Add Claude Sonnet 4.5 support and improve action handling (#362)
🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Qichen Fu
@@ -27,7 +27,7 @@ from datetime import datetime
|
||||
from .tools import ToolResult
|
||||
|
||||
|
||||
COMPUTER_USE_BETA_FLAG = "computer-use-2024-10-22"
|
||||
COMPUTER_USE_BETA_FLAG = "computer-use-2025-01-24"
|
||||
PROMPT_CACHING_BETA_FLAG = "prompt-caching-2024-07-31"
|
||||
|
||||
|
||||
@@ -47,12 +47,25 @@ PROVIDER_TO_DEFAULT_MODEL_NAME: dict[(APIProvider, str), str] = {
|
||||
(APIProvider.ANTHROPIC, "claude-4-opus-20250514"): "claude-4-opus-20250514",
|
||||
(APIProvider.BEDROCK, "claude-4-opus-20250514"): "us.anthropic.claude-opus-4-20250514-v1:0",
|
||||
(APIProvider.VERTEX, "claude-4-opus-20250514"): "claude-4-opus-v1@20250514",
|
||||
# Add mapping for the alternative model name format
|
||||
(APIProvider.ANTHROPIC, "claude-opus-4-20250514"): "claude-opus-4-20250514",
|
||||
(APIProvider.ANTHROPIC, "claude-opus-4-1-20250805"): "claude-opus-4-1-20250805",
|
||||
(APIProvider.ANTHROPIC, "claude-4-sonnet-20250514"): "claude-4-sonnet-20250514",
|
||||
(APIProvider.ANTHROPIC, "claude-sonnet-4-20250514"): "claude-sonnet-4-20250514",
|
||||
(APIProvider.BEDROCK, "claude-4-sonnet-20250514"): "us.anthropic.claude-sonnet-4-20250514-v1:0",
|
||||
(APIProvider.VERTEX, "claude-4-sonnet-20250514"): "claude-sonnet-4-v1@20250514",
|
||||
}
|
||||
|
||||
|
||||
def get_model_name(provider: APIProvider, model_name: str) -> str:
|
||||
"""
|
||||
Get the actual model name to use for API calls.
|
||||
|
||||
Simply returns the model name as-is for direct API usage.
|
||||
"""
|
||||
return model_name
|
||||
|
||||
|
||||
# This system prompt is optimized for the Docker environment in this repository and
|
||||
# specific tool combinations enabled.
|
||||
# We encourage modifying this system prompt to ensure the model has context for the
|
||||
@@ -67,8 +80,15 @@ SYSTEM_PROMPT = f"""<SYSTEM_CAPABILITY>
|
||||
* When viewing a page it can be helpful to zoom out so that you can see everything on the page. Either that, or make sure you scroll down to see everything before deciding something isn't available.
|
||||
* DO NOT ask users for clarification during task execution. DO NOT stop to request more information from users. Always take action using available tools.
|
||||
* When using your computer function calls, they take a while to run and send back to you. Where possible/feasible, try to chain multiple of these calls all into one function calls request.
|
||||
* TASK FEASIBILITY: You can declare a task infeasible at any point during execution - whether at the beginning after taking a screenshot, or later after attempting some actions and discovering barriers. Carefully evaluate whether the task is feasible given the current system state, available applications, and task requirements. If you determine that a task cannot be completed due to:
|
||||
- Missing required applications or dependencies that cannot be installed
|
||||
- Insufficient permissions or system limitations
|
||||
- Contradictory or impossible requirements
|
||||
- Any other fundamental barriers that make completion impossible
|
||||
Then you MUST output exactly "[INFEASIBLE]" (including the square brackets) anywhere in your response to trigger the fail action. The system will automatically detect this pattern and terminate the task appropriately.
|
||||
* The current date is {datetime.today().strftime('%A, %B %d, %Y')}.
|
||||
* Home directory of this Ubuntu system is '/home/user'.
|
||||
* If you need a password for sudo, the password of the computer is 'osworld-public-evaluation'.
|
||||
</SYSTEM_CAPABILITY>
|
||||
|
||||
<IMPORTANT>
|
||||
@@ -82,6 +102,7 @@ SYSTEM_PROMPT_WINDOWS = f"""<SYSTEM_CAPABILITY>
|
||||
* The current date is {datetime.today().strftime('%A, %B %d, %Y')}.
|
||||
* Home directory of this Windows system is 'C:\\Users\\user'.
|
||||
* When you want to open some applications on Windows, please use Double Click on it instead of clicking once.
|
||||
* If you need a password for sudo, The password of the computer is 'osworld-public-evaluation'.
|
||||
</SYSTEM_CAPABILITY>"""
|
||||
|
||||
|
||||
@@ -154,21 +175,30 @@ def _inject_prompt_caching(
|
||||
one cache breakpoint is left for tools/system prompt, to be shared across sessions
|
||||
"""
|
||||
|
||||
breakpoints_remaining = 3
|
||||
breakpoints_remaining = 2 # Use full budget for recent messages
|
||||
messages_processed = 0
|
||||
|
||||
for message in reversed(messages):
|
||||
if message["role"] == "user" and isinstance(
|
||||
content := message["content"], list
|
||||
):
|
||||
if breakpoints_remaining:
|
||||
breakpoints_remaining -= 1
|
||||
messages_processed += 1
|
||||
# Check if this message would fit within the remaining budget
|
||||
if breakpoints_remaining >= len(content):
|
||||
# We have enough budget, spend it and add cache_control
|
||||
breakpoints_remaining -= len(content)
|
||||
# Use type ignore to bypass TypedDict check until SDK types are updated
|
||||
content[-1]["cache_control"] = BetaCacheControlEphemeralParam( # type: ignore
|
||||
{"type": "ephemeral"}
|
||||
)
|
||||
else:
|
||||
content[-1].pop("cache_control", None)
|
||||
# we'll only every have one extra turn per loop
|
||||
break
|
||||
# Check if this is the first message (contains image + text with task description)
|
||||
is_first_message = messages_processed == len([msg for msg in messages if msg["role"] == "user"])
|
||||
|
||||
if not is_first_message:
|
||||
# Not enough budget, remove any existing cache_control from this message
|
||||
content[-1].pop("cache_control", None)
|
||||
# Continue to clean up older messages that might have cache_control from previous turns
|
||||
|
||||
|
||||
def _maybe_filter_to_n_most_recent_images(
|
||||
@@ -220,6 +250,105 @@ def _maybe_filter_to_n_most_recent_images(
|
||||
tool_result["content"] = new_content
|
||||
|
||||
|
||||
def validate_model_support(model_name: str, api_key: str = None, temperature: float = None, top_p: float = None, no_thinking: bool = False, use_isp: bool = False) -> bool:
|
||||
"""
|
||||
Validate model support with the same API call pattern as the main agent.
|
||||
|
||||
Args:
|
||||
model_name: The model name to validate
|
||||
api_key: Optional API key, defaults to ANTHROPIC_API_KEY env var
|
||||
temperature: Optional temperature parameter for testing
|
||||
top_p: Optional top_p parameter for testing
|
||||
no_thinking: Disable thinking mode (matches AnthropicAgent)
|
||||
use_isp: Use interleaved scratchpad mode (matches AnthropicAgent)
|
||||
|
||||
Returns:
|
||||
True if model is supported and API call succeeds, False otherwise
|
||||
"""
|
||||
print(f"🔍 Validating model support: {model_name}")
|
||||
|
||||
try:
|
||||
from anthropic import Anthropic
|
||||
import os
|
||||
import time
|
||||
|
||||
# Same client setup as main agent but with manual retry (max_retries=1 for faster feedback)
|
||||
client = Anthropic(
|
||||
api_key=api_key or os.environ.get("ANTHROPIC_API_KEY"),
|
||||
max_retries=4
|
||||
).with_options(default_headers={"anthropic-beta": COMPUTER_USE_BETA_FLAG})
|
||||
|
||||
# Same message format as main agent - always use structured format with cache_control
|
||||
messages = [{"role": "user", "content": [{"type": "text", "text": "Respond with 'OK'", "cache_control": {"type": "ephemeral"}}]}]
|
||||
|
||||
# Same betas configuration as main agent
|
||||
betas = [COMPUTER_USE_BETA_FLAG]
|
||||
if use_isp:
|
||||
betas.append("interleaved-thinking-2025-05-14")
|
||||
|
||||
system = [{"type": "text", "text": "You are Claude. Respond with 'OK'."}]
|
||||
|
||||
# Same tools configuration as main agent - use modern computer tool for all models
|
||||
tools = [{"name": "computer", "type": "computer_20250124",
|
||||
"display_width_px": 1280, "display_height_px": 720, "display_number": 1}]
|
||||
|
||||
# Same thinking configuration as main agent
|
||||
max_tokens = 50 # Base validation max_tokens
|
||||
if no_thinking:
|
||||
extra_body = {}
|
||||
actual_max_tokens = max_tokens
|
||||
else:
|
||||
budget_tokens = 2048
|
||||
# Same logic as main agent: if max_tokens <= budget_tokens, increase it
|
||||
if max_tokens <= budget_tokens:
|
||||
actual_max_tokens = budget_tokens + 500
|
||||
else:
|
||||
actual_max_tokens = max_tokens
|
||||
extra_body = {
|
||||
"thinking": {"type": "enabled", "budget_tokens": budget_tokens}
|
||||
}
|
||||
|
||||
# Sampling parameters (same logic as main agent)
|
||||
sampling_params = {}
|
||||
if temperature is not None:
|
||||
sampling_params['temperature'] = temperature
|
||||
if top_p is not None:
|
||||
sampling_params['top_p'] = top_p
|
||||
|
||||
# Retry logic with 5 attempts, 5 second delays
|
||||
for attempt in range(5):
|
||||
try:
|
||||
# Same API call pattern as main agent
|
||||
client.beta.messages.create(
|
||||
max_tokens=actual_max_tokens,
|
||||
messages=messages,
|
||||
model=get_model_name(APIProvider.ANTHROPIC, model_name),
|
||||
system=system,
|
||||
tools=tools,
|
||||
betas=betas,
|
||||
extra_body=extra_body,
|
||||
**sampling_params
|
||||
)
|
||||
|
||||
print(f"✅ Model {model_name} validated successfully")
|
||||
return True
|
||||
except Exception as e:
|
||||
if attempt < 4: # Don't print error on final attempt
|
||||
print(f"🔄 Validation attempt {attempt + 1}/5 failed: {e}")
|
||||
print(f"⏳ Retrying in 5 seconds...")
|
||||
time.sleep(5)
|
||||
else:
|
||||
print(f"❌ All validation attempts failed. Final error: {e}")
|
||||
|
||||
return False
|
||||
|
||||
except ValueError:
|
||||
return False
|
||||
except Exception as e:
|
||||
print(f"❌ API validation setup failed: {e}")
|
||||
return False
|
||||
|
||||
|
||||
def _response_to_params(
|
||||
response: BetaMessage,
|
||||
) -> list[BetaContentBlockParam]:
|
||||
|
||||
Reference in New Issue
Block a user