fix(session): skip synthetic delegation nudge in subagent sessions

When a user message contains an @agent part, createUserMessage injects a
synthetic text part that tells the model to call the task tool with that
subagent. In a child/subagent session the orchestrator has already written
an explicit task prompt, so injecting the nudge again causes recursive
nesting bias: the subagent re-reads the @agent mention in its own context
and tries to spawn another layer of delegation.

Guard the injection behind a parentID check so only root sessions receive
the synthetic delegation instruction.

Fixes #17202

Author: vicary

packages/opencode/src/session/prompt.ts

@@ -1267,15 +1267,24 @@ export namespace SessionPrompt {
         }
 
         if (part.type === "agent") {
+          const agentPart = {
+            ...part,
+            messageID: info.id,
+            sessionID: input.sessionID,
+          }
+          // Only inject the delegation nudge in root sessions. Subagent sessions
+          // already have an explicit task prompt; injecting it there causes
+          // recursive nesting bias when the agent sees @agent mentions in its
+          // own context.
+          const currentSession = await Session.get(input.sessionID)
+          if (currentSession.parentID) {
+            return [agentPart]
+          }
           // Check if this agent would be denied by task permission
           const perm = PermissionNext.evaluate("task", part.name, agent.permission)
           const hint = perm.action === "deny" ? " . Invoked by user; guaranteed to exist." : ""
           return [
-            {
-              ...part,
-              messageID: info.id,
-              sessionID: input.sessionID,
-            },
+            agentPart,
             {
               messageID: info.id,
               sessionID: input.sessionID,