agent_expanded.py

import json
import sys
from subprocess import getoutput as run_shell
from urllib.request import Request, urlopen

MODEL = "gpt-5.6"
CONTEXT_WINDOW_TOKENS = 1_050_000

endpoint_url = sys.argv[1]
history = []
request_body = dict(
    model=MODEL,
    input=history,
    tools=[dict(type="custom", name="sh")],
)

while user_prompt := input("> "):
    history += [dict(role="user", content=user_prompt)]
    headers = {"Content-Type": "application/json"}

    while True:
        output_items = (
            response := json.load(
                urlopen(
                    Request(
                        endpoint_url,
                        json.dumps(request_body).encode(),
                        headers,
                    )
                )
            )
        )["output"]
        history += output_items
        tool_calls = [
            item
            for item in output_items
            if item["type"] == "custom_tool_call"
        ]
        context_usage_percent = (
            response["usage"]["total_tokens"] / CONTEXT_WINDOW_TOKENS * 100
        )

        if not tool_calls:
            print(
                output_items[-1]["content"][0]["text"],
                f"\n[{context_usage_percent:06.3f}%]",
            )
            break

        history += [
            dict(
                type="custom_tool_call_output",
                call_id=tool_call["call_id"],
                output=run_shell(tool_call["input"]),
            )
            for tool_call in tool_calls
        ]

see https://news.ycombinator.com/item?id=49006862 and https://gist.github.com/tosh/6e91a9dbf08dd630c535e7345ac7f0b5

Claude and I wrote this version but we didn't test it. It uses http.client (std lib) which keeps the connection open, so less overhead.

import json
import sys
from subprocess import getoutput as run_shell
import http.client
from urllib.parse import urlsplit

url = urlsplit(sys.argv[1])
conn = http.client.HTTPSConnection(url.hostname)
history = []
body = {'model': 'gpt-5.6', 'input': history, 'tools': [{'type': 'custom', 'name': 'sh'}]}

while prompt := input('> '):
    history.append({'role': 'user', 'content': prompt})
    while True:
        conn.request('POST', url.path, json.dumps(body).encode(), {'Content-Type': 'application/json'})
        result = json.load(conn.getresponse())
        output = result['output']
        history.extend(output)
        tool_calls = [item for item in output if item['type'] == 'custom_tool_call']
        if not tool_calls:
            usage_pct = result['usage']['total_tokens'] / 10500
            print(output[-1]['content'][0]['text'], f'\n[{usage_pct:06.3f}%]')
            break
        history.extend({'type': 'custom_tool_call_output', 'call_id': item['call_id'], 'output': run_shell(item['input'])} for item in tool_calls)

also make sure to reuse thread id's for repsonses api

Wow, this is using a lot of clever tricks. I might steal some of them for my surprisingly similar small agent if you don't mind.

I adapted OpenCode's big-pickle model so it runs without having to configure anything. Maybe that could work for you as well.

@swyxio ty re thread id!

@99991 sure, you can consider the code apache 2.0, take what you need

interesting @ big-pickle

Nice, thanks!

a go version of the minimal agent (and continued development as "smol") lives here now:

https://github.com/smol-env/smol