openclaw/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions 9 Packages Projects Releases Wiki Activity

feat(web): wire local/custom endpoints into model assignment

Browse Source 
The runtime resolver reads model.base_url from config and ignores the
OPENAI_BASE_URL env var, so a self-hosted endpoint could not be configured
from the GUI. Two changes enable it:

- POST /api/model/set accepts an optional base_url and persists it as
  model.base_url when provider=custom (still clearing stale base_url for
  hosted providers).
- POST /api/providers/validate now returns the model ids a custom endpoint
  advertises at /v1/models, so the GUI can auto-pick a default without
  asking the user to type a model name.

Refs desktop onboarding "Local / custom endpoint" bug.
This commit is contained in:
xxxigm
2026-06-03 21:41:29 +07:00
committed by Teknium
parent d50741af90
commit ca06715721
2 changed files with 128 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
hermes_cli/web_server.py
View File

@ -608,6 +608,12 @@ class ModelAssignment(BaseModel):
provider: str
model: str
task: str = ""
# Optional OpenAI-compatible endpoint URL. Only honored for custom/local
# providers on the main slot — lets the GUI configure a self-hosted endpoint
# (vLLM, llama.cpp, Ollama, …) that needs no API key. The runtime resolver
# reads model.base_url from config (it ignores OPENAI_BASE_URL), so this is
# the path that actually wires a local endpoint into resolution.
base_url: str = ""
_GATEWAY_HEALTH_URL = os.getenv("GATEWAY_HEALTH_URL")
@ -1954,6 +1960,7 @@ async def set_model_assignment(body: ModelAssignment):
provider = (body.provider or "").strip()
model = (body.model or "").strip()
task = (body.task or "").strip().lower()
base_url = (body.base_url or "").strip()
if scope not in {"main", "auxiliary"}:
raise HTTPException(status_code=400, detail="scope must be 'main' or 'auxiliary'")
@ -1969,8 +1976,14 @@ async def set_model_assignment(body: ModelAssignment):
model_cfg = {}
model_cfg["provider"] = provider
model_cfg["default"] = model
# Clear stale base_url so the resolver picks the provider's own default.
if "base_url" in model_cfg and model_cfg.get("base_url"):
# Custom/local providers are defined by their endpoint URL, so a
# base_url must be persisted here — the runtime resolver reads
# model.base_url from config and no longer consults OPENAI_BASE_URL.
# For every other provider, clear any stale base_url so the
# resolver picks the provider's own default endpoint.
if provider.strip().lower() == "custom" and base_url:
model_cfg["base_url"] = base_url
elif "base_url" in model_cfg and model_cfg.get("base_url"):
model_cfg["base_url"] = ""
# Also clear hardcoded context_length override — new model may have
# a different context window.
@ -2013,6 +2026,7 @@ async def set_model_assignment(body: ModelAssignment):
"scope": "main",
"provider": provider,
"model": model,
"base_url": model_cfg.get("base_url", ""),
"gateway_tools": gateway_tools,
}
@ -2181,6 +2195,33 @@ _CREDENTIAL_PROBES: dict[str, tuple[str, str]] = {
}
def _parse_model_ids(resp: "Any") -> List[str]:
"""Extract model ids from an OpenAI-compatible ``/v1/models`` response.
Tolerant of the common shapes: ``{"data": [{"id": ...}]}`` (OpenAI / vLLM /
llama.cpp) and a bare ``{"data": ["id", ...]}``. Returns ``[]`` on any
parse/HTTP error so a slightly non-standard endpoint never hard-blocks.
"""
try:
if not resp.is_success:
return []
payload = resp.json()
except Exception:
return []
data = payload.get("data") if isinstance(payload, dict) else payload
if not isinstance(data, list):
return []
ids: List[str] = []
for item in data:
if isinstance(item, dict):
mid = str(item.get("id") or "").strip()
else:
mid = str(item or "").strip()
if mid:
ids.append(mid)
return ids
@app.post("/api/providers/validate")
async def validate_provider_credential(body: EnvVarUpdate, request: Request):
"""Live-probe a provider credential before it's saved.
@ -2199,13 +2240,15 @@ async def validate_provider_credential(body: EnvVarUpdate, request: Request):
return {"ok": False, "reachable": True, "message": "Enter a value first."}
# Local / custom endpoint: validate connectivity, not auth — any HTTP
# response (even 401) proves the endpoint is up.
# response (even 401) proves the endpoint is up. Also surface the model
# ids the endpoint advertises (OpenAI ``/v1/models`` shape) so the GUI can
# auto-pick a default without asking the user to type a model name.
if key == "OPENAI_BASE_URL":
url = value.rstrip("/") + "/models"
try:
with httpx.Client(timeout=httpx.Timeout(8.0)) as client:
client.get(url)
return {"ok": True, "reachable": True, "message": ""}
resp = client.get(url)
return {"ok": True, "reachable": True, "message": "", "models": _parse_model_ids(resp)}
except Exception:
return {"ok": False, "reachable": False, "message": f"Could not reach {url}."}

80
tests/hermes_cli/test_web_server.py
View File

@ -1047,6 +1047,86 @@ class TestWebServerEndpoints:
assert data["ok"] is True
assert data.get("gateway_tools", []) == []
def test_parse_model_ids_handles_openai_and_bare_shapes(self):
"""Model discovery must tolerate the common /v1/models shapes and
never raise (so a slightly non-standard local endpoint still works)."""
from hermes_cli.web_server import _parse_model_ids
class FakeResp:
def __init__(self, payload, ok=True):
self._payload = payload
self.is_success = ok
def json(self):
if isinstance(self._payload, Exception):
raise self._payload
return self._payload
# OpenAI / vLLM / llama.cpp shape.
assert _parse_model_ids(
FakeResp({"data": [{"id": "llama-3.1-8b"}, {"id": "qwen2.5-7b"}]})
) == ["llama-3.1-8b", "qwen2.5-7b"]
# Bare list of ids.
assert _parse_model_ids(FakeResp({"data": ["m1", "m2"]})) == ["m1", "m2"]
# Top-level list.
assert _parse_model_ids(FakeResp([{"id": "x"}])) == ["x"]
# Non-success / malformed / exception → [] (never raises).
assert _parse_model_ids(FakeResp({"data": []}, ok=False)) == []
assert _parse_model_ids(FakeResp({"nope": 1})) == []
assert _parse_model_ids(FakeResp(ValueError("bad json"))) == []
def test_set_model_main_custom_persists_base_url(self):
"""Custom/local providers must persist model.base_url so the runtime
resolver (which ignores OPENAI_BASE_URL) can route to a self-hosted
endpoint without an API key. Regression for the desktop onboarding bug
where 'Local / custom endpoint' could never be configured."""
from hermes_cli.config import load_config
resp = self.client.post(
"/api/model/set",
json={
"scope": "main",
"provider": "custom",
"model": "llama-3.1-8b",
"base_url": "http://127.0.0.1:8000/v1",
},
)
assert resp.status_code == 200
data = resp.json()
assert data["ok"] is True
assert data["provider"] == "custom"
assert data["base_url"] == "http://127.0.0.1:8000/v1"
model_cfg = load_config().get("model")
assert isinstance(model_cfg, dict)
assert model_cfg["provider"] == "custom"
assert model_cfg["default"] == "llama-3.1-8b"
assert model_cfg["base_url"] == "http://127.0.0.1:8000/v1"
def test_set_model_main_non_custom_clears_stale_base_url(self):
"""Switching to a hosted provider must clear a stale base_url so the
resolver picks that provider's own default endpoint."""
from hermes_cli.config import load_config, save_config
cfg = load_config()
cfg["model"] = {
"provider": "custom",
"default": "llama-3.1-8b",
"base_url": "http://127.0.0.1:8000/v1",
}
save_config(cfg)
resp = self.client.post(
"/api/model/set",
json={"scope": "main", "provider": "openrouter", "model": "anthropic/claude-opus-4.8"},
)
assert resp.status_code == 200
assert resp.json()["base_url"] == ""
model_cfg = load_config().get("model")
assert model_cfg["provider"] == "openrouter"
assert model_cfg.get("base_url", "") == ""
def test_set_model_main_gateway_failure_does_not_block_save(self, monkeypatch):
"""A Portal/gateway hiccup must never prevent saving the model."""
import hermes_cli.nous_subscription as ns