feat(providers): make all 33 providers pluggable under plugins/model-providers/

Every provider profile is now a self-contained plugin under
plugins/model-providers/<name>/, mirroring the plugins/platforms/
pattern established for IRC and Teams. The ProviderProfile ABC
stays in providers/; the per-provider profile data moves out.

- plugins/model-providers/<name>/__init__.py calls register_provider()
- plugins/model-providers/<name>/plugin.yaml declares kind: model-provider
- providers/__init__.py._discover_providers() lazily scans bundled plugins
  then $HERMES_HOME/plugins/model-providers/<name>/ (user override path)
- User plugins with the same name override bundled ones (last-writer-wins
  in register_provider)
- Legacy providers/<name>.py layout still supported for back-compat with
  out-of-tree editable installs
- Hermes PluginManager: new kind=model-provider; skipped like memory
  plugins (providers/ discovery owns them); standalone plugins with
  register_provider+ProviderProfile in their __init__.py auto-coerce to
  this kind (same heuristic as memory providers)
- skip_names extended to include 'model-providers' so the general
  PluginManager doesn't double-scan the category
- 4 new tests in tests/providers/test_plugin_discovery.py covering
  bundled discovery, user override, and general-loader isolation
- Docs updated: website/docs/developer-guide/adding-providers.md,
  provider-runtime.md, providers/README.md, plugins/model-providers/README.md

No API break: auth.py / config.py / doctor.py / models.py / runtime_provider.py /
model_metadata.py / auxiliary_client.py / chat_completions.py / run_agent.py
all still consume providers via get_provider_profile() / list_providers() —
they just now see plugin-discovered entries instead of pkgutil-iterated ones.

Third parties can now drop a single directory into
~/.hermes/plugins/model-providers/<name>/ to add or override an inference
provider without touching the repo.
This commit is contained in:
Teknium 2026-05-05 13:36:08 -07:00
parent 20a4f79ed1
commit 9022804d78
63 changed files with 585 additions and 309 deletions

View file

@ -173,7 +173,7 @@ def _get_enabled_plugins() -> Optional[set]:
# Data classes # Data classes
# --------------------------------------------------------------------------- # ---------------------------------------------------------------------------
_VALID_PLUGIN_KINDS: Set[str] = {"standalone", "backend", "exclusive", "platform"} _VALID_PLUGIN_KINDS: Set[str] = {"standalone", "backend", "exclusive", "platform", "model-provider"}
@dataclass @dataclass
@ -643,15 +643,17 @@ class PluginManager:
# - flat: ``plugins/disk-cleanup/plugin.yaml`` (standalone) # - flat: ``plugins/disk-cleanup/plugin.yaml`` (standalone)
# - category: ``plugins/image_gen/openai/plugin.yaml`` (backend) # - category: ``plugins/image_gen/openai/plugin.yaml`` (backend)
# #
# ``memory/`` and ``context_engine/`` are skipped at the top level — # ``memory/``, ``context_engine/``, and ``model-providers/`` are
# they have their own discovery systems. ``platforms/`` is a category # skipped at the top level — they have their own discovery systems
# holding platform adapters (scanned one level deeper below). # (plugins/memory/__init__.py, providers/__init__.py). ``platforms/``
# is a category holding platform adapters (scanned one level deeper
# below).
repo_plugins = get_bundled_plugins_dir() repo_plugins = get_bundled_plugins_dir()
manifests.extend( manifests.extend(
self._scan_directory( self._scan_directory(
repo_plugins, repo_plugins,
source="bundled", source="bundled",
skip_names={"memory", "context_engine", "platforms"}, skip_names={"memory", "context_engine", "platforms", "model-providers"},
) )
) )
manifests.extend( manifests.extend(
@ -709,6 +711,21 @@ class PluginManager:
) )
continue continue
# Model provider plugins are loaded by providers/__init__.py
# (its own lazy discovery keyed off first get_provider_profile()
# call). We record the manifest here for introspection but do
# not import the module — a second import would create two
# ProviderProfile instances and break the "last writer wins"
# override semantics between bundled and user plugins.
if manifest.kind == "model-provider":
loaded = LoadedPlugin(manifest=manifest, enabled=True)
self._plugins[lookup_key] = loaded
logger.debug(
"Skipping '%s' (model-provider, handled by providers/ discovery)",
lookup_key,
)
continue
# Built-in backends auto-load — they ship with hermes and must # Built-in backends auto-load — they ship with hermes and must
# just work. Selection among them (e.g. which image_gen backend # just work. Selection among them (e.g. which image_gen backend
# services calls) is driven by ``<category>.provider`` config, # services calls) is driven by ``<category>.provider`` config,
@ -886,6 +903,19 @@ class PluginManager:
"treating as kind='exclusive'", "treating as kind='exclusive'",
key, key,
) )
elif (
"register_provider" in source_text
and "ProviderProfile" in source_text
):
# Model provider plugin (calls register_provider()
# from ``providers`` with a ProviderProfile). Route
# to providers/__init__.py discovery.
kind = "model-provider"
logger.debug(
"Plugin %s: detected model provider, "
"treating as kind='model-provider'",
key,
)
except Exception: except Exception:
pass pass

View file

@ -0,0 +1,70 @@
# Model Provider Plugins
Each subdirectory is a self-contained provider profile plugin. The
directory layout mirrors `plugins/platforms/`:
```
plugins/model-providers/
├── openrouter/
│ ├── __init__.py # registers the ProviderProfile
│ └── plugin.yaml # manifest: name, kind, version, description
├── anthropic/
│ ├── __init__.py
│ └── plugin.yaml
└── ...
```
## How discovery works
`providers/__init__.py._discover_providers()` scans this directory (and
`$HERMES_HOME/plugins/model-providers/`) the first time anything calls
`get_provider_profile()` or `list_providers()`. Each `__init__.py` is
imported and expected to call `providers.register_provider(profile)`.
User plugins at `$HERMES_HOME/plugins/model-providers/<name>/` override
bundled plugins of the same name — last-writer-wins in
`register_provider()`. Drop a file there to replace a built-in.
## Adding a new provider
1. Create `plugins/model-providers/<your_provider>/__init__.py`:
```python
from providers import register_provider
from providers.base import ProviderProfile
my_provider = ProviderProfile(
name="your-provider",
aliases=("alias1", "alias2"),
display_name="Your Provider",
description="One-line description shown in the setup picker",
signup_url="https://your-provider.example.com/keys",
env_vars=("YOUR_PROVIDER_API_KEY", "YOUR_PROVIDER_BASE_URL"),
base_url="https://api.your-provider.example.com/v1",
default_aux_model="your-cheap-model",
)
register_provider(my_provider)
```
2. Create `plugins/model-providers/<your_provider>/plugin.yaml`:
```yaml
name: your-provider-profile
kind: model-provider
version: 1.0.0
description: Short sentence about the provider
author: Your Name
```
Nothing else needs to change. `auth.py`, `config.py`, `models.py`,
`doctor.py`, `model_metadata.py`, `runtime_provider.py`, and the
chat_completions transport all auto-wire from the registry.
## Non-trivial profiles
Override the `ProviderProfile` hooks in a subclass for per-provider
quirks — see `plugins/model-providers/openrouter/__init__.py` for
`build_extra_body` and `build_api_kwargs_extras` examples, and
`plugins/model-providers/gemini/__init__.py` for `thinking_config`
translation.

View file

@ -0,0 +1,5 @@
name: ai-gateway-provider
kind: model-provider
version: 1.0.0
description: Vercel AI Gateway
author: Nous Research

View file

@ -0,0 +1,5 @@
name: alibaba-coding-plan-provider
kind: model-provider
version: 1.0.0
description: Alibaba Cloud Coding Plan
author: Nous Research

View file

@ -0,0 +1,5 @@
name: alibaba-provider
kind: model-provider
version: 1.0.0
description: Alibaba DashScope (international)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: anthropic-provider
kind: model-provider
version: 1.0.0
description: Anthropic (Claude)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: arcee-provider
kind: model-provider
version: 1.0.0
description: Arcee AI
author: Nous Research

View file

@ -0,0 +1,5 @@
name: azure-foundry-provider
kind: model-provider
version: 1.0.0
description: Azure AI Foundry
author: Nous Research

View file

@ -0,0 +1,5 @@
name: bedrock-provider
kind: model-provider
version: 1.0.0
description: AWS Bedrock
author: Nous Research

View file

@ -0,0 +1,5 @@
name: copilot-acp-provider
kind: model-provider
version: 1.0.0
description: GitHub Copilot via ACP subprocess
author: Nous Research

View file

@ -0,0 +1,5 @@
name: copilot-provider
kind: model-provider
version: 1.0.0
description: GitHub Copilot
author: Nous Research

View file

@ -0,0 +1,5 @@
name: custom-provider
kind: model-provider
version: 1.0.0
description: Custom / Ollama / local OpenAI-compatible endpoint
author: Nous Research

View file

@ -0,0 +1,5 @@
name: deepseek-provider
kind: model-provider
version: 1.0.0
description: DeepSeek
author: Nous Research

View file

@ -0,0 +1,5 @@
name: gemini-provider
kind: model-provider
version: 1.0.0
description: Google Gemini (API key + Cloud Code OAuth)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: gmi-provider
kind: model-provider
version: 1.0.0
description: GMI Cloud
author: Nous Research

View file

@ -0,0 +1,5 @@
name: huggingface-provider
kind: model-provider
version: 1.0.0
description: HuggingFace Inference Providers
author: Nous Research

View file

@ -0,0 +1,5 @@
name: kilocode-provider
kind: model-provider
version: 1.0.0
description: Kilo Code
author: Nous Research

View file

@ -0,0 +1,5 @@
name: kimi-coding-provider
kind: model-provider
version: 1.0.0
description: Moonshot Kimi Coding (global + China)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: minimax-provider
kind: model-provider
version: 1.0.0
description: MiniMax M-series (global + China + OAuth)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: nous-provider
kind: model-provider
version: 1.0.0
description: Nous Research Portal
author: Nous Research

View file

@ -0,0 +1,5 @@
name: nvidia-provider
kind: model-provider
version: 1.0.0
description: NVIDIA NIM
author: Nous Research

View file

@ -0,0 +1,5 @@
name: ollama-cloud-provider
kind: model-provider
version: 1.0.0
description: Ollama Cloud
author: Nous Research

View file

@ -0,0 +1,5 @@
name: openai-codex-provider
kind: model-provider
version: 1.0.0
description: OpenAI Codex (Responses API)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: opencode-zen-provider
kind: model-provider
version: 1.0.0
description: OpenCode (Zen + Go)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: openrouter-provider
kind: model-provider
version: 1.0.0
description: OpenRouter aggregator
author: Nous Research

View file

@ -0,0 +1,5 @@
name: qwen-oauth-provider
kind: model-provider
version: 1.0.0
description: Qwen Portal (OAuth)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: stepfun-provider
kind: model-provider
version: 1.0.0
description: StepFun Step Plan
author: Nous Research

View file

@ -0,0 +1,5 @@
name: xai-provider
kind: model-provider
version: 1.0.0
description: xAI Grok (Responses API)
author: Nous Research

View file

@ -0,0 +1,5 @@
name: xiaomi-provider
kind: model-provider
version: 1.0.0
description: Xiaomi MiMo
author: Nous Research

View file

@ -0,0 +1,5 @@
name: zai-provider
kind: model-provider
version: 1.0.0
description: Z.AI / GLM
author: Nous Research

View file

@ -1,307 +1,78 @@
# providers/ # providers/
Single source of truth for every inference provider Hermes knows about. Registry and ABC for every inference provider Hermes knows about.
Each provider is declared once here as a `ProviderProfile`. Every other layer — Each provider is declared once as a `ProviderProfile`. Every other layer —
auth resolution, transport kwargs, model listing, runtime routing — reads from auth resolution, transport kwargs, model listing, runtime routing — reads from
these profiles instead of maintaining its own parallel data. these profiles instead of maintaining its own parallel data.
--- ---
## Directory layout ## Layout
``` ```
providers/ providers/
├── base.py ProviderProfile dataclass + OMIT_TEMPERATURE sentinel ├── base.py ProviderProfile dataclass + OMIT_TEMPERATURE sentinel
├── __init__.py Registry: register_provider(), get_provider_profile() ├── __init__.py Registry: register_provider(), get_provider_profile(), list_providers()
├── README.md This file └── README.md This file
├── # Simple providers — just identity + auth + endpoint
├── alibaba.py Alibaba Cloud DashScope
├── arcee.py Arcee AI
├── bedrock.py AWS Bedrock (api_mode=bedrock_converse)
├── deepseek.py DeepSeek
├── huggingface.py Hugging Face Inference API
├── kilocode.py Kilo Code
├── minimax.py MiniMax (international + CN)
├── nvidia.py NVIDIA NIM (default_max_tokens=16384)
├── ollama_cloud.py Ollama Cloud
├── stepfun.py StepFun
├── xiaomi.py Xiaomi MiMo
├── xai.py xAI Grok (api_mode=codex_responses)
├── zai.py Z.AI / GLM
├── # Medium — one or two quirks
├── anthropic.py Native Anthropic (x-api-key header, api_mode=anthropic_messages)
├── copilot.py GitHub Copilot (auth_type=copilot, reasoning per model)
├── copilot_acp.py Copilot ACP subprocess (api_mode=copilot_acp)
├── custom.py Custom/Ollama local (think=false, num_ctx)
├── gemini.py Google Gemini AI Studio + Cloud Code OAuth
├── kimi.py Kimi Coding (OMIT_TEMPERATURE, thinking, dual endpoint)
├── openai_codex.py OpenAI Codex OAuth (api_mode=codex_responses)
├── opencode.py OpenCode Zen + Go (per-model api_mode routing)
├── # Complex — subclasses with multiple overrides
├── nous.py Nous Portal (tags, attribution, reasoning omit-when-disabled)
├── openrouter.py OpenRouter (provider preferences, public model fetch)
├── qwen.py Qwen OAuth (message normalization, cache_control, vl_hires)
└── vercel.py Vercel AI Gateway (attribution headers, reasoning passthrough)
``` ```
The **profiles themselves** live as plugins under
`plugins/model-providers/<name>/` (bundled in this repo) and
`$HERMES_HOME/plugins/model-providers/<name>/` (per-user overrides). The
registry in `providers/__init__.py` lazily discovers them the first time any
consumer calls `get_provider_profile()` or `list_providers()`. See
`plugins/model-providers/README.md` for the plugin contract and examples.
--- ---
## ProviderProfile fields ## How it wires in
```python The registry is populated on first access. After that, every downstream
@dataclass layer reads from it:
class ProviderProfile:
# Identity
name: str # canonical ID — auto-registered as PROVIDER_REGISTRY key for new api-key providers
api_mode: str # "chat_completions" | "anthropic_messages" |
# "codex_responses" | "bedrock_converse" | "copilot_acp"
aliases: tuple # alternate names resolved by get_provider_profile()
# Auth & endpoints - `hermes_cli/auth.py` extends `PROVIDER_REGISTRY` with every api-key
env_vars: tuple # env var names holding the API key, in priority order profile it sees (skipping `copilot`, `kimi-coding`, `kimi-coding-cn`,
base_url: str # default inference endpoint `zai`, `openrouter`, `custom` — those need bespoke token resolution).
models_url: str # explicit models endpoint; falls back to {base_url}/models - `hermes_cli/models.py` extends `CANONICAL_PROVIDERS` and calls
# set when the models catalog lives at a different URL `profile.fetch_models()` inside `provider_model_ids()`.
# (e.g. OpenRouter: public /api/v1/models vs /api/v1 inference) - `hermes_cli/doctor.py` adds a `/models` health check for each
auth_type: str # "api_key" | "oauth_device_code" | "oauth_external" | `auth_type="api_key"` profile.
# "copilot" | "aws" | "external_process" - `hermes_cli/config.py` injects every `env_var` into
`OPTIONAL_ENV_VARS` so the setup wizard knows about it.
# Client-level quirks - `hermes_cli/runtime_provider.py` reads `profile.api_mode` as a fallback
default_headers: dict # extra HTTP headers sent on every request when URL detection finds nothing.
- `agent/model_metadata.py` maps hostname → provider via
# Request-level quirks `profile.get_hostname()`.
fixed_temperature: Any # None = use caller's default; OMIT_TEMPERATURE = don't send - `agent/auxiliary_client.py` reads `profile.default_aux_model` first
default_max_tokens: int|None # inject max_tokens when caller omits it before falling back to the legacy hardcoded dict.
default_aux_model: str # cheap model for auxiliary tasks (compression, vision, etc.) - `agent/transports/chat_completions.py::_build_kwargs_from_profile()`
# empty string = use main model (default) invokes `profile.prepare_messages()`, `profile.build_extra_body()`,
``` and `profile.build_api_kwargs_extras()` on every call.
- `run_agent.py` passes `provider_profile=<ProviderProfile>` so the
transport takes the profile path instead of the legacy flag path.
--- ---
## Hooks (override in a subclass) ## Adding a provider
| Method | When to override | See `plugins/model-providers/README.md` — drop a new directory there (or
|--------|-----------------| under `$HERMES_HOME/plugins/model-providers/` for a private plugin).
| `prepare_messages(messages)` | Provider needs message pre-processing (Qwen: string → list-of-parts, cache_control) |
| `build_extra_body(*, session_id, **ctx)` | Provider-specific `extra_body` fields (Nous: tags, OpenRouter: provider preferences) |
| `build_api_kwargs_extras(*, reasoning_config, **ctx)` | Returns `(extra_body_additions, top_level_kwargs)` — use when some fields go to `extra_body` and some go top-level (Kimi: `reasoning_effort` top-level; OpenRouter: `reasoning` in extra_body) |
| `fetch_models(*, api_key, timeout)` | Custom model listing (Anthropic: x-api-key header; OpenRouter: public endpoint, no auth; Bedrock/copilot-acp: return None) |
All hooks have safe defaults — only override what differs from the base.
--- ---
## How to add a new provider ## Hooks you can override on `ProviderProfile`
### 1. Simple (standard OpenAI-compatible endpoint) | Hook | Purpose |
|------|---------|
```python | `get_hostname()` | URL-based detection — default derives from `base_url`. |
# providers/myprovider.py | `prepare_messages(msgs)` | Provider-specific message preprocessing (Qwen normalises to list-of-parts, injects `cache_control`). |
from providers import register_provider | `build_extra_body(**ctx)` | Provider-specific `extra_body` (OpenRouter provider prefs, Gemini `thinking_config`). |
from providers.base import ProviderProfile | `build_api_kwargs_extras(**ctx)` | `(extra_body_additions, top_level_kwargs)` — Kimi puts reasoning_effort top-level, Qwen splits `enable_thinking`/`thinking_budget`. |
| `fetch_models(*, api_key)` | Live catalog fetch — default hits `{models_url or base_url}/models` with Bearer auth. Override for no-REST providers (Bedrock), OAuth catalogs (Anthropic), or public catalogs (OpenRouter). |
myprovider = ProviderProfile(
name="myprovider", # must match id in hermes_cli/auth.py PROVIDER_REGISTRY
aliases=("my-provider", "myp"),
api_mode="chat_completions",
env_vars=("MYPROVIDER_API_KEY",),
base_url="https://api.myprovider.com/v1",
auth_type="api_key",
)
register_provider(myprovider)
```
The default `fetch_models()` will call `GET https://api.myprovider.com/v1/models`
with Bearer auth automatically. No override needed for standard `/v1/models`.
### 2. With quirks (subclass)
```python
# providers/myprovider.py
from typing import Any
from providers import register_provider
from providers.base import ProviderProfile
class MyProviderProfile(ProviderProfile):
"""My provider — custom reasoning header."""
def build_api_kwargs_extras(
self,
*,
reasoning_config: dict | None = None,
**ctx: Any,
) -> tuple[dict[str, Any], dict[str, Any]]:
extra_body: dict[str, Any] = {}
if reasoning_config:
extra_body["my_reasoning"] = reasoning_config.get("effort", "medium")
return extra_body, {}
def fetch_models(
self,
*,
api_key: str | None = None,
timeout: float = 8.0,
) -> list[str] | None:
# Override only if your endpoint differs from standard /v1/models
return super().fetch_models(api_key=api_key, timeout=timeout)
myprovider = MyProviderProfile(
name="myprovider",
aliases=("myp",),
env_vars=("MYPROVIDER_API_KEY",),
base_url="https://api.myprovider.com/v1",
)
register_provider(myprovider)
```
### 3. Wire it up
After creating the file, add `name` to the `_PROFILE_ACTIVE_PROVIDERS` set in
`run_agent.py` once you've verified parity against the legacy flag path. Start
with a simple provider (no message prep, no reasoning quirks) and work up.
--- ---
## fetch_models contract ## Configuration fields
```python Full reference in `providers/base.py` dataclass definition.
def fetch_models(
self,
*,
api_key: str | None = None,
timeout: float = 8.0,
) -> list[str] | None:
...
```
- Returns `list[str]`: model IDs from the provider's live endpoint.
- Returns `None`: provider doesn't support REST model listing (Bedrock, copilot-acp),
or the request failed. Callers **must** fall back to `_PROVIDER_MODELS` on `None`.
- Never raises — swallow exceptions and return `None`.
- Default implementation: `GET {base_url}/models` with Bearer auth. Works for any
standard OpenAI-compatible provider.
**Override when:**
- Auth header is not `Bearer` (Anthropic: `x-api-key`)
- Endpoint path differs from `/models` AND you can't just set `models_url` (OpenRouter: public endpoint, pass `api_key=None` explicitly)
- Response format differs (extra wrapping, non-standard `id` field)
- Provider has no REST endpoint (Bedrock, copilot-acp → return `None`)
- Filtering needed post-fetch (only tool-capable models, etc.)
Use `models_url` instead of overriding when the only difference is the URL:
```python
# No subclass needed — just set models_url
myprovider = ProviderProfile(
name="myprovider",
base_url="https://api.myprovider.com/v1",
models_url="https://catalog.myprovider.com/models", # different host
)
```
---
## Debugging
### Check if a provider resolves
```python
from providers import get_provider_profile
p = get_provider_profile("myprovider")
print(p) # ProviderProfile(name='myprovider', ...)
print(p.base_url)
print(p.api_mode)
```
### Check all registered providers
```python
from providers import _REGISTRY
print(list(_REGISTRY.keys()))
```
### Test live model fetch
```python
import os
from providers import get_provider_profile
p = get_provider_profile("myprovider")
key = os.getenv("MYPROVIDER_API_KEY")
models = p.fetch_models(api_key=key, timeout=5.0)
print(models) # list of model IDs, or None on failure
```
### Test alias resolution
```python
from providers import get_provider_profile
# All of these should return the same profile
assert get_provider_profile("openrouter").name == "openrouter"
assert get_provider_profile("or").name == "openrouter"
```
### Run the provider test suite
```bash
# From the repo root
source venv/bin/activate
python -m pytest tests/providers/ -v
```
### Check ruff + ty compliance
```bash
source venv/bin/activate
ruff format providers/*.py
ruff check providers/*.py --select UP,E,F,I,W
ty check providers/*.py
```
---
## Common mistakes
**Wrong `name`** — must be the same string that appears as the key in
`hermes_cli/auth.py` `PROVIDER_REGISTRY`. New api-key providers auto-register
into `PROVIDER_REGISTRY` from the profile, so the name IS the key. For providers
with a pre-existing `PROVIDER_REGISTRY` entry, use the exact `id` field value.
**Wrong `env_vars`** — separate API-key vars from base-URL override vars in the
tuple. Env vars that end with `_BASE_URL` or `_URL` are treated as URL overrides;
everything else is treated as an API key. Getting this wrong causes the doctor
health check to send a URL string as a Bearer token.
**Wrong `base_url`** — several providers have non-obvious paths:
`stepfun: /step_plan/v1`, `opencode-go: /zen/go/v1`. The profile's `base_url`
is also used as the `inference_base_url` when auto-registering into `PROVIDER_REGISTRY`
for new providers, so it must be correct for auth resolution to work.
**Skipping `api_mode`** — defaults to `chat_completions`. Providers that use
`anthropic_messages`, `codex_responses`, `bedrock_converse`, or `copilot_acp`
must set it explicitly.
**Forgetting `register_provider()`** — auto-discovery runs `pkgutil.iter_modules`
over the package and imports each module, but only if `register_provider()` is
called at module level. Without it the profile is never in `_REGISTRY`.
**`fetch_models` returning the wrong shape** — must return `list[str]` (plain
model IDs), not `list[tuple]` or `list[dict]`. Callers expect plain strings.
**Wrong `build_api_kwargs_extras` return shape** — must return a 2-tuple
`(extra_body_dict, top_level_dict)`. Returning a single dict causes a
`ValueError: not enough values to unpack` in the transport.
**`build_api_kwargs_extras` wrong tuple** — must return `(extra_body_dict,
top_level_dict)`. Returning a flat dict or swapping the order silently sends
fields to the wrong place.

View file

@ -1,25 +1,62 @@
"""Provider module registry. """Provider module registry.
Auto-discovers ProviderProfile instances from providers/*.py modules. Provider profiles can live in two places:
Each module should define a module-level PROVIDER or PROVIDERS list.
1. Bundled plugins: ``plugins/model-providers/<name>/`` (shipped with hermes-agent)
2. User plugins: ``$HERMES_HOME/plugins/model-providers/<name>/``
Each plugin directory contains:
- ``__init__.py`` calls ``register_provider(profile)`` at import
- ``plugin.yaml`` manifest (name, kind: model-provider, version, description)
Discovery is lazy: the first call to ``get_provider_profile()`` or
``list_providers()`` scans both locations and imports every plugin. User
plugins override bundled plugins on name collision (last-writer-wins), so
third parties can monkey-patch or replace any built-in profile without
editing the repo.
For backward compatibility, ``providers/*.py`` files (other than ``base.py``
and ``__init__.py``) are still discovered via ``pkgutil.iter_modules``.
This lets out-of-tree users drop a single-file profile into an editable
install without the plugin dir structure. New profiles should prefer the
plugin layout.
Usage::
Usage:
from providers import get_provider_profile from providers import get_provider_profile
profile = get_provider_profile("nvidia") # returns ProviderProfile or None profile = get_provider_profile("nvidia") # ProviderProfile or None
profile = get_provider_profile("kimi") # checks name + aliases profile = get_provider_profile("kimi") # checks name + aliases
""" """
from __future__ import annotations from __future__ import annotations
import importlib
import importlib.util
import logging
import sys
from pathlib import Path
from providers.base import OMIT_TEMPERATURE, ProviderProfile # noqa: F401 from providers.base import OMIT_TEMPERATURE, ProviderProfile # noqa: F401
logger = logging.getLogger(__name__)
_REGISTRY: dict[str, ProviderProfile] = {} _REGISTRY: dict[str, ProviderProfile] = {}
_ALIASES: dict[str, str] = {} _ALIASES: dict[str, str] = {}
_discovered = False _discovered = False
# Repo-root ``plugins/model-providers/`` — populated at discovery time.
_BUNDLED_PLUGINS_DIR = (
Path(__file__).resolve().parent.parent / "plugins" / "model-providers"
)
def register_provider(profile: ProviderProfile) -> None: def register_provider(profile: ProviderProfile) -> None:
"""Register a provider profile by name and aliases.""" """Register a provider profile by name and aliases.
Later registrations with the same name replace earlier ones so user
plugins under ``$HERMES_HOME/plugins/model-providers/`` can override
bundled profiles without editing repo code.
"""
_REGISTRY[profile.name] = profile _REGISTRY[profile.name] = profile
for alias in profile.aliases: for alias in profile.aliases:
_ALIASES[alias] = profile.name _ALIASES[alias] = profile.name
@ -51,26 +88,104 @@ def list_providers() -> list[ProviderProfile]:
return result return result
def _user_plugins_dir() -> Path | None:
"""Return ``$HERMES_HOME/plugins/model-providers/`` if it exists."""
try:
from hermes_constants import get_hermes_home
d = get_hermes_home() / "plugins" / "model-providers"
return d if d.is_dir() else None
except Exception:
return None
def _import_plugin_dir(plugin_dir: Path, source: str) -> None:
"""Import a single plugin directory so it self-registers.
``source`` is "bundled" or "user", used only for log messages.
"""
init_file = plugin_dir / "__init__.py"
if not init_file.exists():
return
# Give bundled plugins a stable import path (``plugins.model_providers.<name>``)
# so relative imports within the plugin work. User plugins load via
# ``importlib.util.spec_from_file_location`` with a unique module name so
# multiple HERMES_HOME profiles don't alias each other.
safe_name = plugin_dir.name.replace("-", "_")
if source == "bundled":
module_name = f"plugins.model_providers.{safe_name}"
else:
module_name = f"_hermes_user_provider_{safe_name}"
if module_name in sys.modules:
return # already imported
try:
spec = importlib.util.spec_from_file_location(
module_name, init_file, submodule_search_locations=[str(plugin_dir)]
)
if spec is None or spec.loader is None:
return
module = importlib.util.module_from_spec(spec)
sys.modules[module_name] = module
spec.loader.exec_module(module)
except Exception as exc:
logger.warning(
"Failed to load %s provider plugin %s: %s", source, plugin_dir.name, exc
)
sys.modules.pop(module_name, None)
def _discover_providers() -> None: def _discover_providers() -> None:
"""Import all provider modules to trigger registration.""" """Populate the registry by importing every provider plugin.
Order:
1. Bundled plugins at ``<repo>/plugins/model-providers/<name>/``
2. User plugins at ``$HERMES_HOME/plugins/model-providers/<name>/``
3. Legacy per-file modules at ``providers/<name>.py`` (back-compat)
Each step imports its plugins, which call ``register_provider()`` at
module-level. Later steps win on name collision.
"""
global _discovered global _discovered
if _discovered: if _discovered:
return return
_discovered = True _discovered = True
import importlib # 1. Bundled plugins — shipped with hermes-agent.
import pkgutil if _BUNDLED_PLUGINS_DIR.is_dir():
for child in sorted(_BUNDLED_PLUGINS_DIR.iterdir()):
if not child.is_dir() or child.name.startswith(("_", ".")):
continue
_import_plugin_dir(child, "bundled")
import providers as _pkg # 2. User plugins — under $HERMES_HOME/plugins/model-providers/<name>/.
# These can override any bundled profile of the same name (last-writer-wins
# in register_provider()).
user_dir = _user_plugins_dir()
if user_dir is not None:
for child in sorted(user_dir.iterdir()):
if not child.is_dir() or child.name.startswith(("_", ".")):
continue
_import_plugin_dir(child, "user")
for _importer, modname, _ispkg in pkgutil.iter_modules(_pkg.__path__): # 3. Legacy single-file profiles at providers/<name>.py. Kept for
if modname.startswith("_") or modname == "base": # back-compat — if someone drops a ``providers/foo.py`` into an
continue # editable install, it still works without the plugin layout.
try: try:
importlib.import_module(f"providers.{modname}") import pkgutil
except ImportError as e:
import logging
logging.getLogger(__name__).warning( import providers as _pkg
"Failed to import provider module %s: %s", modname, e
) for _importer, modname, _ispkg in pkgutil.iter_modules(_pkg.__path__):
if modname.startswith("_") or modname == "base":
continue
try:
importlib.import_module(f"providers.{modname}")
except ImportError as exc:
logger.warning(
"Failed to import legacy provider module %s: %s", modname, exc
)
except Exception:
pass

View file

@ -0,0 +1,145 @@
"""Tests for the model-providers plugin discovery system.
Verifies that:
1. All bundled providers at plugins/model-providers/<name>/ are discovered
2. User plugins at $HERMES_HOME/plugins/model-providers/<name>/ override bundled
3. plugin.yaml manifests with kind=model-provider are correctly categorized
"""
from __future__ import annotations
import importlib
import sys
from pathlib import Path
import pytest
REPO_ROOT = Path(__file__).resolve().parents[2]
def _clear_provider_caches():
"""Force providers/__init__.py to re-discover on next list_providers()."""
import providers as _pkg
_pkg._REGISTRY.clear()
_pkg._ALIASES.clear()
_pkg._discovered = False
# Evict any cached plugin modules so the next import re-executes.
for mod in list(sys.modules.keys()):
if (
mod.startswith("plugins.model_providers")
or mod.startswith("_hermes_user_provider")
):
del sys.modules[mod]
def test_bundled_plugins_discovered():
"""Every plugins/model-providers/<name>/ should contain a plugin.yaml + __init__.py."""
plugins_dir = REPO_ROOT / "plugins" / "model-providers"
assert plugins_dir.is_dir(), f"Missing {plugins_dir}"
child_dirs = [c for c in plugins_dir.iterdir() if c.is_dir()]
assert len(child_dirs) >= 28, f"Expected at least 28 provider plugins, found {len(child_dirs)}"
for child in child_dirs:
assert (child / "__init__.py").exists(), f"{child.name} missing __init__.py"
assert (child / "plugin.yaml").exists(), f"{child.name} missing plugin.yaml"
def test_all_33_profiles_register():
"""After discovery, the registry must contain exactly 33 distinct profiles."""
_clear_provider_caches()
from providers import list_providers
profiles = list_providers()
names = sorted(p.name for p in profiles)
assert len(names) == 33, f"Expected 33 profiles, got {len(names)}: {names}"
# Spot-check representative providers from different categories
for required in (
"openrouter", "anthropic", "custom", "bedrock", "openai-codex",
"minimax-oauth", "gmi", "xiaomi", "alibaba-coding-plan",
):
assert required in names, f"Missing profile: {required}"
def test_user_plugin_overrides_bundled(tmp_path, monkeypatch):
"""A user plugin with the same name must override the bundled profile."""
# Point HERMES_HOME at a fresh temp dir
hermes_home = tmp_path / ".hermes"
hermes_home.mkdir()
monkeypatch.setenv("HERMES_HOME", str(hermes_home))
# get_hermes_home() may be module-cached depending on codebase; ensure the
# env var is the source of truth. Most code paths re-read it each call.
# Drop a user plugin that replaces 'gmi'
user_gmi = hermes_home / "plugins" / "model-providers" / "gmi"
user_gmi.mkdir(parents=True)
(user_gmi / "__init__.py").write_text(
"from providers import register_provider\n"
"from providers.base import ProviderProfile\n"
"\n"
"custom_gmi = ProviderProfile(\n"
' name="gmi",\n'
' aliases=("gmi-user-override-test",),\n'
' env_vars=("GMI_API_KEY",),\n'
' base_url="https://user-override.example.com/v1",\n'
' auth_type="api_key",\n'
")\n"
"register_provider(custom_gmi)\n"
)
(user_gmi / "plugin.yaml").write_text(
"name: gmi-user-override\n"
"kind: model-provider\n"
"version: 0.0.1\n"
"description: Test user override\n"
)
_clear_provider_caches()
from providers import get_provider_profile
gmi = get_provider_profile("gmi")
assert gmi is not None
assert gmi.base_url == "https://user-override.example.com/v1", (
f"User override not applied; got base_url={gmi.base_url!r}"
)
assert "gmi-user-override-test" in gmi.aliases
# Clean up: reset discovery state so other tests see the bundled version
_clear_provider_caches()
def test_general_plugin_manager_skips_model_provider_kind(tmp_path, monkeypatch):
"""The general PluginManager must NOT import model-provider plugins
(providers/__init__.py handles them). It records the manifest only."""
from hermes_cli import plugins as plugin_mod
hermes_home = tmp_path / ".hermes"
hermes_home.mkdir()
monkeypatch.setenv("HERMES_HOME", str(hermes_home))
# Create a user-installed plugin with an explicit kind: model-provider.
user_plugin = hermes_home / "plugins" / "test-model-provider"
user_plugin.mkdir(parents=True)
(user_plugin / "plugin.yaml").write_text(
"name: test-model-provider\n"
"kind: model-provider\n"
"version: 0.0.1\n"
)
(user_plugin / "__init__.py").write_text(
# Intentionally broken import — if the general loader tries to
# import this module, the test will fail with ImportError.
"raise AssertionError('model-provider plugins must not be imported by PluginManager')\n"
)
# Fresh manager
manager = plugin_mod.PluginManager()
manager.discover_and_load(force=True)
# The manifest should be recorded but not loaded
loaded = manager._plugins.get("test-model-provider")
assert loaded is not None
assert loaded.manifest.kind == "model-provider"
# No import means the module must NOT be in the plugins list as a loaded one.
# We check that the general loader didn't crash and didn't raise from the
# broken __init__.py.

View file

@ -99,10 +99,12 @@ If your provider is just an OpenAI-compatible endpoint that authenticates with a
All you need is: All you need is:
1. A file in `providers/` (e.g. `providers/myprovider.py`) that calls `register_provider()` with the provider config. 1. A plugin directory under `plugins/model-providers/<your-provider>/` containing:
2. That's it. `auth.py` auto-registers every file in `providers/` at startup via a module-level import sweep. - `__init__.py` — calls `register_provider(profile)` at module-level
- `plugin.yaml` — manifest (name, kind: model-provider, version, description)
2. That's it. Provider plugins auto-load the first time anything calls `get_provider_profile()` or `list_providers()` — bundled plugins (this repo) and user plugins at `$HERMES_HOME/plugins/model-providers/` both get picked up.
When you add a `providers/*.py` file and call `register_provider()`, the following wire up automatically: When you add a plugin and it calls `register_provider()`, the following wire up automatically:
1. `PROVIDER_REGISTRY` entry in `auth.py` (credential resolution, env-var lookup) 1. `PROVIDER_REGISTRY` entry in `auth.py` (credential resolution, env-var lookup)
2. `api_mode` set to `chat_completions` 2. `api_mode` set to `chat_completions`
@ -117,7 +119,9 @@ When you add a `providers/*.py` file and call `register_provider()`, the followi
11. `HERMES_INFERENCE_PROVIDER` env-var override accepts the provider id 11. `HERMES_INFERENCE_PROVIDER` env-var override accepts the provider id
12. Fallback model activation can switch into the provider cleanly 12. Fallback model activation can switch into the provider cleanly
See `providers/nvidia.py` or `providers/gmi.py` as a template. User plugins at `$HERMES_HOME/plugins/model-providers/<name>/` override bundled plugins of the same name (last-writer-wins in `register_provider()`) — so third parties can monkey-patch or replace any built-in profile without editing the repo.
See `plugins/model-providers/nvidia/` or `plugins/model-providers/gmi/` as a template, and `plugins/model-providers/README.md` for the full contract.
## Full path: OAuth and complex providers ## Full path: OAuth and complex providers

View file

@ -20,9 +20,10 @@ Primary implementation:
- `hermes_cli/auth.py` — provider registry, `resolve_provider()` - `hermes_cli/auth.py` — provider registry, `resolve_provider()`
- `hermes_cli/model_switch.py` — shared `/model` switch pipeline (CLI + gateway) - `hermes_cli/model_switch.py` — shared `/model` switch pipeline (CLI + gateway)
- `agent/auxiliary_client.py` — auxiliary model routing - `agent/auxiliary_client.py` — auxiliary model routing
- `providers/` — declarative source for `api_mode`, `base_url`, `env_vars`, `fallback_models` (auto-registered into `auth.py` `PROVIDER_REGISTRY` at startup) - `providers/` — ABC + registry entry points (`ProviderProfile`, `register_provider`, `get_provider_profile`, `list_providers`)
- `plugins/model-providers/<name>/` — per-provider plugins (bundled) that declare `api_mode`, `base_url`, `env_vars`, `fallback_models` and register themselves into the registry on first access. User plugins at `$HERMES_HOME/plugins/model-providers/<name>/` override bundled ones of the same name.
`get_provider_profile()` in `providers/` returns a typed dict for a given provider id. `runtime_provider.py` calls this at resolution time to get the canonical `base_url`, `env_vars` priority list, `api_mode`, and `fallback_models` without needing to duplicate that data in multiple files. Adding a new `providers/*.py` file that calls `register_provider()` is enough for `runtime_provider.py` to pick it up — no branch needed in the resolver itself. `get_provider_profile()` in `providers/` returns a `ProviderProfile` for a given provider id. `runtime_provider.py` calls this at resolution time to get the canonical `base_url`, `env_vars` priority list, `api_mode`, and `fallback_models` without needing to duplicate that data in multiple files. Adding a new plugin under `plugins/model-providers/<your-provider>/` (or `$HERMES_HOME/plugins/model-providers/<your-provider>/`) that calls `register_provider()` is enough for `runtime_provider.py` to pick it up — no branch needed in the resolver itself.
If you are trying to add a new first-class inference provider, read [Adding Providers](./adding-providers.md) alongside this page. If you are trying to add a new first-class inference provider, read [Adding Providers](./adding-providers.md) alongside this page.