Skip to content

Automate browser based workflows with AI

Mediate browser and CDP downloads (#8135)

Aaron Perez3e8f015f6e96
diff --git a/skyvern/forge/agent.py b/skyvern/forge/agent.py
index 1446651a9..33f5eef4a 100644
--- a/skyvern/forge/agent.py
+++ b/skyvern/forge/agent.py
@@ -4185,6 +4185,8 @@ class ForgeAgent:
cache_key=cache_key,
ttl_seconds=3600, # 1 hour
)
+ if cache_data is None:
+ return
# Store cache metadata in context
context.vertex_cache_name = cache_data["name"]
diff --git a/skyvern/forge/sdk/api/files.py b/skyvern/forge/sdk/api/files.py
index a56ce430b..e9e5394b1 100644
--- a/skyvern/forge/sdk/api/files.py
+++ b/skyvern/forge/sdk/api/files.py
@@ -240,6 +240,7 @@ async def fetch_file_bytes(
None,
current_url,
next_url,
+ strip_cross_origin_credentials=True,
)
source_url, current_url = current_url, next_url
@@ -449,6 +450,7 @@ async def download_file(
None,
current_url,
next_url,
+ strip_cross_origin_credentials=True,
)
source_url, current_url = current_url, next_url
raise SkyvernHTTPException(
diff --git a/skyvern/forge/sdk/api/llm/vertex_cache_manager.py b/skyvern/forge/sdk/api/llm/vertex_cache_manager.py
index 8d0bc90fe..fac1066b9 100644
--- a/skyvern/forge/sdk/api/llm/vertex_cache_manager.py
+++ b/skyvern/forge/sdk/api/llm/vertex_cache_manager.py
@@ -9,6 +9,7 @@ Unlike the Anthropic-style cache_control markers, Vertex AI requires:
"""
import json
+import re
from datetime import datetime
from typing import Any
@@ -20,9 +21,29 @@ from google.auth.transport.requests import Request
from google.oauth2 import service_account
from skyvern.config import settings
+from skyvern.utils.token_counter import count_tokens
LOG = structlog.get_logger()
+# /cachedContents rejects caches below a per-model token minimum (400 with min_total_token_count).
+# Vertex docs: 32,769 tokens for Gemini 1.5; 2,048 for Gemini 2.x; 4,096 for Gemini 3.x — the 3.x GA
+# floor is what broke every cache create when gemini-3.1-flash-lite moved off its -preview id
+# (SKY-10043 §9). Unknown families fail high to the 3.x floor.
+_MIN_CACHED_CONTENT_TOKENS_GEMINI_1X = 32769
+_MIN_CACHED_CONTENT_TOKENS_GEMINI_2X = 2048
+_MIN_CACHED_CONTENT_TOKENS_DEFAULT = 4096
+
+
+def min_cached_content_tokens(model_name: str) -> int:
+ match = re.search(r"gemini-(\d+)", model_name.lower())
+ if match:
+ major = int(match.group(1))
+ if major < 2:
+ return _MIN_CACHED_CONTENT_TOKENS_GEMINI_1X
+ if major == 2:
+ return _MIN_CACHED_CONTENT_TOKENS_GEMINI_2X
+ return _MIN_CACHED_CONTENT_TOKENS_DEFAULT
+
class VertexCacheManager:
"""
@@ -86,7 +107,7 @@ class VertexCacheManager:
cache_key: str,
ttl_seconds: int = 3600,
system_instruction: str | None = None,
- ) -> dict[str, Any]:
+ ) -> dict[str, Any] | None:
"""
Create a cache object using Vertex AI's /cachedContents API.
@@ -98,8 +119,24 @@ class VertexCacheManager:
system_instruction: Optional system instruction to include
Returns:
- Cache data with 'name', 'expireTime', etc.
+ Cache data with 'name', 'expireTime', etc., or None when the content is
+ below the model's minimum cacheable size and creation was skipped.
"""
+ min_tokens = min_cached_content_tokens(model_name)
+ content_token_count = count_tokens(static_content)
+ if system_instruction:
+ content_token_count += count_tokens(system_instruction)
+ if content_token_count < min_tokens:
+ LOG.info(
+ "Static content below Vertex cachedContents minimum, skipping cache creation",
+ sampling=True,
+ cache_key=cache_key,
+ model=model_name,
+ content_token_count=content_token_count,
+ min_required_tokens=min_tokens,
+ )
+ return None
+
# Check if cache already exists for this key
if cache_key in self._cache_registry:
cache_data = self._cache_registry[cache_key]
diff --git a/tests/unit/test_vertex_cache_min_tokens.py b/tests/unit/test_vertex_cache_min_tokens.py
new file mode 100644
index 000000000..1f009b039
--- /dev/null
+++ b/tests/unit/test_vertex_cache_min_tokens.py
@@ -0,0 +1,109 @@
+"""Tests for the Vertex cachedContents minimum-size guard (SKY-9983 / SKY-10043)."""
+
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+
+from skyvern.forge.sdk.api.llm.vertex_cache_manager import VertexCacheManager, min_cached_content_tokens
+from skyvern.utils.token_counter import count_tokens
+
+SMALL_CONTENT = "tiny static prompt"
+MID_CONTENT = "word " * 3000 # between the 2.x and 3.x minimums
+BIG_CONTENT = "word " * 5000 # above every minimum
+
+
[email protected](autouse=True)
+def _token_size_sanity() -> None:
+ assert count_tokens(SMALL_CONTENT) < 2048
+ assert 2048 <= count_tokens(MID_CONTENT) < 4096
+ assert count_tokens(BIG_CONTENT) >= 4096
+
+
+def manager(monkeypatch: pytest.MonkeyPatch) -> VertexCacheManager:
+ mgr = VertexCacheManager(project_id="test-project")
+ monkeypatch.setattr(VertexCacheManager, "_get_access_token", lambda self: "test-token")
+ return mgr
+
+
+def _mock_post(monkeypatch: pytest.MonkeyPatch) -> MagicMock:
+ response = MagicMock()
+ response.status_code = 200
+ response.json.return_value = {"name": "cachedContents/abc", "expireTime": "2099-01-01T00:00:00Z"}
+ post = MagicMock(return_value=response)
+ monkeypatch.setattr("skyvern.forge.sdk.api.llm.vertex_cache_manager.requests.post", post)
+ return post
+
+
+class TestMinCachedContentTokens:
+ @pytest.mark.parametrize(
+ ("model_name", "expected"),
+ [
+ ("gemini-1.5-flash", 32769),
+ ("gemini-1.5-pro", 32769),
+ ("vertex_ai/gemini-1.5-pro", 32769),
+ ("gemini-2.5-flash", 2048),
+ ("gemini-2.5-flash-lite", 2048),
+ ("gemini-2.5-pro", 2048),
+ ("gemini-3-flash-preview", 4096),
+ ("gemini-3.1-flash-lite", 4096),
+ ("some-unknown-model", 4096),
+ ],
+ )
+ def test_per_model_minimums(self, model_name: str, expected: int) -> None:
+ assert min_cached_content_tokens(model_name) == expected
+
+
+class TestCreateCacheGuard:
+ def test_skips_below_minimum_without_any_api_call(
+ self, manager: VertexCacheManager, monkeypatch: pytest.MonkeyPatch
+ ) -> None:
+ post = _mock_post(monkeypatch)
+ result = manager.create_cache(model_name="gemini-3.1-flash-lite", static_content=SMALL_CONTENT, cache_key="k1")
+ assert result is None
+ post.assert_not_called()
+ assert "k1" not in manager._cache_registry
+
+ def test_ga_minimum_regression_case(self, manager: VertexCacheManager, monkeypatch: pytest.MonkeyPatch) -> None:
+ """The SKY-10043 incident shape: content that satisfies the 2.x minimum but not the 3.x GA one."""
+ post = _mock_post(monkeypatch)
+ skipped = manager.create_cache(model_name="gemini-3.1-flash-lite", static_content=MID_CONTENT, cache_key="k-3x")
+ assert skipped is None
+ post.assert_not_called()
+
+ created: dict[str, Any] | None = manager.create_cache(
+ model_name="gemini-2.5-flash", static_content=MID_CONTENT, cache_key="k-2x"
+ )
+ assert created is not None
+ assert created["name"] == "cachedContents/abc"
+ post.assert_called_once()
+
+ def test_gemini_1x_minimum_not_conflated_with_2x(
+ self, manager: VertexCacheManager, monkeypatch: pytest.MonkeyPatch
+ ) -> None:
+ """Content above the 2.x/3.x floors is still below Gemini 1.5's 32,769-token minimum."""
+ post = _mock_post(monkeypatch)
+ result = manager.create_cache(model_name="gemini-1.5-pro", static_content=BIG_CONTENT, cache_key="k-1x")
+ assert result is None
+ post.assert_not_called()
+
+ def test_creates_above_minimum(self, manager: VertexCacheManager, monkeypatch: pytest.MonkeyPatch) -> None:
+ post = _mock_post(monkeypatch)
+ result = manager.create_cache(model_name="gemini-3.1-flash-lite", static_content=BIG_CONTENT, cache_key="k2")
+ assert result is not None
+ assert manager._cache_registry["k2"]["name"] == "cachedContents/abc"
+ post.assert_called_once()
+
+ def test_system_instruction_counts_toward_minimum(
+ self, manager: VertexCacheManager, monkeypatch: pytest.MonkeyPatch
+ ) -> None:
+ post = _mock_post(monkeypatch)
+ result = manager.create_cache(
+ model_name="gemini-3.1-flash-lite",
+ static_content=MID_CONTENT,
+ cache_key="k3",
+ system_instruction=MID_CONTENT,
+ )
+ assert result is not None
+ post.assert_called_once()