gui+server: subagent prompt injection-model pick (API + dashboard select)

- /api/injection-model GET/PUT: pick one routed/native model whose info is
  injected into the v1 ultra proactive prompt; disabled models filtered
- dashboard injection-model select + i18n (en/ko/zh); server wiring
- tests: multi-agent-compat coverage for the injection pick (16 pass)

Author: lidge-jun

gui/src/i18n/en.ts

@@ -52,6 +52,10 @@ export const en = {
   "dash.visionModelHint": "Model used to describe images for text-only routed models. Requires ChatGPT login.",
   "dash.sidecarSaved": "Sidecar settings saved. Applied on the next request.",
   "dash.sidecarSaveFailed": "Failed to save sidecar settings.",
+  "dash.injectionLabel": "Sub-agent delegation",
+  "dash.injectionHint": "Pick a routed model to inject into the delegation prompt. The agent will be told to use it for sub-tasks.",
+  "dash.injectionActive": "Active",
+  "dash.injectionNone": "None",
   "dash.maintenance": "Maintenance",
   "dash.maintenanceHint": "Refresh Codex's model catalog or install a newer opencodex release.",
   "dash.syncModels": "Sync models",

gui/src/i18n/ko.ts

@@ -52,7 +52,11 @@ export const ko: Record<TKey, string> = {
   "dash.visionModelHint": "텍스트 전용 라우팅 모델에 이미지를 설명하는 데 사용되는 모델입니다. ChatGPT 로그인 필요.",
   "dash.sidecarSaved": "사이드카 설정이 저장됐습니다. 다음 요청부터 적용됩니다.",
   "dash.sidecarSaveFailed": "사이드카 설정 저장에 실패했습니다.",
-  "dash.maintenance": "유지보수",
+  "dash.injectionLabel": "서브에이전트 위임",
+  "dash.injectionHint": "위임 프롬프트에 주입할 라우팅 모델을 선택합니다. 에이전트가 서브태스크에 이 모델을 사용하도록 안내됩니다.",
+  "dash.injectionActive": "활성",
+  "dash.injectionNone": "없음",
+    "dash.maintenance": "유지보수",
   "dash.maintenanceHint": "Codex 모델 카탈로그를 새로고침하거나 최신 opencodex 릴리스를 설치합니다.",
   "dash.syncModels": "모델 동기화",
   "dash.syncing": "동기화 중…",

gui/src/i18n/zh.ts

@@ -52,7 +52,11 @@ export const zh: Record<TKey, string> = {
   "dash.visionModelHint": "为纯文本路由模型描述图像的模型。需要 ChatGPT 登录。",
   "dash.sidecarSaved": "附属设置已保存。将在下一个请求时生效。",
   "dash.sidecarSaveFailed": "保存附属设置失败。",
-  "dash.maintenance": "维护",
+  "dash.injectionLabel": "子代理委托",
+  "dash.injectionHint": "选择要注入委托提示的路由模型。代理将被告知在子任务中使用它。",
+  "dash.injectionActive": "已激活",
+  "dash.injectionNone": "无",
+    "dash.maintenance": "维护",
   "dash.maintenanceHint": "刷新 Codex 模型目录，或安装新的 opencodex 版本。",
   "dash.syncModels": "同步模型",
   "dash.syncing": "同步中…",

gui/src/pages/Dashboard.tsx

@@ -86,6 +86,9 @@ export default function Dashboard({ apiBase }: { apiBase: string }) {
   const [maMode, setMaMode] = useState<"v1" | "default" | "v2">("default");
   const [maBusy, setMaBusy] = useState(false);
   const [maHelpOpen, setMaHelpOpen] = useState(false);
+  const [injectionModel, setInjectionModel] = useState<string>("");
+  const [injectionAvailable, setInjectionAvailable] = useState<Array<{ provider: string; model: string; namespaced: string }>>([]);
+  const [injectionSaving, setInjectionSaving] = useState(false);
   const [syncResult, setSyncResult] = useState<SyncResult | null>(null);
   const [syncError, setSyncError] = useState<string | null>(null);
   const [projectConfigWarnings, setProjectConfigWarnings] = useState<ProjectCodexConfigGroup[]>([]);
@@ -124,6 +127,14 @@ export default function Dashboard({ apiBase }: { apiBase: string }) {
             else setMaMode("default");
           }
         } catch { /* old server */ }
+        try {
+          const imRes = await fetch(`${apiBase}/api/injection-model`);
+          if (imRes.ok) {
+            const imData = await imRes.json() as { model?: string | null; available?: Array<{ provider: string; model: string; namespaced: string }> };
+            setInjectionModel(imData.model ?? "");
+            setInjectionAvailable(imData.available ?? []);
+          }
+        } catch { /* old server */ }
       } catch {
         setError(true);
       }
@@ -422,6 +433,44 @@ export default function Dashboard({ apiBase }: { apiBase: string }) {
         </div>
       )}
 
+      <div className="panel" style={{ marginBottom: 24 }}>
+        <div className="spread setting-row" style={{ alignItems: "flex-start" }}>
+          <div className="setting-copy" style={{ flex: 1 }}>
+            <div style={{ display: "flex", alignItems: "center", gap: 8 }}>
+              <span style={{ fontWeight: 650 }}>{t("dash.injectionLabel")}</span>
+              {injectionModel && <span className="badge badge-green" style={{ fontSize: 10 }}>{t("dash.injectionActive")}</span>}
+            </div>
+            <div className="muted setting-hint">{t("dash.injectionHint")}</div>
+          </div>
+          <Select
+            value={injectionModel}
+            options={[
+              { value: "", label: t("dash.injectionNone") },
+              ...injectionAvailable.map(m => ({ value: m.namespaced, label: `${m.provider} / ${m.model}` })),
+            ]}
+            onChange={async (v) => {
+              if (injectionSaving) return;
+              setInjectionSaving(true);
+              setInjectionModel(v);
+              try {
+                const res = await fetch(`${apiBase}/api/injection-model`, {
+                  method: "PUT",
+                  headers: { "Content-Type": "application/json" },
+                  body: JSON.stringify({ model: v || null }),
+                });
+                if (res.ok) {
+                  const data = await res.json() as { model?: string | null };
+                  setInjectionModel(data.model ?? "");
+                }
+              } catch { /* ignore */ }
+              finally { setInjectionSaving(false); }
+            }}
+            disabled={injectionSaving}
+            label={t("dash.injectionLabel")}
+          />
+        </div>
+      </div>
+
       <div className="panel maintenance-panel" style={{ marginBottom: 24 }}>
         <div className="spread maintenance-head">
           <div>

gui/src/ui.tsx

@@ -23,13 +23,15 @@ export function Notice({ tone, children }: { tone: "ok" | "err"; children: React
 
 export interface SelectOption { value: string; label: React.ReactNode }
 
-export function Select({ value, options, onChange, disabled, label, style }: {
+export function Select({ value, options, onChange, disabled, label, style, align, dropdownStyle }: {
   value: string;
   options: SelectOption[];
   onChange: (value: string) => void;
   disabled?: boolean;
   label?: string;
   style?: CSSProperties;
+  align?: "left" | "right";
+  dropdownStyle?: CSSProperties;
 }) {
   const [open, setOpen] = useState(false);
   const ref = useRef<HTMLDivElement>(null);
@@ -59,7 +61,7 @@ export function Select({ value, options, onChange, disabled, label, style }: {
         <IconChevron style={{ width: 12, height: 12, color: "var(--muted)", transform: open ? "rotate(90deg)" : "none", transition: "transform .12s" }} />
       </button>
       {open && (
-        <div className="select-dropdown" role="listbox" aria-label={label}>
+        <div className="select-dropdown" role="listbox" aria-label={label} style={{ ...(align === "right" ? { left: "auto", right: 0 } : {}), ...dropdownStyle }}>
           {options.map(o => (
             <button
               key={o.value}

src/server/index.ts

@@ -158,11 +158,11 @@ export function startServer(port?: number) {
     config.subagentModels = [...DEFAULT_SUBAGENT_MODELS];
     saveConfig(config);
   }
-  // Sidecar model migration (KST 2026-06-10 06:00 = UTC 2026-06-09 21:00): auto-migrate the old
+  // Sidecar model migration (KST 2026-07-10 06:00 = UTC 2026-07-09 21:00): auto-migrate the old
   // gpt-5.4-mini default to gpt-5.6-luna for both search and vision sidecars. Only touches configs
   // still on the old default — explicit user choices are preserved.
   {
-    const SIDECAR_MIGRATION_CUTOFF = Date.UTC(2026, 5, 9, 21, 0); // June 9 21:00 UTC = KST June 10 06:00
+    const SIDECAR_MIGRATION_CUTOFF = Date.UTC(2026, 6, 9, 21, 0); // July 9 21:00 UTC = KST July 10 06:00
     if (Date.now() >= SIDECAR_MIGRATION_CUTOFF) {
       let migrated = false;
       if (config.webSearchSidecar?.model === "gpt-5.4-mini") {

src/server/management-api.ts

@@ -517,6 +517,31 @@ export async function handleManagementAPI(req: Request, url: URL, config: OcxCon
     return jsonResponse({ providers: deriveProviderPresets() });
   }
 
+  // Subagent prompt injection model: single routed model whose info is dynamically
+  // injected into the v1 ultra proactive prompt. GET returns the current pick + available
+  // routed models; PUT sets or clears the pick.
+  if (url.pathname === "/api/injection-model" && req.method === "GET") {
+    const models = await fetchAllModels(config);
+    const disabled = new Set(config.disabledModels ?? []);
+    const { listCatalogNativeSlugs } = await import("../codex/catalog");
+    const nativeModels = listCatalogNativeSlugs()
+      .filter(slug => !disabled.has(slug))
+      .map(slug => ({ provider: "openai", model: slug, namespaced: slug }));
+    const routedModels = models
+      .map(m => ({ provider: m.provider, model: m.id, namespaced: `${m.provider}/${m.id}` }))
+      .filter(m => !disabled.has(m.namespaced));
+    return jsonResponse({ model: config.injectionModel ?? null, available: [...nativeModels, ...routedModels] });
+  }
+  if (url.pathname === "/api/injection-model" && req.method === "PUT") {
+    let body: { model?: unknown };
+    try { body = await req.json(); } catch { return jsonResponse({ error: "invalid JSON body" }, 400); }
+    const model = typeof body.model === "string" && body.model.length > 0 ? body.model : undefined;
+    if (model) config.injectionModel = model;
+    else delete config.injectionModel;
+    saveConfig(config);
+    return jsonResponse({ ok: true, model: config.injectionModel ?? null });
+  }
+
   // Subagent model picker: which ≤5 routed models Codex's spawn_agent advertises (it shows the
   // first 5 routed catalog entries). PUT reorders the injected catalog so the chosen ones lead.
   if (url.pathname === "/api/subagent-models" && req.method === "GET") {

src/server/responses.ts

@@ -75,8 +75,13 @@ export function buildToolBridgeMaps(parsed: OcxParsedRequest): {
 }
 
 /** Verbatim upstream Proactive text (codex-rs core/src/context/multi_agent_mode_instructions.rs). */
-const PROACTIVE_MULTI_AGENT_MODE_TEXT =
-  "Proactive multi-agent delegation is active. Any earlier instruction requiring an explicit user request before spawning sub-agents no longer applies. Use sub-agents when parallel work would materially improve speed or quality. This mode remains active until a later multi-agent mode developer message changes it.";
+const PROACTIVE_MULTI_AGENT_MODE_TEXT = [
+  "Proactive multi-agent delegation is active.",
+  "Any earlier instruction requiring an explicit user request before spawning sub-agents no longer applies.",
+  "Delegate independent sub-tasks to sub-agents whenever parallel work would materially improve speed or quality — do not serialize work that can run concurrently.",
+  "Each sub-agent runs in its own context and can use all available tools; prefer spawning specialists over doing everything yourself.",
+  "This mode remains active until a later multi-agent mode developer message changes it.",
+].join(" ");
 
 /**
  * True when this turn runs the v1 collab surface, judged from the request's own tool list
@@ -98,19 +103,39 @@ export function isV1CollabSurface(parsed: OcxParsedRequest): boolean {
 }
 
 /**
- * Multi-agent guidance for this turn, or null when nothing applies. codex-rs only emits
- * its Proactive delegation developer message on the v2 surface, so when a v1-surface turn
- * arrives at the synthetic top tier (codex converts ultra -> max on the wire, so max
- * arrival means the user picked the top rung) the proxy supplies the same one-liner,
- * wrapped in codex's own <multi_agent_mode> tags (v1 turns never carry that fragment, so
- * there is nothing to collide with). Ultra is always advertised, so the guidance fires
- * regardless of the multi_agent_v2 toggle.
+ * Multi-agent guidance for this turn, or null when nothing applies.
+ *
+ * codex-rs only emits its Proactive delegation developer message on the v2 surface,
+ * so when a v1-surface turn arrives at the synthetic top tier (codex converts
+ * ultra -> max on the wire, so max arrival means the user picked the top rung) the
+ * proxy supplies the same one-liner, wrapped in codex's own <multi_agent_mode> tags
+ * (v1 turns never carry that fragment, so there is nothing to collide with).
+ * Ultra is always advertised, so the guidance fires regardless of the multi_agent_v2
+ * toggle.
+ *
+ * Dynamic model injection: when the user has configured a specific injectionModel,
+ * the prompt names it so the agent knows which routed model to delegate to.
+ *
+ * Effort gate relaxation: when an injectionModel is set, the prompt fires at every
+ * effort level, not just max/ultra — the user opted into delegation.
  */
-export async function multiAgentGuidanceText(parsed: OcxParsedRequest): Promise<string | null> {
+export async function multiAgentGuidanceText(parsed: OcxParsedRequest, injectionModel?: string): Promise<string | null> {
   if (!isV1CollabSurface(parsed)) return null;
   const effort = parsed.options.reasoning;
-  if (effort !== "max" && effort !== "ultra") return null;
-  return `<multi_agent_mode>${PROACTIVE_MULTI_AGENT_MODE_TEXT}</multi_agent_mode>`;
+  // When the user has selected a specific injection model, fire the delegation prompt
+  // at ANY effort level. Otherwise preserve the original gate: top tier only (max/ultra).
+  if (!injectionModel && effort !== "max" && effort !== "ultra") return null;
+
+  let text = PROACTIVE_MULTI_AGENT_MODE_TEXT;
+
+  // Append the selected model when the user has configured a specific injection target.
+  if (injectionModel) {
+    text += `\n\nA routed model is configured as the preferred sub-agent: "${injectionModel}". `
+      + "To delegate work to it, use spawn_agent with this model id. "
+      + "Use it for tasks that benefit from this model's strengths.";
+  }
+
+  return `<multi_agent_mode>${text}</multi_agent_mode>`;
 }
 
 /**
@@ -322,7 +347,7 @@ export async function handleResponses(
       const rewritten = sanitizeEncryptedContentInPlace(raw?.input);
       if (rewritten > 0) console.warn(`[opencodex] ${route.modelId}: rewrote ${rewritten} plaintext encrypted_content part(s) to input_text (routed-parent spawn compatibility)`);
     }
-    const guidance = await multiAgentGuidanceText(parsed);
+    const guidance = await multiAgentGuidanceText(parsed, config.injectionModel);
     if (guidance) injectDeveloperMessage(parsed, guidance);
   }
 

src/types.ts

@@ -244,6 +244,7 @@ export interface OcxConfig {
    * Codex's spawn_agent only advertises the first 5 routed models, so this picks which 5 appear.
    */
   subagentModels?: string[];
+  injectionModel?: string;
   /**
    * Models hidden from Codex. Routed ids are namespaced ("<provider>/<model>") and are excluded
    * from the catalog + /v1/models entirely. BARE ids (no "/") are native GPT passthrough slugs:

tests/multi-agent-compat.test.ts

@@ -77,6 +77,53 @@ describe("multiAgentGuidanceText", () => {
     }));
     expect(text).toContain("<multi_agent_mode>");
   });
+
+  test("injectionModel is named in the dynamic section", async () => {
+    codexHomeFixture(V2_OFF);
+    const text = await multiAgentGuidanceText(
+      parsedFixture({ reasoning: "max", tools: [{ name: "spawn_agent", namespace: "agents" }] }),
+      "anthropic/claude-sonnet-5",
+    );
+    expect(text).toContain("Proactive multi-agent delegation is active");
+    expect(text).toContain('"anthropic/claude-sonnet-5"');
+    expect(text).toContain("spawn_agent");
+  });
+
+  test("no injectionModel produces base prompt only at max", async () => {
+    codexHomeFixture(V2_OFF);
+    const text = await multiAgentGuidanceText(
+      parsedFixture({ reasoning: "max", tools: [{ name: "spawn_agent", namespace: "agents" }] }),
+    );
+    expect(text).toContain("Proactive multi-agent delegation is active");
+    expect(text).not.toContain("routed model");
+  });
+
+  test("injectionModel fires prompt even at low effort", async () => {
+    codexHomeFixture(V2_OFF);
+    const text = await multiAgentGuidanceText(
+      parsedFixture({ reasoning: "high", tools: [{ name: "spawn_agent", namespace: "agents" }] }),
+      "opencode-go/glm-5.2",
+    );
+    expect(text).toContain("Proactive multi-agent delegation is active");
+    expect(text).toContain('"opencode-go/glm-5.2"');
+  });
+
+  test("injectionModel fires prompt even without effort set", async () => {
+    codexHomeFixture(V2_OFF);
+    const text = await multiAgentGuidanceText(
+      parsedFixture({ tools: [{ name: "spawn_agent", namespace: "agents" }] }),
+      "anthropic/claude-opus-4-6",
+    );
+    expect(text).toContain("Proactive multi-agent delegation is active");
+  });
+
+  test("without injectionModel, low effort stays silent", async () => {
+    codexHomeFixture(V2_OFF);
+    const v1Tools = [{ name: "spawn_agent", namespace: "agents" }];
+    expect(await multiAgentGuidanceText(parsedFixture({ reasoning: "high", tools: v1Tools }))).toBeNull();
+    expect(await multiAgentGuidanceText(parsedFixture({ reasoning: "medium", tools: v1Tools }))).toBeNull();
+    expect(await multiAgentGuidanceText(parsedFixture({ reasoning: "max", tools: v1Tools }))).not.toBeNull();
+  });
 });
 
 describe("injectDeveloperMessage", () => {