fix(voice): single scoped resolver — STT/TTS keys fall back to the credential pool

Rework of #68509 per triage: hoist the duplicated per-tool
_resolve_provider_key helpers into one owner,
tools.tool_backend_helpers.resolve_provider_secret(), and migrate every
STT/TTS key lookup site to it.

Resolution order: explicit config.yaml value > profile secret scope /
env / ~/.hermes/.env > credential pool (checks both '<provider>' and
'custom:<provider>' pool keys, so keys added via 'hermes auth add
mistral' or declared under providers.<name> both resolve). Under an
active multiplex turn the profile scope stays authoritative — no pool
or .env fallback that could borrow another profile's key (composes with
the #69469 scope fix).

Coverage now includes GROQ_API_KEY, MISTRAL_API_KEY, ELEVENLABS_API_KEY,
DEEPINFRA_API_KEY, MINIMAX_API_KEY, GEMINI_API_KEY/GOOGLE_API_KEY, the
XAI_API_KEY fallback in resolve_xai_http_credentials, and the OpenAI
audio key (resolve_openai_audio_api_key now pool-aware for
OPENAI_API_KEY via 'hermes auth add openai-api').

Unit tests: fake pool entry proves each provider resolves from the pool
when env is empty; env still wins when set; config wins over both; a
multiplex scope miss never borrows the pool; pool read failures never
raise; tool-level wiring for STT, TTS, xAI, and OpenAI audio.

Fixes #68003
This commit is contained in:
Teknium 2026-07-28 09:37:57 -07:00
parent 79bcfc23ab
commit c136400c9e
5 changed files with 391 additions and 51 deletions

View file

@ -0,0 +1,239 @@
"""Tests for ``resolve_provider_secret`` — the single owner of STT/TTS
provider key resolution (#68003).
Keys added via ``hermes auth add <provider>`` live in the credential pool /
auth store and used to be invisible to the voice tools, which only read
``os.environ`` + ``~/.hermes/.env`` via ``get_env_value``. The shared
resolver falls back to the pool; env still wins when set; an explicit
config.yaml value wins over both; and under a multiplexed gateway turn the
profile secret scope stays authoritative (no pool borrow).
"""
from types import SimpleNamespace
from unittest.mock import patch
import pytest
from tools.tool_backend_helpers import resolve_provider_secret
def _fake_pool(key: str = "", *, has: bool = True):
entry = SimpleNamespace(runtime_api_key=key, access_token=key) if key else None
return SimpleNamespace(
has_credentials=lambda: has and bool(key),
peek=lambda: entry,
)
@pytest.fixture(autouse=True)
def _clean_env(monkeypatch):
for var in (
"GROQ_API_KEY", "MISTRAL_API_KEY", "ELEVENLABS_API_KEY",
"DEEPINFRA_API_KEY", "MINIMAX_API_KEY", "GEMINI_API_KEY",
"GOOGLE_API_KEY", "XAI_API_KEY", "OPENAI_API_KEY",
"VOICE_TOOLS_OPENAI_KEY",
):
monkeypatch.delenv(var, raising=False)
yield
@pytest.fixture(autouse=True)
def _no_dotenv(monkeypatch):
"""Keep the developer's real ~/.hermes/.env out of these tests."""
import hermes_cli.config as config_mod
monkeypatch.setattr(config_mod, "load_env", lambda: {})
yield
class TestPoolFallback:
"""Each provider's key resolves from the credential pool when env is empty."""
@pytest.mark.parametrize(
"env_var,provider_id",
[
("GROQ_API_KEY", "groq"),
("MISTRAL_API_KEY", "mistral"),
("ELEVENLABS_API_KEY", "elevenlabs"),
("DEEPINFRA_API_KEY", "deepinfra"),
("MINIMAX_API_KEY", "minimax"),
("GEMINI_API_KEY", "gemini"),
("XAI_API_KEY", "xai"),
("OPENAI_API_KEY", "openai-api"),
],
)
def test_pool_entry_resolves_when_env_empty(self, env_var, provider_id):
pool_key_seen = []
def fake_load_pool(pid):
pool_key_seen.append(pid)
if pid == provider_id:
return _fake_pool(f"pool-key-{provider_id}")
return _fake_pool("")
with patch("agent.credential_pool.load_pool", side_effect=fake_load_pool):
assert resolve_provider_secret(env_var, provider_id) == (
f"pool-key-{provider_id}"
)
assert provider_id in pool_key_seen
def test_custom_pool_key_fallback(self):
"""A provider pooled under ``custom:<name>`` (config.yaml providers)
is found when the plain pool id is empty the issue's
``custom:mistral`` scenario."""
def fake_load_pool(pid):
if pid == "custom:mistral":
return _fake_pool("custom-mistral-key")
return _fake_pool("")
with patch("agent.credential_pool.load_pool", side_effect=fake_load_pool):
assert (
resolve_provider_secret("MISTRAL_API_KEY", "mistral")
== "custom-mistral-key"
)
def test_no_key_anywhere_returns_empty(self):
with patch(
"agent.credential_pool.load_pool", return_value=_fake_pool("")
):
assert resolve_provider_secret("MISTRAL_API_KEY", "mistral") == ""
def test_pool_read_failure_never_raises(self):
with patch(
"agent.credential_pool.load_pool", side_effect=Exception("disk error")
):
assert resolve_provider_secret("MISTRAL_API_KEY", "mistral") == ""
class TestEnvPrecedence:
"""Env / .env still wins over the pool when set — unchanged behaviour."""
def test_env_wins_over_pool(self, monkeypatch):
monkeypatch.setenv("ELEVENLABS_API_KEY", "env-key")
with patch(
"agent.credential_pool.load_pool",
return_value=_fake_pool("pool-key"),
) as lp:
assert (
resolve_provider_secret("ELEVENLABS_API_KEY", "elevenlabs")
== "env-key"
)
lp.assert_not_called()
def test_env_getter_is_consulted(self):
"""Callers can pass their module-level get_env_value wrapper."""
with patch(
"agent.credential_pool.load_pool", return_value=_fake_pool("")
):
assert (
resolve_provider_secret(
"GROQ_API_KEY",
"groq",
env_getter=lambda name: "dotenv-key",
)
== "dotenv-key"
)
class TestConfigPrecedence:
def test_config_value_wins_over_env_and_pool(self, monkeypatch):
monkeypatch.setenv("MISTRAL_API_KEY", "env-key")
with patch(
"agent.credential_pool.load_pool",
return_value=_fake_pool("pool-key"),
):
assert (
resolve_provider_secret(
"MISTRAL_API_KEY", "mistral", config_value="cfg-key"
)
== "cfg-key"
)
class TestMultiplexScope:
"""Under multiplexing the profile scope is authoritative — no pool borrow."""
@pytest.fixture(autouse=True)
def _reset_multiplex(self):
from agent import secret_scope as ss
ss.set_multiplex_active(False)
yield
ss.set_multiplex_active(False)
def test_scope_value_wins(self, monkeypatch):
from agent import secret_scope as ss
monkeypatch.setenv("MISTRAL_API_KEY", "sk-other-profile")
ss.set_multiplex_active(True)
token = ss.set_secret_scope({"MISTRAL_API_KEY": "sk-this-profile"})
try:
assert (
resolve_provider_secret("MISTRAL_API_KEY", "mistral")
== "sk-this-profile"
)
finally:
ss.reset_secret_scope(token)
def test_scope_miss_does_not_fall_through_to_pool(self):
from agent import secret_scope as ss
ss.set_multiplex_active(True)
token = ss.set_secret_scope({"UNRELATED": "x"})
try:
with patch(
"agent.credential_pool.load_pool",
return_value=_fake_pool("pool-key"),
) as lp:
assert resolve_provider_secret("MISTRAL_API_KEY", "mistral") == ""
lp.assert_not_called()
finally:
ss.reset_secret_scope(token)
class TestToolWiring:
"""The tools' module-level helpers delegate to the shared resolver."""
def test_transcription_tools_delegates(self):
from tools import transcription_tools as tt
def fake_load_pool(pid):
return _fake_pool("stt-pool-key" if pid == "groq" else "")
with patch("agent.credential_pool.load_pool", side_effect=fake_load_pool):
assert tt._resolve_provider_key("GROQ_API_KEY", "groq") == "stt-pool-key"
def test_tts_tool_delegates(self):
from tools import tts_tool
def fake_load_pool(pid):
return _fake_pool("tts-pool-key" if pid == "minimax" else "")
with patch("agent.credential_pool.load_pool", side_effect=fake_load_pool):
assert (
tts_tool._resolve_provider_key("MINIMAX_API_KEY", "minimax")
== "tts-pool-key"
)
def test_xai_env_fallback_consults_pool(self):
from tools.xai_http import resolve_xai_http_credentials
def fake_load_pool(pid):
# xai-oauth pool empty → OAuth path yields no token;
# manual `hermes auth add xai` pool has the key.
return _fake_pool("xai-pool-key" if pid == "xai" else "")
with patch("agent.credential_pool.load_pool", side_effect=fake_load_pool):
creds = resolve_xai_http_credentials()
assert creds["api_key"] == "xai-pool-key"
assert creds["provider"] == "xai"
def test_openai_audio_key_falls_back_to_pool(self):
from tools.tool_backend_helpers import resolve_openai_audio_api_key
def fake_load_pool(pid):
return _fake_pool("oai-pool-key" if pid == "openai-api" else "")
with patch("agent.credential_pool.load_pool", side_effect=fake_load_pool):
assert resolve_openai_audio_api_key() == "oai-pool-key"

View file

@ -2,12 +2,15 @@
from __future__ import annotations
import logging
import os
from pathlib import Path
from typing import Any, Dict
from utils import is_truthy_value
logger = logging.getLogger(__name__)
_DEFAULT_BROWSER_PROVIDER = "local"
_DEFAULT_MODAL_MODE = "auto"
@ -152,6 +155,103 @@ def _scoped_credential(name: str) -> str:
return (os.getenv(name, "") or "").strip()
def resolve_provider_secret(
env_var: str,
provider_id: str,
config_value: str = "",
env_getter=None,
) -> str:
"""Resolve a voice-provider API key. Single owner for STT/TTS key lookup.
Resolution order (fixes #68003 — keys added via ``hermes auth add
<provider>`` were invisible to the voice tools, which only consulted
env/.env):
1. An explicit ``config_value`` from config.yaml, when the caller has one.
2. The environment / ``~/.hermes/.env``. Under a multiplexed gateway turn
this reads the active profile's secret scope (authoritative — a scope
miss must NOT borrow another profile's ``os.environ``; see
``agent/secret_scope.py``). Outside multiplexing it reads
``hermes_cli.config.get_env_value`` (os.environ, then ``.env``),
matching the tools' historical behaviour exactly.
3. The credential pool / auth store for ``provider_id`` (``hermes auth
add <provider_id>``). Skipped under an active multiplex turn, where
only the profile scope is authoritative for credentials.
Never raises credential resolution must not hard-fail on a pool or
config read; returns ``""`` when no key is found anywhere.
``env_getter`` lets callers supply their module-level ``get_env_value``
wrapper (transcription_tools / tts_tool expose one that tests patch);
when omitted, ``hermes_cli.config.get_env_value`` is used directly.
"""
value = str(config_value or "").strip()
if value:
return value
# Scope-aware env read: under a multiplexed gateway turn this reads the
# active profile's secret scope (authoritative); otherwise it reads the
# scope overlay then os.environ (see ``agent.secret_scope.get_secret``).
key = _scoped_credential(env_var)
if key:
return key
try:
from agent.secret_scope import is_multiplex_active
if is_multiplex_active():
# Under multiplexing the profile scope is authoritative: do not
# fall through to the process-global .env or credential pool,
# which may belong to a different profile than the current turn.
return ""
except Exception: # pragma: no cover — secret_scope is in-repo
pass
if env_getter is not None:
key = str(env_getter(env_var) or "").strip()
else:
try:
from hermes_cli.config import get_env_value
key = str(get_env_value(env_var) or "").strip()
except ImportError: # pragma: no cover — config is in-repo
key = ""
if key:
return key
if not provider_id:
return ""
try:
from agent.credential_pool import load_pool
# `hermes auth add <provider>` keys a registry provider by its plain
# id, but a provider declared via config.yaml ``providers.<name>`` /
# ``custom_providers`` is pooled under ``custom:<name>`` (see
# agent/credential_pool.py CUSTOM_POOL_PREFIX). Check both.
for pool_key in (provider_id, f"custom:{provider_id}"):
pool = load_pool(pool_key)
if pool is None or not pool.has_credentials():
continue
entry = pool.peek()
if entry is None:
continue
key = str(
getattr(entry, "runtime_api_key", "")
or getattr(entry, "access_token", "")
or ""
).strip()
if key:
return key
except Exception as exc:
logger.debug(
"Could not read %s credential pool for %s: %s",
provider_id,
env_var,
exc,
)
return ""
def resolve_openai_audio_api_key() -> str:
"""Prefer the voice-tools key, but fall back to the normal OpenAI key.
@ -163,10 +263,15 @@ def resolve_openai_audio_api_key() -> str:
and get billed against a different profile's OpenAI account. Same
routing the WeChat send path and ``agent/vertex_adapter`` already use; see
``agent/secret_scope.py``.
Outside a multiplexed turn, ``OPENAI_API_KEY`` additionally falls back to
the credential pool (``hermes auth add openai-api``) via
``resolve_provider_secret`` same #68003 fix as the other voice
providers. The dedicated voice-tools override remains env/scope-only.
"""
return (
_scoped_credential("VOICE_TOOLS_OPENAI_KEY")
or _scoped_credential("OPENAI_API_KEY")
resolve_provider_secret("VOICE_TOOLS_OPENAI_KEY", "")
or resolve_provider_secret("OPENAI_API_KEY", "openai-api")
)

View file

@ -64,27 +64,19 @@ def get_env_value(name, default=None):
def _resolve_provider_key(env_var: str, provider_id: str) -> str:
"""Resolve an API key from env, .env, or the credential pool.
"""Resolve an STT provider API key via the shared voice-key resolver.
Used by TTS/STT providers (Mistral, ElevenLabs) that store keys
via ``hermes auth add <provider_id>``.
Delegates to ``tools.tool_backend_helpers.resolve_provider_secret``
the single owner of STT/TTS key resolution (config > env/.env > the
credential pool populated by ``hermes auth add <provider_id>``).
Resolved at call time so tests that reload the helpers module see the
live function.
"""
key = get_env_value(env_var)
if key:
return key
try:
from agent.credential_pool import load_pool
pool = load_pool(provider_id)
if pool and pool.has_credentials():
entry = pool.peek()
if entry:
key = getattr(entry, "access_token", "") or getattr(entry, "runtime_api_key", "")
key = str(key).strip()
if key:
return key
except Exception:
pass
return ""
from tools.tool_backend_helpers import resolve_provider_secret
except ImportError: # pragma: no cover — helpers are in-repo
return str(get_env_value(env_var) or "").strip()
return resolve_provider_secret(env_var, provider_id, env_getter=get_env_value)
# ---------------------------------------------------------------------------
# Optional imports — graceful degradation
@ -851,7 +843,7 @@ def _get_provider(stt_config: dict) -> str:
return "none"
if provider == "groq":
if _HAS_OPENAI and get_env_value("GROQ_API_KEY"):
if _HAS_OPENAI and _resolve_provider_key("GROQ_API_KEY", "groq"):
return "groq"
logger.warning(
"STT provider 'groq' configured but GROQ_API_KEY not set"
@ -894,7 +886,7 @@ def _get_provider(stt_config: dict) -> str:
return "none"
if provider == "deepinfra":
if _HAS_OPENAI and (get_env_value("DEEPINFRA_API_KEY") or "").strip():
if _HAS_OPENAI and _resolve_provider_key("DEEPINFRA_API_KEY", "deepinfra"):
return "deepinfra"
logger.warning(
"STT provider 'deepinfra' configured but DEEPINFRA_API_KEY not set "
@ -919,7 +911,7 @@ def _get_provider(stt_config: dict) -> str:
# Try lazy-install before falling through to cloud providers
if _try_lazy_install_stt():
return "local"
if _HAS_OPENAI and get_env_value("GROQ_API_KEY"):
if _HAS_OPENAI and _resolve_provider_key("GROQ_API_KEY", "groq"):
logger.info("No local STT available, using Groq Whisper API")
return "groq"
if _HAS_OPENAI and _has_openai_audio_backend():
@ -942,7 +934,7 @@ def _get_provider(stt_config: dict) -> str:
if _resolve_provider_key("ELEVENLABS_API_KEY", "elevenlabs"):
logger.info("No local STT available, using ElevenLabs Scribe STT API")
return "elevenlabs"
if _HAS_OPENAI and (get_env_value("DEEPINFRA_API_KEY") or "").strip():
if _HAS_OPENAI and _resolve_provider_key("DEEPINFRA_API_KEY", "deepinfra"):
logger.info("No local STT available, using DeepInfra Whisper API")
return "deepinfra"
return "none"
@ -1362,7 +1354,7 @@ def _transcribe_groq(file_path: str, model_name: str) -> Dict[str, Any]:
``HERMES_LOCAL_STT_LANGUAGE`` (env). When none is set, Groq
Whisper auto-detects.
"""
api_key = get_env_value("GROQ_API_KEY")
api_key = _resolve_provider_key("GROQ_API_KEY", "groq")
if not api_key:
return {"success": False, "transcript": "", "error": "GROQ_API_KEY not set"}
@ -1752,7 +1744,7 @@ def _transcribe_deepinfra(file_path: str, model_name: str) -> Dict[str, Any]:
``hermes_cli.models`` helpers so every DeepInfra surface resolves the
base URL and model ids identically.
"""
api_key = (get_env_value("DEEPINFRA_API_KEY") or "").strip()
api_key = _resolve_provider_key("DEEPINFRA_API_KEY", "deepinfra")
if not api_key:
return {"success": False, "transcript": "", "error": "DEEPINFRA_API_KEY not set"}

View file

@ -72,27 +72,20 @@ def get_env_value(name, default=None):
def _resolve_provider_key(env_var: str, provider_id: str) -> str:
"""Resolve an API key from env, .env, or the credential pool.
"""Resolve a TTS provider API key via the shared voice-key resolver.
Used by TTS providers (Mistral, ElevenLabs) that store keys
via ``hermes auth add <provider_id>``.
Delegates to ``tools.tool_backend_helpers.resolve_provider_secret``
the single owner of STT/TTS key resolution (config > env/.env > the
credential pool populated by ``hermes auth add <provider_id>``).
Resolved at call time so tests that reload the helpers module see the
live function.
"""
key = get_env_value(env_var)
if key:
return key
try:
from agent.credential_pool import load_pool
pool = load_pool(provider_id)
if pool and pool.has_credentials():
entry = pool.peek()
if entry:
key = getattr(entry, "access_token", "") or getattr(entry, "runtime_api_key", "")
key = str(key).strip()
if key:
return key
except Exception:
pass
return ""
from tools.tool_backend_helpers import resolve_provider_secret
except ImportError: # pragma: no cover — helpers are in-repo
return str(get_env_value(env_var) or "").strip()
return resolve_provider_secret(env_var, provider_id, env_getter=get_env_value)
from tools.managed_tool_gateway import resolve_managed_tool_gateway
from tools.tool_backend_helpers import (
managed_nous_tools_enabled,
@ -1285,7 +1278,7 @@ def _generate_deepinfra_tts(text: str, output_path: str, tts_config: Dict[str, A
the shared ``hermes_cli.models`` helpers so every DeepInfra surface
resolves them identically.
"""
api_key = (get_env_value("DEEPINFRA_API_KEY") or "").strip()
api_key = _resolve_provider_key("DEEPINFRA_API_KEY", "deepinfra")
if not api_key:
raise ValueError(
"DEEPINFRA_API_KEY not set. Run `hermes setup` to configure, "
@ -1566,7 +1559,7 @@ def _generate_minimax_tts(text: str, output_path: str, tts_config: Dict[str, Any
"""
import requests
api_key = (get_env_value("MINIMAX_API_KEY") or "")
api_key = (_resolve_provider_key("MINIMAX_API_KEY", "minimax") or "")
if not api_key:
raise ValueError("MINIMAX_API_KEY not set. Get one at https://platform.minimax.io/")
@ -1936,7 +1929,10 @@ def _generate_gemini_tts(text: str, output_path: str, tts_config: Dict[str, Any]
"""
import requests
api_key = (get_env_value("GEMINI_API_KEY") or get_env_value("GOOGLE_API_KEY") or "").strip()
api_key = (
_resolve_provider_key("GEMINI_API_KEY", "gemini")
or _resolve_provider_key("GOOGLE_API_KEY", "gemini")
)
if not api_key:
raise ValueError(
"GEMINI_API_KEY not set. Get one at https://aistudio.google.com/app/apikey"
@ -2782,9 +2778,9 @@ def check_tts_requirements() -> bool:
_import_openai_client()
except ImportError:
return False
return bool(get_env_value("DEEPINFRA_API_KEY"))
return bool(_resolve_provider_key("DEEPINFRA_API_KEY", "deepinfra"))
if provider == "minimax":
return bool(get_env_value("MINIMAX_API_KEY"))
return bool(_resolve_provider_key("MINIMAX_API_KEY", "minimax"))
if provider == "xai":
try:
from tools.xai_http import resolve_xai_http_credentials
@ -2793,7 +2789,10 @@ def check_tts_requirements() -> bool:
except Exception:
return False
if provider == "gemini":
return bool(get_env_value("GEMINI_API_KEY") or get_env_value("GOOGLE_API_KEY"))
return bool(
_resolve_provider_key("GEMINI_API_KEY", "gemini")
or _resolve_provider_key("GOOGLE_API_KEY", "gemini")
)
if provider == "mistral":
try:
_import_mistral_client()

View file

@ -311,7 +311,12 @@ def resolve_xai_http_credentials(
except Exception:
pass
api_key = str(get_env_value("XAI_API_KEY") or "").strip()
try:
from tools.tool_backend_helpers import resolve_provider_secret
api_key = resolve_provider_secret("XAI_API_KEY", "xai", env_getter=get_env_value)
except ImportError: # pragma: no cover — helpers are in-repo
api_key = str(get_env_value("XAI_API_KEY") or "").strip()
base_url = str(get_env_value("XAI_BASE_URL") or "https://api.x.ai/v1").strip().rstrip("/")
return {
"provider": "xai",