mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-04-25 00:51:20 +00:00
05d8f11085
/model gpt-5.5 on openai-codex showed 'Context: 1,050,000 tokens' because the display block used ModelInfo.context_window directly from models.dev. Codex OAuth actually enforces 272K for the same slug, and the agent's compressor already runs at 272K via get_model_context_length() — so the banner + real context budget said 272K while /model lied with 1M. Route the display context through a new resolve_display_context_length() helper that always prefers agent.model_metadata.get_model_context_length (which knows about Codex OAuth, Copilot, Nous caps) and only falls back to models.dev when that returns nothing. Fix applied to all 3 /model display sites: cli.py _handle_model_switch gateway/run.py picker on_model_selected callback gateway/run.py text-fallback confirmation Reported by @emilstridell (Telegram, April 2026).
90 lines
3.3 KiB
Python
90 lines
3.3 KiB
Python
"""Regression test for /model context-length display on provider-capped models.
|
|
|
|
Bug (April 2026): `/model gpt-5.5` on openai-codex (ChatGPT OAuth) showed
|
|
"Context: 1,050,000 tokens" because the display code used the raw models.dev
|
|
``ModelInfo.context_window`` (which reports the direct-OpenAI API value) instead
|
|
of the provider-aware resolver. The agent was actually running at 272K — Codex
|
|
OAuth's enforced cap — so the display was lying to the user.
|
|
|
|
Fix: ``resolve_display_context_length()`` prefers
|
|
``agent.model_metadata.get_model_context_length`` (which knows about Codex OAuth,
|
|
Copilot, Nous, etc.) and falls back to models.dev only if that returns nothing.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
from unittest.mock import patch
|
|
|
|
from hermes_cli.model_switch import resolve_display_context_length
|
|
|
|
|
|
class _FakeModelInfo:
|
|
def __init__(self, ctx):
|
|
self.context_window = ctx
|
|
|
|
|
|
class TestResolveDisplayContextLength:
|
|
def test_codex_oauth_overrides_models_dev(self):
|
|
"""gpt-5.5 on openai-codex must show Codex's 272K cap, not models.dev's 1.05M."""
|
|
fake_mi = _FakeModelInfo(1_050_000) # what models.dev reports
|
|
with patch(
|
|
"agent.model_metadata.get_model_context_length",
|
|
return_value=272_000, # what Codex OAuth actually enforces
|
|
):
|
|
ctx = resolve_display_context_length(
|
|
"gpt-5.5",
|
|
"openai-codex",
|
|
base_url="https://chatgpt.com/backend-api/codex",
|
|
api_key="",
|
|
model_info=fake_mi,
|
|
)
|
|
assert ctx == 272_000, (
|
|
"Codex OAuth's 272K cap must win over models.dev's 1.05M for gpt-5.5"
|
|
)
|
|
|
|
def test_falls_back_to_model_info_when_resolver_returns_none(self):
|
|
fake_mi = _FakeModelInfo(1_048_576)
|
|
with patch(
|
|
"agent.model_metadata.get_model_context_length", return_value=None
|
|
):
|
|
ctx = resolve_display_context_length(
|
|
"some-model",
|
|
"some-provider",
|
|
model_info=fake_mi,
|
|
)
|
|
assert ctx == 1_048_576
|
|
|
|
def test_returns_none_when_both_sources_empty(self):
|
|
with patch(
|
|
"agent.model_metadata.get_model_context_length", return_value=None
|
|
):
|
|
ctx = resolve_display_context_length(
|
|
"unknown-model",
|
|
"unknown-provider",
|
|
model_info=None,
|
|
)
|
|
assert ctx is None
|
|
|
|
def test_resolver_exception_falls_back_to_model_info(self):
|
|
fake_mi = _FakeModelInfo(200_000)
|
|
with patch(
|
|
"agent.model_metadata.get_model_context_length",
|
|
side_effect=RuntimeError("network down"),
|
|
):
|
|
ctx = resolve_display_context_length(
|
|
"x", "y", model_info=fake_mi
|
|
)
|
|
assert ctx == 200_000
|
|
|
|
def test_prefers_resolver_even_when_model_info_has_larger_value(self):
|
|
"""Invariant: provider-aware resolver is authoritative, even if models.dev
|
|
reports a bigger window."""
|
|
fake_mi = _FakeModelInfo(2_000_000)
|
|
with patch(
|
|
"agent.model_metadata.get_model_context_length", return_value=128_000
|
|
):
|
|
ctx = resolve_display_context_length(
|
|
"capped-model",
|
|
"capped-provider",
|
|
model_info=fake_mi,
|
|
)
|
|
assert ctx == 128_000
|