff --git a/packages/i18n/src/locales/en/index.ts b/packages/i18n/src/locales/en/index.ts
index 54eb8cbf3c..548f07dda0 100644
--- a/packages/i18n/src/locales/en/index.ts
+++ b/packages/i18n/src/locales/en/index.ts
@@ -1816,6 +1816,18 @@ sklm: {
     contextWindowCatalogHint: "Follows models.dev; editing pins your value.",
     availableForSubagents: "Available for AI delegation",
     availableForSubagentsHint: "Allow AI to use this model when delegating tasks to subagents",
+    subagentModels: "Subagent models",
+    subagentModelsUnset: "No model is open to subagents yet",
+    subagentModelAdd: "Add models from a service's model settings — Advanced → Available for AI delegation",
+    subagentModelEdit: "Edit",
+    subagentModelsMore: "+{{count}} more",
+    subagentModelRemove: "Remove {{model}} from subagent models",
+    subagentModelSelected: "Model available for delegation",
+    subagentModelUnselected: "Model removed from delegation",
+    subagentModelSaveFailed: "Could not save the delegation choice",
+    subagentModelPluginManaged: "Managed by a plugin",
+    subagentModelProviderDisabled: "Service is off",
+    subagentModelCredentialMissing: "No credential",
     nativeWebSearch: "Native web search",
     nativeWebSearchHint: "Attach the provider-hosted web search tool to requests for this model. Requires an endpoint that supports it; billed by the provider per search.",
     nativeWebSearchUnsupported: "This app has not integrated native search for this connection. Website search or another API may still be available.",
diff --git a/packages/i18n/src/locales/es/index.ts b/packages/i18n/src/locales/es/index.ts
index 40c01094a7..67a5817782 100644
--- a/packages/i18n/src/locales/es/index.ts
+++ b/packages/i18n/src/locales/es/index.ts
@@ -1796,6 +1796,18 @@ sklm: {
     "contextWindowCatalogHint": "Sigue a models.dev; al editarlo se fija tu valor.",
     "availableForSubagents": "Disponible para delegación de AI",
     "availableForSubagentsHint": "Permitir que AI use este modelo al delegar tareas a subagentes",
+    "subagentModels": "Modelos de subagentes",
+    "subagentModelsUnset": "Aún no hay ningún modelo abierto a subagentes",
+    "subagentModelAdd": "Añade modelos en los ajustes de modelos del servicio — Avanzado → Disponible para delegación de AI",
+    "subagentModelEdit": "Editar",
+    "subagentModelsMore": "+{{count}} más",
+    "subagentModelRemove": "Quitar {{model}} de los modelos de subagente",
+    "subagentModelSelected": "Modelo disponible para delegación",
+    "subagentModelUnselected": "Modelo retirado de la delegación",
+    "subagentModelSaveFailed": "No se pudo guardar la elección de delegación",
+    "subagentModelPluginManaged": "Gestionado por un complemento",
+    "subagentModelProviderDisabled": "El servicio está desactivado",
+    "subagentModelCredentialMissing": "Sin credencial",
     "nativeWebSearch": "Búsqueda web nativa",
     "nativeWebSearchHint": "Adjunta la herramienta de búsqueda web alojada por el proveedor a las solicitudes de este modelo. Requiere un punto de conexión que la admita; el proveedor la factura por búsqueda.",
     nativeWebSearchUnsupported: "La aplicación aún no integra búsqueda nativa para esta conexión. El sitio web u otra API pueden ofrecerla.",
diff --git a/packages/i18n/src/locales/fr/index.ts b/packages/i18n/src/locales/fr/index.ts
index 723e5ba265..e2304efb8d 100644
--- a/packages/i18n/src/locales/fr/index.ts
+++ b/packages/i18n/src/locales/fr/index.ts
@@ -1796,6 +1796,18 @@ sklm: {
     "contextWindowCatalogHint": "Suit models.dev ; une modification fixe votre valeur.",
     "availableForSubagents": "Disponible pour la délégation de l'IA",
     "availableForSubagentsHint": "Autoriser l'IA à utiliser ce modèle lors de la délégation de tâches à des sous-agents",
+    "subagentModels": "Modèles de sous-agents",
+    "subagentModelsUnset": "Aucun modèle ouvert aux sous-agents pour l'instant",
+    "subagentModelAdd": "Ajoutez des modèles dans les réglages du service — Avancé → Disponible pour la délégation de l'IA",
+    "subagentModelEdit": "Modifier",
+    "subagentModelsMore": "+{{count}} de plus",
+    "subagentModelRemove": "Retirer {{model}} des modèles de sous-agent",
+    "subagentModelSelected": "Modèle disponible pour la délégation",
+    "subagentModelUnselected": "Modèle retiré de la délégation",
+    "subagentModelSaveFailed": "Impossible d'enregistrer le choix de délégation",
+    "subagentModelPluginManaged": "Géré par un plugin",
+    "subagentModelProviderDisabled": "Le service est désactivé",
+    "subagentModelCredentialMissing": "Aucun justificatif",
     "nativeWebSearch": "Recherche web native",
     "nativeWebSearchHint": "Attache l'outil de recherche web hébergé par le fournisseur aux requêtes pour ce modèle. Nécessite un point de terminaison qui le prend en charge ; facturé par le fournisseur à chaque recherche.",
     nativeWebSearchUnsupported: "La recherche native n’est pas encore intégrée pour cette connexion. Le site web ou une autre API peut la proposer.",
diff --git a/packages/i18n/src/locales/ko/index.ts b/packages/i18n/src/locales/ko/index.ts
index fd95c8096b..d9c908f202 100644
--- a/packages/i18n/src/locales/ko/index.ts
+++ b/packages/i18n/src/locales/ko/index.ts
@@ -1812,6 +1812,18 @@ sklm: {
     contextWindowCatalogHint: "models.dev를 따릅니다. 수정하면 내 값으로 고정됩니다.",
     availableForSubagents: "AI 위임에 사용 가능",
     availableForSubagentsHint: "서브에이전트에 작업을 위임할 때 AI가 이 모델을 사용하도록 허용",
+    subagentModels: "서브에이전트 모델",
+    subagentModelsUnset: "아직 서브에이전트에 열린 모델이 없습니다",
+    subagentModelAdd: "서비스 모델 설정에서 추가하세요 — 고급 → AI 위임에 사용 가능",
+    subagentModelEdit: "편집",
+    subagentModelsMore: "외 {{count}}개",
+    subagentModelRemove: "{{model}}을(를) 서브에이전트 모델에서 제거",
+    subagentModelSelected: "위임에 사용할 수 있는 모델",
+    subagentModelUnselected: "위임 목록에서 제거된 모델",
+    subagentModelSaveFailed: "위임 선택을 저장하지 못했습니다",
+    subagentModelPluginManaged: "플러그인이 관리함",
+    subagentModelProviderDisabled: "서비스가 꺼져 있음",
+    subagentModelCredentialMissing: "자격 증명 없음",
     nativeWebSearch: "네이티브 웹 검색",
     nativeWebSearchHint: "이 모델에 대한 요청에 공급자 호스팅 웹 검색 도구를 첨부합니다. 이를 지원하는 엔드포인트가 필요하며, 검색 건수로 공급자에게 과금됩니다.",
     nativeWebSearchUnsupported: "이 연결의 네이티브 검색은 아직 앱에 통합되지 않았습니다. 웹사이트나 다른 API에서는 검색을 지원할 수 있습니다.",
diff --git a/packages/i18n/src/locales/pt-BR/index.ts b/packages/i18n/src/locales/pt-BR/index.ts
index 12a50fe18a..8b7a19a7b9 100644
--- a/packages/i18n/src/locales/pt-BR/index.ts
+++ b/packages/i18n/src/locales/pt-BR/index.ts
@@ -1766,6 +1766,18 @@ export const ptBR = {
     contextWindowCatalogHint: "Segue models.dev; editar fixa o seu valor.",
     availableForSubagents: "Disponível para subagentes",
     availableForSubagentsHint: "Permite que a IA use este modelo ao delegar tarefas a subagentes",
+    subagentModels: "Modelos de subagentes",
+    subagentModelsUnset: "Nenhum modelo aberto para subagentes ainda",
+    subagentModelAdd: "Adicione modelos nas configurações de modelos do serviço — Avançado → Disponível para delegação",
+    subagentModelEdit: "Editar",
+    subagentModelsMore: "+{{count}} mais",
+    subagentModelRemove: "Remover {{model}} dos modelos de subagente",
+    subagentModelSelected: "Modelo disponível para delegação",
+    subagentModelUnselected: "Modelo removido da delegação",
+    subagentModelSaveFailed: "Não foi possível salvar a escolha de delegação",
+    subagentModelPluginManaged: "Gerenciado por um plugin",
+    subagentModelProviderDisabled: "O serviço está desligado",
+    subagentModelCredentialMissing: "Sem credencial",
     nativeWebSearch: "Pesquisa nativa na web",
     nativeWebSearchHint: "Anexa às solicitações deste modelo a ferramenta de pesquisa na web do provedor. Exige um endpoint compatível; o provedor cobra por pesquisa.",
     nativeWebSearchUnsupported: "O aplicativo ainda não integra pesquisa nativa para esta conexão. O site ou outra API ainda pode oferecer pesquisa.",
diff --git a/packages/i18n/src/locales/tr/index.ts b/packages/i18n/src/locales/tr/index.ts
index 8f89672c7d..abbe7c5a3c 100644
--- a/packages/i18n/src/locales/tr/index.ts
+++ b/packages/i18n/src/locales/tr/index.ts
@@ -1806,6 +1806,18 @@ sklm: {
     contextWindowCatalogHint: "models.dev'i takip eder; düzenlerseniz değeriniz sabitlenir.",
     availableForSubagents: "AI devri için kullanılabilir",
     availableForSubagentsHint: "Görevler alt ajanlara devredilirken AI’nin bu modeli kullanmasına izin ver",
+    subagentModels: "Alt ajan modelleri",
+    subagentModelsUnset: "Henüz alt ajanlara açılmış model yok",
+    subagentModelAdd: "Modelleri hizmetin model ayarlarından ekleyin — Gelişmiş → AI devri için kullanılabilir",
+    subagentModelEdit: "Düzenle",
+    subagentModelsMore: "+{{count}} tane daha",
+    subagentModelRemove: "{{model}} öğesini alt ajan modellerinden çıkar",
+    subagentModelSelected: "Model devretmeye açık",
+    subagentModelUnselected: "Model devretme listesinden çıkarıldı",
+    subagentModelSaveFailed: "Devretme seçimi kaydedilemedi",
+    subagentModelPluginManaged: "Bir eklenti tarafından yönetiliyor",
+    subagentModelProviderDisabled: "Hizmet kapalı",
+    subagentModelCredentialMissing: "Kimlik bilgisi yok",
     nativeWebSearch: "Yerel web araması",
     nativeWebSearchHint: "Bu modele yapılan isteklere sağlayıcının barındırdığı web arama aracını ekler. Bunu destekleyen bir uç nokta gerekir; arama başına sağlayıcı tarafından faturalandırılır.",
     nativeWebSearchUnsupported: "Bu bağlantı için yerel arama henüz uygulamaya entegre edilmedi. Web sitesi veya başka bir API aramayı destekleyebilir.",
diff --git a/packages/i18n/src/locales/zh-CN/index.ts b/packages/i18n/src/locales/zh-CN/index.ts
index b8602e172a..33b61c4919 100644
--- a/packages/i18n/src/locales/zh-CN/index.ts
+++ b/packages/i18n/src/locales/zh-CN/index.ts
@@ -1791,6 +1791,18 @@ sklm: {
     contextWindowCatalogHint: "跟随 models.dev；手动修改后会固定为你的值",
     availableForSubagents: "可供 AI 自动调度",
     availableForSubagentsHint: "允许 AI 在委派子任务时自动选用此模型",
+    subagentModels: "子代理模型",
+    subagentModelsUnset: "还没有可供子代理调用的模型",
+    subagentModelAdd: "在服务的模型设置里添加：高级 → 可供 AI 自动调度",
+    subagentModelEdit: "编辑",
+    subagentModelsMore: "另有 {{count}} 个",
+    subagentModelRemove: "将 {{model}} 从子代理模型移除",
+    subagentModelSelected: "该模型可供 AI 调度",
+    subagentModelUnselected: "该模型已从调度列表移除",
+    subagentModelSaveFailed: "保存子代理模型选择失败",
+    subagentModelPluginManaged: "由插件管理",
+    subagentModelProviderDisabled: "服务已关闭",
+    subagentModelCredentialMissing: "尚未配置凭据",
     nativeWebSearch: "原生联网搜索",
     nativeWebSearchHint: "为该模型的请求附加提供商托管的联网搜索工具。需要端点实际支持；按次计费由提供商收取。",
     nativeWebSearchUnsupported: "应用尚未适配此连接的原生搜索；不代表厂商官网或其他 API 不支持搜索。",
diff --git a/packages/i18n/src/locales/zh-TW/index.ts b/packages/i18n/src/locales/zh-TW/index.ts
index 9483adbce6..d3460c34d3 100644
--- a/packages/i18n/src/locales/zh-TW/index.ts
+++ b/packages/i18n/src/locales/zh-TW/index.ts
@@ -1791,6 +1791,18 @@ sklm: {
     contextWindowCatalogHint: "跟隨 models.dev；手動修改後會固定為你的值",
     availableForSubagents: "可供 AI 自動排程",
     availableForSubagentsHint: "允許 AI 在委派子任務時自動選用此模型",
+    subagentModels: "子代理模型",
+    subagentModelsUnset: "尚無可提供子代理呼叫的模型",
+    subagentModelAdd: "在服務的模型設定裡新增：進階 → 可供 AI 自動調度",
+    subagentModelEdit: "編輯",
+    subagentModelsMore: "另有 {{count}} 個",
+    subagentModelRemove: "將 {{model}} 從子代理模型移除",
+    subagentModelSelected: "此模型可供 AI 調度",
+    subagentModelUnselected: "此模型已從調度清單移除",
+    subagentModelSaveFailed: "儲存子代理模型選擇失敗",
+    subagentModelPluginManaged: "由外掛管理",
+    subagentModelProviderDisabled: "服務已關閉",
+    subagentModelCredentialMissing: "尚未設定憑證",
     nativeWebSearch: "原生網路搜尋",
     nativeWebSearchHint: "為該模型的請求附加提供商託管的網路搜尋工具。需要端點實際支援；按次計費由提供商收取。",
     nativeWebSearchUnsupported: "應用程式尚未整合此連線的原生搜尋；不代表廠商官網或其他 API 不支援搜尋。",