From a0f58c1128d59d653f3c28c7c714ce36079ccb3b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 23 Jul 2026 20:47:29 -0700
Subject: [PATCH] Unsloth start: keep Claude subagents on the local model
 (#7333)

Add CLAUDE_CODE_SUBAGENT_MODEL=inherit to the session-only claude settings overlay so built-in subagents stay on the loaded local model.
---
 unsloth_cli/commands/start.py   | 14 +++++++++++---
 unsloth_cli/tests/test_start.py |  7 ++++++-
 2 files changed, 17 insertions(+), 4 deletions(-)

diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 7c768d16f5..3d407576f8 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -1422,10 +1422,18 @@ _DYNAMIC_SECTIONS_FLAG = "--exclude-dynamic-system-prompt-sections"
 
 def _claude_settings_overlay(model_id: str) -> str:
     # Session-only `claude --settings` overlay (command-line tier, no ~/.claude write):
-    # suppress the attribution header, and pin availableModels to the served model so a
-    # user allowlist can't reject it. The pin must be non-empty; [] is ignored.
+    # suppress the attribution header, keep every subagent on the served model (a user
+    # CLAUDE_CODE_SUBAGENT_MODEL pin would otherwise route delegated work off the local
+    # endpoint), and pin availableModels to the served model so a user allowlist can't
+    # reject it. The pin must be non-empty; [] is ignored.
     return json.dumps(
-        {"env": {"CLAUDE_CODE_ATTRIBUTION_HEADER": "0"}, "availableModels": [model_id]}
+        {
+            "env": {
+                "CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
+                "CLAUDE_CODE_SUBAGENT_MODEL": "inherit",
+            },
+            "availableModels": [model_id],
+        }
     )
 
 
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index 36a5e51938..8ce0bcc0d4 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -127,6 +127,8 @@ def test_claude_settings_overlay_pins_served_model():
     assert overlay["availableModels"] == [MODEL["id"]]
     # The attribution-header suppression is preserved alongside it.
     assert overlay["env"]["CLAUDE_CODE_ATTRIBUTION_HEADER"] == "0"
+    # Subagents fall through to the served model instead of a user's opus/sonnet pin.
+    assert overlay["env"]["CLAUDE_CODE_SUBAGENT_MODEL"] == "inherit"
 
 
 def test_install_agent_prompts_then_installs(monkeypatch):
@@ -784,7 +786,10 @@ def test_connect_claude_no_launch(fake_studio):
     _assert_env_set(result.output, "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE", "90")
     assert f"claude --model {MODEL['id']} --exclude-dynamic-system-prompt-sections" in result.output
     # Overlay is passed inline (session-only), not a path into the user's ~/.claude.
-    assert "--settings" in result.output
+    command = _launch_command(result.output)
+    settings = json.loads(command[command.index("--settings") + 1])
+    assert settings["env"]["CLAUDE_CODE_SUBAGENT_MODEL"] == "inherit"
+    assert "--plugin-dir" not in command
     assert ".claude/settings.json" not in result.output