From 1b8dab204061b236c57d9acffa85764f2514de5b Mon Sep 17 00:00:00 2001
From: paulbettner <paulbettner@playfulstudios.com>
Date: Mon, 16 Feb 2026 02:51:27 -0500
Subject: [PATCH] fix(create): apply tier updateArgs when model ID selected
 (#960)

---
 src/agent/create.ts | 25 ++++++++++++++++++-------
 1 file changed, 18 insertions(+), 7 deletions(-)

diff --git a/src/agent/create.ts b/src/agent/create.ts
index 23c5cbc..cfe5ec1 100644
--- a/src/agent/create.ts
+++ b/src/agent/create.ts
@@ -257,9 +257,14 @@ export async function createAgent(
     blockProvenance.push({ label: blockId, source: "shared" });
   }
 
-  // Get the model's context window from its configuration (if known)
-  // First try models.json, then fall back to API-cached context window for BYOK models
-  const modelUpdateArgs = getModelUpdateArgs(modelHandle);
+  // Get the model's context window from its configuration (if known).
+  // If the caller specified a model *ID* (e.g. gpt-5.3-codex-plus-pro-high),
+  // use that identifier to preserve tier-specific updateArgs like reasoning_effort.
+  // Otherwise, fall back to the resolved handle.
+  const modelIdentifierForDefaults = options.model ?? modelHandle;
+  const modelUpdateArgs = options.model
+    ? getModelUpdateArgs(modelIdentifierForDefaults)
+    : undefined;
   const contextWindow =
     (modelUpdateArgs?.context_window as number | undefined) ??
     (await getModelContextWindow(modelHandle));
@@ -325,11 +330,17 @@ export async function createAgent(
 
   // Note: Preflight check above falls back to 'memory' when 'memory_apply_patch' is unavailable.
 
-  // Apply updateArgs if provided (e.g., context_window, reasoning_effort, verbosity, etc.)
-  // We intentionally pass context_window through so updateAgentLLMConfig can set
+  // Apply updateArgs if provided (e.g., context_window, reasoning_effort, verbosity, etc.).
+  // Also apply tier defaults from models.json when the caller explicitly selected a model.
+  //
+  // Note: we intentionally pass context_window through so updateAgentLLMConfig can set
   // context_window_limit using the latest server API, avoiding any fallback.
-  if (options.updateArgs && Object.keys(options.updateArgs).length > 0) {
-    await updateAgentLLMConfig(agent.id, modelHandle, options.updateArgs);
+  const mergedUpdateArgs = {
+    ...(modelUpdateArgs ?? {}),
+    ...(options.updateArgs ?? {}),
+  };
+  if (Object.keys(mergedUpdateArgs).length > 0) {
+    await updateAgentLLMConfig(agent.id, modelHandle, mergedUpdateArgs);
   }
 
   // Always retrieve the agent to ensure we get the full state with populated memory blocks