molecule-ai/hermes-agent
35
0
Fork 0
Code Issues 2 Pull Requests 4 Actions 9 Packages Projects Releases Wiki Activity

fix(cli): /model picker honors provider-specific context caps (#16030)

Browse Source 
`_apply_model_switch_result` (the interactive `/model` picker's
confirmation path) printed `ModelInfo.context_window` straight from
models.dev, which reports the vendor-wide value (1.05M for gpt-5.5 on
openai). ChatGPT Codex OAuth caps the same slug at 272K, so the picker
showed 1M while the runtime (compressor, gateway `/model`, typed
`/model <name>`) correctly used 272K — the classic 'sometimes 1M,
sometimes 272K' mismatch on a single model.

Both display paths now go through `resolve_display_context_length()`,
matching the fix that `_handle_model_switch` received earlier.

Also bump the stale last-resort fallback in DEFAULT_CONTEXT_LENGTHS
(`gpt-5.5: 400000 -> 1050000`) to match the real OpenAI API value; the
272K Codex cap is already enforced via the Codex-OAuth branch, so the
fallback now reflects what every non-Codex probe-miss should see.

Tests: adds `test_apply_model_switch_result_context.py` with three
scenarios (Codex cap wins, OpenRouter shows 1.05M, resolver-empty falls
back to ModelInfo). Updates the existing non-Codex fallback test to
assert 1.05M (the correct value).

## Validation
| path                          | before    | after     |
|-------------------------------|-----------|-----------|
| picker -> gpt-5.5 on Codex    | 1,050,000 | 272,000   |
| picker -> gpt-5.5 on OpenAI   | 1,050,000 | 1,050,000 |
| picker -> gpt-5.5 on OpenRouter | 1,050,000 | 1,050,000 |
| typed /model gpt-5.5 on Codex | 272,000   | 272,000   |
This commit is contained in:
Teknium 2026-04-26 05:43:31 -07:00 committed by GitHub
parent 2ccdadcca6
commit 438db0c7b0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 177 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
agent/model_metadata.py
View File

@ -145,10 +145,11 @@ DEFAULT_CONTEXT_LENGTHS = {
"claude": 200000,
# OpenAI — GPT-5 family (most have 400k; specific overrides first)
# Source: https://developers.openai.com/api/docs/models
# GPT-5.5 (launched Apr 23 2026). 400k is the fallback for providers we
# can't probe live. ChatGPT Codex OAuth actually caps lower (272k as of
# Apr 2026) and is resolved via _resolve_codex_oauth_context_length().
"gpt-5.5": 400000,
# GPT-5.5 (launched Apr 23 2026) is 1.05M on the direct OpenAI API and
# ChatGPT Codex OAuth caps it at 272K; both paths resolve via their own
# provider-aware branches (_resolve_codex_oauth_context_length + models.dev).
# This hardcoded value is only reached when every probe misses.
"gpt-5.5": 1050000,
"gpt-5.4-nano": 400000, # 400k (not 1.05M like full 5.4)
"gpt-5.4-mini": 400000, # 400k (not 1.05M like full 5.4)
"gpt-5.4": 1050000, # GPT-5.4, GPT-5.4 Pro (1.05M context)

30
cli.py
View File

@ -5153,27 +5153,29 @@ class HermesCLI:
_cprint(f" ✓ Model switched: {result.new_model}")
_cprint(f" Provider: {provider_label}")
# Context: always resolve via the provider-aware chain so Codex OAuth,
# Copilot, and Nous-enforced caps win over the raw models.dev entry
# (e.g. gpt-5.5 is 1.05M on openai but 272K on Codex OAuth).
mi = result.model_info
try:
from hermes_cli.model_switch import resolve_display_context_length
ctx = resolve_display_context_length(
result.new_model,
result.target_provider,
base_url=result.base_url or self.base_url or "",
api_key=result.api_key or self.api_key or "",
model_info=mi,
)
if ctx:
_cprint(f" Context: {ctx:,} tokens")
except Exception:
pass
if mi:
if mi.context_window:
_cprint(f" Context: {mi.context_window:,} tokens")
if mi.max_output:
_cprint(f" Max output: {mi.max_output:,} tokens")
if mi.has_cost_data():
_cprint(f" Cost: {mi.format_cost()}")
_cprint(f" Capabilities: {mi.format_capabilities()}")
else:
try:
from agent.model_metadata import get_model_context_length
ctx = get_model_context_length(
result.new_model,
base_url=result.base_url or self.base_url,
api_key=result.api_key or self.api_key,
provider=result.target_provider,
)
_cprint(f" Context: {ctx:,} tokens")
except Exception:
pass
cache_enabled = (
(base_url_host_matches(result.base_url or "", "openrouter.ai") and "claude" in result.new_model.lower())

6
tests/agent/test_model_metadata.py
View File

@ -340,7 +340,9 @@ class TestCodexOAuthContextLength:
from agent.model_metadata import get_model_context_length
# OpenRouter — should hit its own catalog path first; when mocked
# empty, falls through to hardcoded DEFAULT_CONTEXT_LENGTHS (400k).
# empty, falls through to hardcoded DEFAULT_CONTEXT_LENGTHS (1.05M,
# matching the real direct-API value — Codex OAuth's 272k cap is
# provider-specific and must not leak here).
with patch("agent.model_metadata.fetch_model_metadata", return_value={}), \
patch("agent.model_metadata.fetch_endpoint_model_metadata", return_value={}), \
patch("agent.model_metadata.get_cached_context_length", return_value=None), \
@ -351,7 +353,7 @@ class TestCodexOAuthContextLength:
api_key="",
provider="openrouter",
)
assert ctx == 400_000, (
assert ctx == 1_050_000, (
f"Non-Codex gpt-5.5 resolved to {ctx}; Codex 272k override "
"leaked outside openai-codex provider"
)

152
tests/hermes_cli/test_apply_model_switch_result_context.py Normal file
View File

@ -0,0 +1,152 @@
"""Regression test for the `/model` picker confirmation display.
Bug (April 2026): after choosing a model from the interactive `/model` picker,
``HermesCLI._apply_model_switch_result()`` printed ``ModelInfo.context_window``
straight from models.dev, which always reports the vendor-wide value (e.g.
gpt-5.5 = 1,050,000 on ``openai``). That ignored provider-specific caps — in
particular, ChatGPT Codex OAuth enforces 272K on the same slug. The sibling
``_handle_model_switch()`` (typed ``/model <name>``) was already fixed to use
``resolve_display_context_length()``; the picker path was missed, causing
"sometimes 1M, sometimes 272K" for the same model across sibling UI paths.
Fix: both display paths now go through ``resolve_display_context_length()``.
"""
from __future__ import annotations
from unittest.mock import patch
from hermes_cli.model_switch import ModelSwitchResult
class _FakeModelInfo:
context_window = 1_050_000
max_output = 0
def has_cost_data(self):
return False
def format_capabilities(self):
return ""
class _StubCLI:
"""Minimum attrs ``_apply_model_switch_result`` reads on ``self``."""
agent = None
model = ""
provider = ""
requested_provider = ""
api_key = ""
_explicit_api_key = ""
base_url = ""
_explicit_base_url = ""
api_mode = ""
_pending_model_switch_note = ""
def _run_display(monkeypatch, result):
import cli as cli_mod
captured: list[str] = []
monkeypatch.setattr(cli_mod, "_cprint", lambda s, *a, **k: captured.append(str(s)))
# Avoid writing to ~/.hermes/config.yaml during the test.
monkeypatch.setattr(cli_mod, "save_config_value", lambda *a, **k: None)
cli_mod.HermesCLI._apply_model_switch_result(_StubCLI(), result, False)
return captured
def test_picker_path_uses_provider_aware_context_on_codex(monkeypatch):
"""``_apply_model_switch_result`` must prefer the provider-aware resolver
(272K on Codex) over the raw models.dev value (1.05M for gpt-5.5).
"""
result = ModelSwitchResult(
success=True,
new_model="gpt-5.5",
target_provider="openai-codex",
provider_changed=True,
api_key="",
base_url="https://chatgpt.com/backend-api/codex",
api_mode="codex_responses",
warning_message="",
provider_label="ChatGPT Codex",
resolved_via_alias=False,
capabilities=None,
model_info=_FakeModelInfo(), # models.dev says 1.05M
is_global=False,
)
with patch(
"agent.model_metadata.get_model_context_length",
return_value=272_000,
):
lines = _run_display(monkeypatch, result)
ctx_line = next((l for l in lines if "Context:" in l), "")
assert "272,000" in ctx_line, (
f"picker-path display must show Codex's 272K cap, got: {ctx_line!r}"
)
assert "1,050,000" not in ctx_line, (
f"picker-path display leaked models.dev's 1.05M for Codex: {ctx_line!r}"
)
def test_picker_path_shows_vendor_value_when_no_provider_cap(monkeypatch):
"""On providers with no enforced cap (e.g. OpenRouter), the picker path
should surface the real 1.05M context for gpt-5.5 — resolver and models.dev
agree here.
"""
result = ModelSwitchResult(
success=True,
new_model="openai/gpt-5.5",
target_provider="openrouter",
provider_changed=True,
api_key="",
base_url="https://openrouter.ai/api/v1",
api_mode="chat_completions",
warning_message="",
provider_label="OpenRouter",
resolved_via_alias=False,
capabilities=None,
model_info=_FakeModelInfo(),
is_global=False,
)
with patch(
"agent.model_metadata.get_model_context_length",
return_value=1_050_000,
):
lines = _run_display(monkeypatch, result)
ctx_line = next((l for l in lines if "Context:" in l), "")
assert "1,050,000" in ctx_line, (
f"OpenRouter gpt-5.5 should show 1.05M context, got: {ctx_line!r}"
)
def test_picker_path_falls_back_to_model_info_when_resolver_empty(monkeypatch):
"""If ``get_model_context_length`` returns nothing (rare — truly unknown
endpoint), the display still surfaces ``ModelInfo.context_window`` so the
user sees *something* rather than a silent blank.
"""
result = ModelSwitchResult(
success=True,
new_model="some-model",
target_provider="some-provider",
provider_changed=True,
api_key="",
base_url="",
api_mode="chat_completions",
warning_message="",
provider_label="Some Provider",
resolved_via_alias=False,
capabilities=None,
model_info=_FakeModelInfo(), # context_window = 1_050_000
is_global=False,
)
with patch(
"agent.model_metadata.get_model_context_length",
return_value=None,
):
lines = _run_display(monkeypatch, result)
ctx_line = next((l for l in lines if "Context:" in l), "")
assert "1,050,000" in ctx_line, (
f"resolver-empty path should fall back to ModelInfo, got: {ctx_line!r}"
)