diff --git a/api/agents/shield/explainer.py b/api/agents/shield/explainer.py index efa0c05..6504cda 100644 --- a/api/agents/shield/explainer.py +++ b/api/agents/shield/explainer.py @@ -52,7 +52,7 @@ def _timeout() -> float: def llm_available() -> bool: """Whether a better explanation than the template could be generated.""" - return bool(ai_clients.anthropic(_timeout()) or ai_clients.openai(_timeout())) + return bool(ai_clients.huawei(_timeout()) or ai_clients.anthropic(_timeout()) or ai_clients.openai(_timeout())) # --------------------------------------------------------------------------- @@ -167,7 +167,7 @@ def generate_llm_explanation(signals: list[str], amount: float, action: str) -> "Write the explanation now." ) - text = _try_anthropic(user_prompt) or _try_openai(user_prompt) + text = _try_huawei(user_prompt) or _try_anthropic(user_prompt) or _try_openai(user_prompt) elapsed_ms = round((time.perf_counter() - started) * 1000, 2) if text: @@ -214,6 +214,29 @@ def model_facts(signals: list[str]) -> list[str]: return out[:12] +def _try_huawei(user_prompt: str) -> str | None: + """Huawei MaaS on the Forge cluster. Thinking is off, as the platform + guide asks: GLM reasons by default and max_tokens does not cap it.""" + client = ai_clients.huawei(_timeout()) + if not client: + return None + try: + resp = client.chat.completions.create( + model=ai_clients.huawei_model(), + temperature=0.3, + max_tokens=160, + messages=[ + {"role": "system", "content": SYSTEM_PROMPT}, + {"role": "user", "content": user_prompt}, + ], + extra_body={"chat_template_kwargs": {"thinking": False}}, + ) + text = (resp.choices[0].message.content or "").strip() + return text or None + except Exception: + return None # the template stays; a model error never touches the decision + + def _try_anthropic(user_prompt: str) -> str | None: client = ai_clients.anthropic(_timeout()) if not client: diff --git a/api/core/ai_clients.py b/api/core/ai_clients.py index 2788ab5..333b598 100644 --- a/api/core/ai_clients.py +++ b/api/core/ai_clients.py @@ -8,6 +8,7 @@ """ from __future__ import annotations +import os import threading from core import secrets @@ -53,6 +54,26 @@ def build(key, t): return _cached("anthropic", secrets.get("ANTHROPIC_API_KEY"), timeout, build) +def huawei(timeout: float = 30): + """Huawei MaaS through the Forge gateway: OpenAI-compatible, with the team + key in an `apikey` header. Only set up where FORGE_MAAS_BASE_URL is.""" + base = os.getenv("FORGE_MAAS_BASE_URL", "").strip() + key = os.getenv("GATEWAY_KEY", "").strip() + if not base: + return None + + def build(k, t): + from openai import OpenAI + + return OpenAI(base_url=base, api_key="unused", default_headers={"apikey": k}, timeout=t) + + return _cached("huawei", key, timeout, build) + + +def huawei_model() -> str: + return os.getenv("FORGE_MAAS_MODEL", "").strip() or "glm-5.2" + + def openai_model() -> str: return secrets.get("OPENAI_MODEL") @@ -64,6 +85,7 @@ def anthropic_model() -> str: def configured() -> bool: """Whether any AI provider key is in force (for /health). Never raises.""" try: - return bool(secrets.get("ANTHROPIC_API_KEY") or secrets.get("OPENAI_API_KEY")) + return bool(secrets.get("ANTHROPIC_API_KEY") or secrets.get("OPENAI_API_KEY") + or (os.getenv("FORGE_MAAS_BASE_URL") and os.getenv("GATEWAY_KEY"))) except Exception: # noqa: BLE001 return False diff --git a/api/tests/test_huawei_maas.py b/api/tests/test_huawei_maas.py new file mode 100644 index 0000000..518cc1b --- /dev/null +++ b/api/tests/test_huawei_maas.py @@ -0,0 +1,44 @@ +"""Huawei MaaS writes the explanation wording on the Forge cluster, never the decision.""" +from types import SimpleNamespace + + +def test_huawei_client_sends_the_team_key_in_the_apikey_header(monkeypatch): + from core import ai_clients + + monkeypatch.setenv("FORGE_MAAS_BASE_URL", "http://maas.test/openai/v1") + monkeypatch.setenv("GATEWAY_KEY", "team-key-for-test") + client = ai_clients.huawei(5) + assert client is not None + assert str(client.base_url).startswith("http://maas.test/openai/v1") + assert client.default_headers.get("apikey") == "team-key-for-test" + assert ai_clients.huawei_model() == "glm-5.2" + + +def test_no_huawei_client_off_the_cluster(monkeypatch): + from core import ai_clients + + monkeypatch.delenv("FORGE_MAAS_BASE_URL", raising=False) + assert ai_clients.huawei(5) is None + + +def test_explanation_turns_thinking_off_and_falls_back_on_error(monkeypatch): + from agents.shield import explainer + from core import ai_clients + + seen = {} + + def create(**kwargs): + seen.update(kwargs) + return SimpleNamespace(choices=[SimpleNamespace(message=SimpleNamespace(content="Paused: new payee, unusual amount."))]) + + fake = SimpleNamespace(chat=SimpleNamespace(completions=SimpleNamespace(create=create))) + monkeypatch.setattr(ai_clients, "huawei", lambda _t=30: fake) + assert explainer._try_huawei("decision") == "Paused: new payee, unusual amount." + assert seen["extra_body"] == {"chat_template_kwargs": {"thinking": False}} + assert seen["model"] == "glm-5.2" + + def boom(**_kw): + raise RuntimeError("gateway down") + + fake.chat.completions.create = boom + assert explainer._try_huawei("decision") is None