Merge pull request #104284 from NousResearch/fix/nous-anthropic-chat-wire
fix(nous): route anthropic/* over chat/completions by default; the native Portal route re-writes the cache on 14-20% of calls (nous.anthropic_wire)
This commit is contained in:
@@ -2284,6 +2284,13 @@ DEFAULT_CONFIG = {
|
||||
# server-issued credential lifetime (raising above it has no effect). 0 disables the
|
||||
# keepalive thread.
|
||||
"keepalive_interval_seconds": 900,
|
||||
# anthropic_wire: which Portal route carries anthropic/* models. "chat" =
|
||||
# /v1/chat/completions (default for now); "native" = /v1/messages, the Anthropic
|
||||
# Messages wire (signed thinking passthrough, native cache_control scopes). Native is the
|
||||
# better wire but re-writes the previous turn's cache on 14-20% of consecutive calls in
|
||||
# concurrent tool loops (measured 2026-09-06; NousResearch/api#227), so chat is the
|
||||
# default until that is fixed.
|
||||
"anthropic_wire": "chat",
|
||||
},
|
||||
# Google Vertex AI (Gemini). Auth is OAuth2 from a service-account JSON or ADC, NOT an API key;
|
||||
# the credential path lives in .env (VERTEX_CREDENTIALS_PATH / GOOGLE_APPLICATION_CREDENTIALS).
|
||||
|
||||
+23
-4
@@ -309,14 +309,33 @@ def host_mandated_api_mode(base_url: str = "") -> Optional[str]:
|
||||
|
||||
|
||||
def nous_api_mode(model: str = "") -> str:
|
||||
"""Wire protocol for a Nous Portal model: Portal serves its ``anthropic/*`` catalog on a native
|
||||
Messages route alongside OpenAI-compatible chat/completions for everything else. Empty/unknown
|
||||
model defaults to ``chat_completions`` (the historical Nous transport) as the safer path."""
|
||||
"""Wire protocol for a Nous Portal model. Portal serves its ``anthropic/*`` catalog on a native
|
||||
Messages route alongside OpenAI-compatible chat/completions for everything else.
|
||||
|
||||
``anthropic/*`` rides chat/completions by default for now (``nous.anthropic_wire``). Measured
|
||||
2026-09-06, 20 concurrent sessions x 6 tool calls on Fable 5.1, same account and hour: the
|
||||
native route re-wrote the previous turn on 14-20% of consecutive calls (4 runs; the cache read
|
||||
stopped at the prior breakpoint with byte-identical prefixes), chat/completions 0 of 320 pairs.
|
||||
That is 15-20% of a fan-out's cache-write bill. The cause is inside the portal's native route
|
||||
(NousResearch/api#227 carries the diagnostics); flip the default back to ``native`` when it is
|
||||
fixed. Cost of ``chat``: prior-turn thinking travels as OpenAI-style reasoning fields instead of
|
||||
signed native blocks, and cache_control scopes are translated by the portal's adapter.
|
||||
Empty/unknown model defaults to ``chat_completions`` (the historical Nous transport)."""
|
||||
if str(model or "").strip().lower().startswith("anthropic/"):
|
||||
return "anthropic_messages"
|
||||
return "anthropic_messages" if _nous_anthropic_wire() == "native" else "chat_completions"
|
||||
return "chat_completions"
|
||||
|
||||
|
||||
def _nous_anthropic_wire() -> str:
|
||||
"""``nous.anthropic_wire``: ``"chat"`` (default) or ``"native"``. Anything else reads as ``chat``."""
|
||||
try:
|
||||
from hermes_cli.config import load_config_readonly
|
||||
value = str(((load_config_readonly().get("nous") or {}).get("anthropic_wire")) or "chat").strip().lower()
|
||||
except Exception:
|
||||
return "chat"
|
||||
return "native" if value == "native" else "chat"
|
||||
|
||||
|
||||
def determine_api_mode(provider: str, base_url: str = "", model: str = "") -> str:
|
||||
"""API mode (wire protocol) for a provider/endpoint: host-mandated mode, then Nous dual-wire
|
||||
(model-derived — the overlay alone says openai_chat and would pin Claude on the wrong wire),
|
||||
|
||||
@@ -21,8 +21,18 @@ from unittest.mock import MagicMock, patch
|
||||
import pytest
|
||||
|
||||
from hermes_cli import runtime_provider as rp
|
||||
from hermes_cli import providers as _providers
|
||||
from hermes_cli.providers import nous_api_mode
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def _native_wire_selected(monkeypatch):
|
||||
"""These contracts describe the native wire, which is now opt-in (``nous.anthropic_wire:
|
||||
native``; default ``chat`` since 2026-09-06, see ``nous_api_mode``). Select it here so the
|
||||
wire keeps working for the flip-back; the default's own contract is in
|
||||
``test_nous_anthropic_wire_default.py``."""
|
||||
monkeypatch.setattr(_providers, "_nous_anthropic_wire", lambda: "native")
|
||||
|
||||
PORTAL_URL = "https://inference-api.nousresearch.com/v1"
|
||||
# Staging / preview hosts used via NOUS_INFERENCE_BASE_URL — not the prod
|
||||
# hostname, so Portal behaviour must key off provider=nous.
|
||||
|
||||
@@ -0,0 +1,54 @@
|
||||
"""``nous.anthropic_wire`` selects the Portal route for ``anthropic/*``: ``chat`` (default) rides
|
||||
/v1/chat/completions, ``native`` rides /v1/messages. Read through the real config loader against a
|
||||
temp HERMES_HOME (the loader is keyed on config path + mtime, so a fresh home is a fresh read), and
|
||||
through ``resolve_runtime_provider`` so the api_mode a live agent gets is what is asserted."""
|
||||
from __future__ import annotations
|
||||
|
||||
import pytest
|
||||
|
||||
from hermes_cli import providers as _providers
|
||||
from hermes_cli import runtime_provider as rp
|
||||
|
||||
PORTAL = "https://inference-api.nousresearch.com/v1"
|
||||
|
||||
|
||||
def _cfg(tmp_path, body: str, monkeypatch):
|
||||
monkeypatch.setenv("HERMES_HOME", str(tmp_path))
|
||||
(tmp_path / "config.yaml").write_text(body, encoding="utf-8")
|
||||
|
||||
|
||||
def _portal_creds(monkeypatch):
|
||||
monkeypatch.setattr(rp, "resolve_nous_runtime_credentials",
|
||||
lambda **kw: {"base_url": PORTAL, "api_key": "jwt", "source": "portal", "expires_at": None})
|
||||
monkeypatch.setattr(rp, "_get_model_config", lambda: {"provider": "nous"})
|
||||
|
||||
|
||||
def test_default_is_chat_for_anthropic_and_unchanged_for_everything_else(tmp_path, monkeypatch):
|
||||
_cfg(tmp_path, "model:\n default: anthropic/claude-fable-5.1\n", monkeypatch)
|
||||
assert _providers.nous_api_mode("anthropic/claude-fable-5.1") == "chat_completions"
|
||||
assert _providers.nous_api_mode("openai/gpt-5.6-sol") == "chat_completions"
|
||||
assert _providers.determine_api_mode("nous", PORTAL, "anthropic/claude-fable-5.1") == "chat_completions"
|
||||
|
||||
|
||||
def test_native_opt_in_restores_the_messages_wire_for_anthropic_only(tmp_path, monkeypatch):
|
||||
_cfg(tmp_path, "nous:\n anthropic_wire: native\n", monkeypatch)
|
||||
assert _providers.nous_api_mode("anthropic/claude-fable-5.1") == "anthropic_messages"
|
||||
assert _providers.nous_api_mode("openai/gpt-5.6-sol") == "chat_completions"
|
||||
|
||||
|
||||
@pytest.mark.parametrize("raw", ["''", "CHAT", "messages", "true", "1"])
|
||||
def test_anything_but_native_reads_as_chat(tmp_path, monkeypatch, raw):
|
||||
_cfg(tmp_path, f"nous:\n anthropic_wire: {raw}\n", monkeypatch)
|
||||
assert _providers.nous_api_mode("anthropic/claude-fable-5.1") == "chat_completions"
|
||||
|
||||
|
||||
def test_runtime_resolution_hands_a_live_agent_the_selected_wire(tmp_path, monkeypatch):
|
||||
"""The path an AIAgent takes: provider=nous + anthropic model -> api_mode, both settings."""
|
||||
_portal_creds(monkeypatch)
|
||||
_cfg(tmp_path, "model:\n provider: nous\n", monkeypatch)
|
||||
resolved = rp.resolve_runtime_provider(requested="nous", target_model="anthropic/claude-fable-5.1")
|
||||
assert (resolved["api_mode"], resolved["base_url"]) == ("chat_completions", PORTAL)
|
||||
|
||||
_cfg(tmp_path, "model:\n provider: nous\nnous:\n anthropic_wire: native\n", monkeypatch)
|
||||
resolved = rp.resolve_runtime_provider(requested="nous", target_model="anthropic/claude-fable-5.1")
|
||||
assert resolved["api_mode"] == "anthropic_messages"
|
||||
@@ -215,13 +215,16 @@ class TestFallbackChainAdvancement:
|
||||
assert mock_rpc.call_args.kwargs["explicit_api_key"] == "env-secret"
|
||||
|
||||
|
||||
def test_nous_anthropic_fallback_uses_the_messages_wire(self):
|
||||
"""Portal Claude fallbacks must not stay on chat_completions.
|
||||
def test_nous_anthropic_fallback_uses_the_messages_wire(self, monkeypatch):
|
||||
"""Portal Claude fallbacks must not stay on chat_completions when the native wire is selected.
|
||||
|
||||
``resolve_provider_client`` still returns an OpenAI client for Nous;
|
||||
activation has to re-derive api_mode from the model and rebuild the
|
||||
Anthropic client — otherwise the turn POSTs /chat/completions.
|
||||
Anthropic client — otherwise the turn POSTs /chat/completions. The wire
|
||||
is opt-in since 2026-09-06 (``nous.anthropic_wire``, see ``nous_api_mode``).
|
||||
"""
|
||||
from hermes_cli import providers as _providers
|
||||
monkeypatch.setattr(_providers, "_nous_anthropic_wire", lambda: "native")
|
||||
portal = "https://inference-api.nousresearch.com/v1"
|
||||
fbs = [
|
||||
{
|
||||
|
||||
@@ -423,7 +423,12 @@ class TestDelegateTask(unittest.TestCase):
|
||||
|
||||
def test_nous_child_rederives_api_mode_from_model(self):
|
||||
"""Portal is dual-wire — same provider + different model prefix must
|
||||
not inherit the parent's Messages/chat_completions mode verbatim."""
|
||||
not inherit the parent's Messages/chat_completions mode verbatim.
|
||||
Native wire selected (opt-in since 2026-09-06, ``nous.anthropic_wire``)."""
|
||||
with patch("hermes_cli.providers._nous_anthropic_wire", return_value="native"):
|
||||
self._nous_child_rederives_api_mode_from_model()
|
||||
|
||||
def _nous_child_rederives_api_mode_from_model(self):
|
||||
parent = _make_mock_parent(depth=0)
|
||||
parent.base_url = "https://inference-api.nousresearch.com/v1"
|
||||
parent.api_key = "portal-jwt"
|
||||
|
||||
@@ -234,6 +234,17 @@ omitted, Hermes keeps its normal provider and model capability detection.
|
||||
Older configs used a top-level `custom_providers:` list (with `base_url` instead of `api`). It still works and is auto-migrated to the `providers:` dict on `hermes update` (config v12).
|
||||
:::
|
||||
|
||||
### Nous Portal: which wire carries Claude
|
||||
|
||||
Nous Portal serves its `anthropic/*` models on two routes: OpenAI-compatible `/v1/chat/completions` and the native Anthropic Messages wire `/v1/messages`. `nous.anthropic_wire` picks one:
|
||||
|
||||
```yaml
|
||||
nous:
|
||||
anthropic_wire: chat # default. "native" = the Anthropic Messages wire
|
||||
```
|
||||
|
||||
`chat` is the default for now. The native wire is the better transport (signed thinking blocks pass through unchanged, native `cache_control` scopes), but in concurrent tool loops it currently re-writes the previous turn's prompt cache on 14–20% of consecutive calls, which is 15–20% of a fan-out's cache-write bill; the chat route measured 0 on the same test. Set `native` to opt back in (for example once the portal-side fix has shipped). Only `anthropic/*` models are affected; everything else on Nous already uses chat/completions.
|
||||
|
||||
## When does it take effect?
|
||||
|
||||
- **CLI** (`hermes chat`): next `hermes chat` invocation.
|
||||
|
||||
Reference in New Issue
Block a user