mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-06 13:26:40 +00:00
Teknium
aefffa077f
The official Codex client has sent zstd request bodies to chatgpt.com/backend-api/codex/responses under ChatGPT auth by default since openai/codex#10102. Hermes sent raw JSON, so large-context sessions uploaded ~3.6x more bytes than the client OpenAI's edge is tuned for. The keepalive httpx client now installs a request hook for the official Codex base URL only: POST .../responses bodies are compressed with the stdlib compression.zstd codec (Python 3.14, the installer default) and sent with Content-Encoding: zstd. Custom proxies, other providers, other endpoints, and interpreters without the codec keep sending raw JSON. Lossless transport encoding: input tokens and prompt caching are unchanged.
47 lines
1.7 KiB
Python
47 lines
1.7 KiB
Python
"""Codex Responses bodies go out zstd-compressed to the official backend only (official-client parity)."""
|
|
|
|
import json
|
|
|
|
import pytest
|
|
|
|
from agent.process_bootstrap import build_keepalive_http_client
|
|
|
|
zstd = pytest.importorskip("compression.zstd")
|
|
|
|
OFFICIAL = "https://chatgpt.com/backend-api/codex"
|
|
|
|
|
|
def _send_through_hooks(client, url: str, payload: dict):
|
|
request = client.build_request("POST", url, json=payload)
|
|
for hook in client.event_hooks["request"]:
|
|
hook(request)
|
|
return request
|
|
|
|
|
|
def test_official_codex_responses_body_is_zstd_and_lossless():
|
|
payload = {"model": "gpt-6.1-sol", "input": [{"role": "user", "content": "hi " * 500}]}
|
|
client = build_keepalive_http_client(OFFICIAL)
|
|
try:
|
|
request = _send_through_hooks(client, f"{OFFICIAL}/responses", payload)
|
|
finally:
|
|
client.close()
|
|
assert request.headers["Content-Encoding"] == "zstd"
|
|
assert int(request.headers["Content-Length"]) == len(request.content)
|
|
assert json.loads(zstd.decompress(request.content)) == payload
|
|
|
|
|
|
@pytest.mark.parametrize("base_url, path", [
|
|
("https://my-proxy.example.com/backend-api/codex", "/responses"), # custom proxy: never compressed
|
|
("https://api.openai.com/v1", "/responses"),
|
|
(OFFICIAL, "/models"), # official host, non-Responses endpoint
|
|
])
|
|
def test_other_destinations_stay_raw_json(base_url, path):
|
|
payload = {"model": "x", "input": "hello"}
|
|
client = build_keepalive_http_client(base_url)
|
|
try:
|
|
request = _send_through_hooks(client, f"{base_url}{path}", payload)
|
|
finally:
|
|
client.close()
|
|
assert "Content-Encoding" not in request.headers
|
|
assert json.loads(request.content) == payload
|