51609d827b
Provider discovery imports every bundled plugin's __init__, and it runs in interpreters without the app's HTTP stack: PM's runtime calls lazy_installs_allowed() -> hermes_cli.config, whose import-time _inject_profile_env_vars() lists providers. solstice/__init__ imported transport.py at the top, which pulls agent.gemini_native_adapter -> httpx, so every install logged "Failed to load bundled provider plugin solstice: No module named 'httpx'" and solstice went unregistered in that process. INFERENCE_BASE_URL now lives in __init__ and SolsticeClient is imported inside create_client, matching the other provider plugins.
66 lines
3.2 KiB
Python
66 lines
3.2 KiB
Python
"""Solstice provider profile: a user's own per-user-quota Gemini subscription, signed in with OAuth.
|
|
|
|
Pre-release (``hidden=True``): no picker, setup list or accounts tab offers it until the user signs in
|
|
by name (``hermes auth add solstice``); after that it lists like any signed-in provider. Login and
|
|
token refresh go through the generic PKCE plugin flow with a Nous-portal broker as the token endpoint
|
|
(``auth.py``); inference goes straight to Google on the per-user-quota methods (``transport.py``).
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from dataclasses import replace
|
|
from typing import Any
|
|
|
|
from hermes_cli.auth_oauth_pkce_plugin import OAuthPKCEConfig, pkce_auth_handler, pkce_refresh_credential
|
|
from providers import register_provider
|
|
from providers.base import ProviderProfile
|
|
|
|
from .auth import broker_token_request, discover_client
|
|
|
|
# Lives here, not in transport.py: discovery imports this module from interpreters without the app's
|
|
# HTTP stack (PM's runtime reads config, which lists providers), so the transport loads on first use.
|
|
INFERENCE_BASE_URL = "https://generativelanguage.googleapis.com/v1alpha"
|
|
|
|
# Verified on the per-user-quota endpoint, which has no model listing a user token may read (its
|
|
# /models answers 403 ACCESS_TOKEN_SCOPE_INSUFFICIENT). Quota is per model, so lite stays last as the
|
|
# reserve once the flash family is exhausted.
|
|
FALLBACK_MODELS = ("gemini-3.5-flash", "gemini-flash-latest", "gemini-flash-lite-latest")
|
|
|
|
# The broker owns client_id/scope (discovered per login); refresh needs neither, the broker adds them.
|
|
_BASE_CFG = OAuthPKCEConfig(
|
|
client_id="", authorize_url="", token_url="", label="Solstice",
|
|
redirect_path="/gemini-auth/callback", token_request=broker_token_request,
|
|
# A refresh token is issued only for offline access, and Google omits it on a repeat consent.
|
|
extra_authorize_params={"access_type": "offline", "prompt": "consent"},
|
|
)
|
|
|
|
|
|
def _auth_handler(action: str, args: Any) -> bool:
|
|
cfg = _BASE_CFG
|
|
if action == "add":
|
|
client = discover_client()
|
|
cfg = replace(_BASE_CFG, client_id=client["client_id"], authorize_url=client["authorize_url"],
|
|
scopes=tuple(client["scope"].split()))
|
|
return pkce_auth_handler(cfg)(action, args)
|
|
|
|
|
|
class SolsticeProfile(ProviderProfile):
|
|
def create_client(self, **client_kwargs: Any) -> Any:
|
|
from .transport import SolsticeClient
|
|
|
|
allowed = {"api_key", "base_url", "default_headers", "timeout", "http_client"}
|
|
return SolsticeClient(**{k: v for k, v in client_kwargs.items() if k in allowed})
|
|
|
|
def get_max_tokens(self, model: str | None) -> int | None:
|
|
return None # output caps are the backend's; a Hermes default would truncate long answers
|
|
|
|
|
|
register_provider(SolsticeProfile(
|
|
name="solstice", aliases=("solstice-oauth",), display_name="Solstice",
|
|
description="Solstice (your own per-user quota, OAuth sign-in)",
|
|
auth_type="oauth_external", base_url=INFERENCE_BASE_URL, hidden=True,
|
|
supports_health_check=False, supports_model_listing=False, supports_vision=True,
|
|
fallback_models=FALLBACK_MODELS, default_aux_model="gemini-flash-lite-latest",
|
|
auth_handler=_auth_handler, refresh_credential=pkce_refresh_credential(_BASE_CFG),
|
|
))
|