utils.py•6.57 kB
import platform
from datetime import datetime
from typing import cast
from anthropic.types.beta import (
BetaCacheControlEphemeralParam,
BetaImageBlockParam,
BetaMessage,
BetaMessageParam,
BetaTextBlock,
BetaTextBlockParam,
BetaToolResultBlockParam,
BetaToolUseBlockParam,
)
from tools import ToolResult
SYSTEM_PROMPT = f"""<SYSTEM_CAPABILITY>
* You are utilising an Ubuntu virtual machine using {platform.machine()} architecture with internet access.
* You can feel free to install Ubuntu applications with your bash tool. Use curl instead of wget.
* To open firefox, please just click on the firefox icon. Note, firefox-esr is what is installed on your system.
* Using bash tool you can start GUI applications, but you need to set export DISPLAY=:1 and use a subshell. For example "(DISPLAY=:1 xterm &)". GUI apps run with bash tool will appear within your desktop environment, but they may take some time to appear. Take a screenshot to confirm it did.
* When using your bash tool with commands that are expected to output very large quantities of text, redirect into a tmp file and use str_replace_editor or `grep -n -B <lines before> -A <lines after> <query> <filename>` to confirm output.
* 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.
* 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.
* The current date is {datetime.today().strftime("%A, %B %-d, %Y")}.
</SYSTEM_CAPABILITY>
<IMPORTANT>
* When using Firefox, if a startup wizard appears, IGNORE IT. Do not even click "skip this step". Instead, click on the address bar where it says "Search or enter address", and enter the appropriate search term or URL there.
* If the item you are looking at is a pdf, if after taking a single screenshot of the pdf it seems that you want to read the entire document instead of trying to continue to read the pdf from your screenshots + navigation, determine the URL, use curl to download the pdf, install and use pdftotext to convert it to a text file, and then read that text file directly with your StrReplaceEditTool.
</IMPORTANT>"""
def _maybe_prepend_system_tool_result(result: ToolResult, result_text: str):
if result.system:
result_text = f"<system>{result.system}</system>\n{result_text}"
return result_text
def make_api_tool_result(result: ToolResult, tool_use_id: str) -> BetaToolResultBlockParam:
"""Convert an agent ToolResult to an API ToolResultBlockParam."""
tool_result_content: list[BetaTextBlockParam | BetaImageBlockParam] | str = []
is_error = False
if result.error:
is_error = True
tool_result_content = _maybe_prepend_system_tool_result(result, result.error)
else:
if result.output:
tool_result_content.append(
{
"type": "text",
"text": _maybe_prepend_system_tool_result(result, result.output),
}
)
if result.base64_image:
tool_result_content.append(
{
"type": "image",
"source": {
"type": "base64",
"media_type": "image/png",
"data": result.base64_image,
},
}
)
return {
"type": "tool_result",
"content": tool_result_content,
"tool_use_id": tool_use_id,
"is_error": is_error,
}
def inject_prompt_caching(
messages: list[BetaMessageParam],
):
"""
Set cache breakpoints for the 3 most recent turns
one cache breakpoint is left for tools/system prompt, to be shared across sessions
"""
breakpoints_remaining = 3
for message in reversed(messages):
if message["role"] == "user" and isinstance(content := message["content"], list):
if breakpoints_remaining:
breakpoints_remaining -= 1
content[-1]["cache_control"] = BetaCacheControlEphemeralParam({"type": "ephemeral"})
else:
content[-1].pop("cache_control", None)
# we'll only every have one extra turn per loop
break
def maybe_filter_to_n_most_recent_images(
messages: list[BetaMessageParam],
images_to_keep: int,
min_removal_threshold: int,
):
"""
With the assumption that images are screenshots that are of diminishing value as
the conversation progresses, remove all but the final `images_to_keep` tool_result
images in place, with a chunk of min_removal_threshold to reduce the amount we
break the implicit prompt cache.
"""
if images_to_keep is None:
return messages
tool_result_blocks = cast(
list[BetaToolResultBlockParam],
[
item
for message in messages
for item in (message["content"] if isinstance(message["content"], list) else [])
if isinstance(item, dict) and item.get("type") == "tool_result"
],
)
total_images = sum(
1
for tool_result in tool_result_blocks
for content in tool_result.get("content", [])
if isinstance(content, dict) and content.get("type") == "image"
)
images_to_remove = total_images - images_to_keep
# for better cache behavior, we want to remove in chunks
images_to_remove -= images_to_remove % min_removal_threshold
for tool_result in tool_result_blocks:
if isinstance(tool_result.get("content"), list):
new_content = []
for content in tool_result.get("content", []):
if isinstance(content, dict) and content.get("type") == "image":
if images_to_remove > 0:
images_to_remove -= 1
continue
new_content.append(content)
tool_result["content"] = new_content
def response_to_params(
response: BetaMessage,
) -> list[BetaTextBlockParam | BetaToolUseBlockParam]:
res: list[BetaTextBlockParam | BetaToolUseBlockParam] = []
for block in response.content:
if isinstance(block, BetaTextBlock):
res.append({"type": "text", "text": block.text})
else:
res.append(cast(BetaToolUseBlockParam, block.model_dump()))
return res