fix(subagents): inherit parent model when recommended model is an auto router alias

Builtin subagents like history-analyzer pin 'model: auto' in their
frontmatter. resolveSubagentModel treated any non-inherit recommended
model as an explicit pin that beats parent inheritance, so subagents
spawned by /init ran on the server-side auto router's pick instead of
the session's selected model, contradicting the Task tool's documented
'inherits from parent' contract.

Treat a recommended model that resolves to a router alias (letta/auto,
letta/auto-fast, letta/auto-chat, letta/auto-memory) as non-pinning:
prefer the parent's model when one exists. The alias still applies when
no parent model is set, free-tier defaults are unchanged, and the
reflection carve-out is untouched.

Fixes #3248

Author: abhay-codes07

src/agent/subagent-model-resolution.test.ts

@@ -644,6 +644,46 @@ describe("resolveSubagentModel", () => {
     expect(result).toBe("anthropic/test-model");
   });
 
+  test("recommended auto router alias inherits parent model instead", async () => {
+    const result = await resolveSubagentModel({
+      recommendedModel: "auto",
+      parentModelHandle: "openai/gpt-5.4-mini",
+      availableHandles: new Set(["letta/auto", "openai/gpt-5.4-mini"]),
+    });
+
+    expect(result).toBe("openai/gpt-5.4-mini");
+  });
+
+  test("recommended auto-fast router alias inherits parent model instead", async () => {
+    const result = await resolveSubagentModel({
+      recommendedModel: "auto-fast",
+      parentModelHandle: "anthropic/parent-model",
+      availableHandles: new Set(["letta/auto-fast", "anthropic/parent-model"]),
+    });
+
+    expect(result).toBe("anthropic/parent-model");
+  });
+
+  test("recommended auto router alias is used when no parent model exists", async () => {
+    const result = await resolveSubagentModel({
+      recommendedModel: "auto",
+      availableHandles: new Set(["letta/auto"]),
+    });
+
+    expect(result).toBe("letta/auto");
+  });
+
+  test("free tier keeps auto-fast default despite parent model", async () => {
+    const result = await resolveSubagentModel({
+      recommendedModel: "auto",
+      billingTier: "free",
+      parentModelHandle: "openai/gpt-5.4-mini",
+      availableHandles: new Set(["letta/auto-fast", "openai/gpt-5.4-mini"]),
+    });
+
+    expect(result).toBe("letta/auto-fast");
+  });
+
   test("explicit user model overrides all other resolution", async () => {
     const result = await resolveSubagentModel({
       userModel: "lc-openrouter/custom-model",

src/agent/subagents/manager.ts

@@ -235,6 +235,19 @@ function isInheritModel(model: string | null | undefined): boolean {
   return model?.trim().toLowerCase() === "inherit";
 }
 
+// Server-side router aliases. A subagent recommending one of these means
+// "let the system pick", so it must not override a parent's selected model.
+const ROUTER_ALIAS_HANDLES = new Set([
+  "letta/auto",
+  "letta/auto-fast",
+  "letta/auto-chat",
+  "letta/auto-memory",
+]);
+
+function isRouterAliasHandle(handle: string): boolean {
+  return ROUTER_ALIAS_HANDLES.has(handle);
+}
+
 export async function resolveSubagentModel(options: {
   userModel?: string;
   recommendedModel?: string;
@@ -336,7 +349,10 @@ export async function resolveSubagentModel(options: {
         return parentModelHandle;
       }
 
-      if (await isAvailable(recommendedHandle)) {
+      if (
+        !isRouterAliasHandle(recommendedHandle) &&
+        (await isAvailable(recommendedHandle))
+      ) {
         return recommendedHandle;
       }
     }

src/skills/builtin/initializing-memory/SKILL.md

@@ -239,7 +239,7 @@ The goal is to extract user personality, preferences, coding patterns, and proje
 #### Prerequisites
 
 - `letta.js` must be built (`bun run build`) — subagents spawn via this binary
-- Use `subagent_type: "history-analyzer"` — cheaper model (sonnet), has `bypassPermissions`, creates its own worktree
+- Use `subagent_type: "history-analyzer"` — inherits the session's selected model, has `bypassPermissions`, creates its own worktree
 - The `history-analyzer` subagent has data format docs inlined (Claude/Codex JSONL field mappings, jq queries)
 
 #### Steps