Skip to content

Automate browser based workflows with AI

fix(SKY-9560): drop colliding timeout kwarg in Vertex cache primary call (#5858)

Celal Zamanoğlu60a35e51c2d1
diff --git a/skyvern-frontend/src/components/ui/css-vars.test.ts b/skyvern-frontend/src/components/ui/css-vars.test.ts
index 266ff5991..b2d0d7871 100644
--- a/skyvern-frontend/src/components/ui/css-vars.test.ts
+++ b/skyvern-frontend/src/components/ui/css-vars.test.ts
@@ -13,17 +13,23 @@ import { fileURLToPath } from "node:url";
// :root and .dark prevents a regression where one entrypoint silently
// drops a token that another entrypoint still ships.
-// skyvern-frontend is an ESM package, so __dirname is undefined here.
-const __filename = fileURLToPath(import.meta.url);
-const __dirname = dirname(__filename);
-const REPO_ROOT = resolve(__dirname, "../../../..");
-// cloud/ and eval/ entrypoints only exist in the cloud sync target; in OSS
-// they are absent and must be filtered out so the suite still runs.
+const HERE = dirname(fileURLToPath(import.meta.url));
+const REPO_ROOT = resolve(HERE, "../../../..");
+// Gate by directory, not file: a deleted index.css inside an existing
+// entrypoint dir is the regression we want to fail loudly on.
+function entrypoint(dir: string): string | null {
+ if (!existsSync(resolve(REPO_ROOT, dir))) {
+ return null;
+ }
+ return `${dir}/index.css`;
+}
const CSS_FILES = [
- "skyvern-frontend/src/index.css",
- "skyvern-frontend/cloud/index.css",
- "skyvern-frontend/eval/index.css",
-].filter((file) => existsSync(resolve(REPO_ROOT, file)));
+ "skyvern-frontend/src",
+ "skyvern-frontend/cloud",
+ "skyvern-frontend/eval",
+]
+ .map(entrypoint)
+ .filter((f): f is string => f !== null);
function load(file: string): string {
return readFileSync(resolve(REPO_ROOT, file), "utf-8");
diff --git a/skyvern/forge/sdk/api/llm/api_handler_factory.py b/skyvern/forge/sdk/api/llm/api_handler_factory.py
index 7c25fdf64..d53c87ba5 100644
--- a/skyvern/forge/sdk/api/llm/api_handler_factory.py
+++ b/skyvern/forge/sdk/api/llm/api_handler_factory.py
@@ -828,6 +828,9 @@ class LLMAPIHandlerFactory:
active_params = copy.deepcopy(litellm_params)
active_params.update(parameters)
active_params["cached_content"] = cache_name
+ # Deployment-level timeout (flex tiers carry their own) wins; passing `timeout`
+ # as an explicit kwarg as well would collide with this entry on unpacking.
+ active_params.setdefault("timeout", settings.LLM_CONFIG_TIMEOUT)
request_model = active_params.pop("model", primary_model_dict.get("model_name", main_model_group))
# Clone messages to avoid modifying original list which is needed for fallback
@@ -858,7 +861,6 @@ class LLMAPIHandlerFactory:
response = await litellm.acompletion(
model=request_model,
messages=active_messages,
- timeout=settings.LLM_CONFIG_TIMEOUT,
drop_params=True,
**active_params,
)
@@ -1985,10 +1987,11 @@ class LLMCaller:
t_llm_request = time.perf_counter()
try:
+ # `timeout` may already live in active_parameters via litellm_params (flex configs
+ # carry their own); passing it explicitly too collides on kwarg unpacking.
response = await self._dispatch_llm_call(
messages=messages,
tools=tools,
- timeout=settings.LLM_CONFIG_TIMEOUT,
**active_parameters,
)
if use_message_history: