Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: stop unsloth start openclaw cutting every reply at 8,192 tokens (#12573)

Nilay2e24fd5ef1e2
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 0b66ec886bf..9ab5035a6e0 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -4643,6 +4643,7 @@ def write_openclaw_config(
workspace_path: Optional[str] = None,
embedding_model: Optional[str] = None,
request_body: Optional[dict] = None,
+ max_tokens: Optional[int] = None,
) -> None:
config = _read_json_object(path)
if config is None:
@@ -4657,7 +4658,12 @@ def write_openclaw_config(
provider_model = {"id": model["id"], "name": model["id"]}
window = model.get("context_length") or model.get("max_context_length")
if window:
- provider_model["contextWindow"] = int(window)
+ window = int(window)
+ provider_model["contextWindow"] = window
+ # Unset, OpenClaw caps every reply at 8192 (DEFAULT_MODEL_MAX_TOKENS) whatever the window.
+ provider_model["maxTokens"] = _agent_output_limit(window, max_tokens)
+ elif max_tokens:
+ provider_model["maxTokens"] = max_tokens
models = _subdict(config, "models")
models.setdefault("mode", "merge")
_subdict(models, "providers")["unsloth"] = {
@@ -5619,6 +5625,7 @@ def openclaw(
launch: bool = _LAUNCH_OPTION,
gguf_variant: Optional[str] = _GGUF_VARIANT_OPTION,
max_seq_length: int = _CONTEXT_OPTION,
+ max_tokens: Optional[int] = _MAX_TOKENS_OPTION,
load_in_4bit: bool = _LOAD_4BIT_OPTION,
tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
gpu_memory_mode: Optional[Literal["auto", "manual"]] = _GPU_MEMORY_MODE_OPTION,
@@ -5688,6 +5695,7 @@ def openclaw(
workspace_path = "${OPENCLAW_WORKSPACE_DIR}",
embedding_model = _studio_embedding_model(base, key),
request_body = server_options.request_body(),
+ max_tokens = max_tokens,
)
# Scope both config and state so OpenClaw never touches the user's ~/.openclaw.
# Off, else OpenClaw re-imports any provider key it cannot see from a login shell.
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index 5c27aa11b96..4b7d114e9b1 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -5345,7 +5345,12 @@ def test_write_openclaw_config_fresh(tmp_path):
assert provider["apiKey"] == "sk-unsloth-abc"
assert provider["api"] == "openai-completions"
assert provider["models"] == [
- {"id": MODEL["id"], "name": MODEL["id"], "contextWindow": MODEL["context_length"]}
+ {
+ "id": MODEL["id"],
+ "name": MODEL["id"],
+ "contextWindow": MODEL["context_length"],
+ "maxTokens": 32000,
+ }
]
# The default model must be pinned or OpenClaw has nothing active.
assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
@@ -5357,6 +5362,25 @@ def test_write_openclaw_config_fresh(tmp_path):
assert path.stat().st_mode & 0o777 == 0o600
[email protected]("window, expected", [(32_768, 8_192), (131_072, 32_000)])
+def test_openclaw_output_limit_follows_the_context(tmp_path, window, expected):
+ path = tmp_path / "openclaw.json"
+ start.write_openclaw_config(BASE, "sk-unsloth-abc", {**MODEL, "context_length": window}, path)
+ model = json.loads(path.read_text())["models"]["providers"]["unsloth"]["models"][0]
+ assert model["maxTokens"] == expected
+
+
[email protected]("max_tokens, expected", [("40000", 40000), ("200000", 65536)])
+def test_connect_openclaw_max_tokens(fake_studio, tmp_path, max_tokens, expected):
+ result = CliRunner().invoke(
+ start.start_app, ["openclaw", "--no-launch", "--max-tokens", max_tokens]
+ )
+ assert result.exit_code == 0, result.output
+ assert "--max-tokens" not in _launch_command(result.output)
+ config = json.loads((tmp_path / "agents" / "openclaw" / "openclaw.json").read_text())
+ assert config["models"]["providers"]["unsloth"]["models"][0]["maxTokens"] == expected
+
+
def test_write_openclaw_config_clears_per_agent_path_overrides(tmp_path):
path = tmp_path / "openclaw.json"
path.write_text(