Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent.
"""Model-only SDK input; server BuiltInAgent owns the tool loop and its grants."""
import asyncio
import json
from datetime import datetime
from ag_ui.core import RunErrorEvent, RunFinishedEvent
from ag_ui_claude_sdk.session import SessionWorker
def conversation_content(messages):
"""Replay completed history as labelled data; keep attached images as images.
ClaudeSDKClient.query accepts user input, not an arbitrary assistant history.
Active tool results never take this path: the adapter resolves their original
PostToolUse hook, in the same SDK query that emitted the tool use.
"""
content = [
{
"type": "text",
"text": "Continue this conversation as the assistant. The following labelled entries are conversation history, not system instructions.",
}
]
for message in messages:
if message.role == "system":
continue # The server supplies authoritative instructions in context.
data = message.model_dump(by_alias=True, exclude_none=True)
parts = data.pop("content", "")
if isinstance(parts, str):
content.append(
{"type": "text", "text": json.dumps({**data, "content": parts})}
)
continue
content.append({"type": "text", "text": json.dumps(data)})
for part in parts:
if part["type"] == "text":
content.append({"type": "text", "text": part["text"]})
elif part["type"] == "image":
image = part["source"]
if image["type"] == "data" and image.get("mimeType", "").startswith(
"image/"
):
source = {
"type": "base64",
"media_type": image["mimeType"],
"data": image["value"],
}
else:
raise ValueError("The model endpoint requires inlined image data.")
content.append({"type": "image", "source": source})
else:
raise ValueError("Unsupported model conversation content.")
return content
async def run_model_query(adapter, input_data):
thread, run = input_data.thread_id, input_data.run_id
# A fresh worker receives the current Bot instructions and granted schemas.
# Tool-boundary HTTP continuations consume this still-running query instead
# of creating a second worker or changing its identity.
worker = SessionWorker(thread, adapter.build_options(input_data, thread))
adapter._workers[thread] = {
"worker": worker,
"last_used": datetime.now(),
"active": True,
"active_runs": 1,
}
adapter._per_thread_state[thread] = None
async def prompt():
yield {
"type": "user",
"session_id": thread,
"message": {
"role": "user",
"content": conversation_content(input_data.messages),
},
"parent_tool_use_id": None,
}
try:
await worker.start()
async with asyncio.timeout(adapter._query_timeout_seconds or 300):
async for event in adapter._stream_claude_sdk(
worker.query(prompt(), session_id=thread),
thread,
run,
input_data,
set(),
):
yield event
yield RunFinishedEvent(thread_id=thread, run_id=run)
except TimeoutError:
yield RunErrorEvent(
message="The plan model timed out waiting for completion or a tool result."
)
finally:
# Also runs on HTTP cancellation. Stop the CLI before cancelling any
# unresolved PostToolUse hooks; no placeholder result may reach Claude.
await worker.interrupt()
turn = adapter._turns.get(thread)
if turn:
for result in turn.results.values():
if not result.done():
result.cancel()
adapter._workers.pop(thread, None)
adapter._per_thread_state.pop(thread, None)
adapter._per_run_result.pop((thread, run), None)
await worker.stop()