feat(plugins): pluggable image_gen backends + OpenAI provider (#13799)

* feat(plugins): pluggable image_gen backends + OpenAI provider

Adds a ImageGenProvider ABC so image generation backends register as
bundled plugins under `plugins/image_gen/<name>/`. The plugin scanner
gains three primitives to make this work generically:

- `kind:` manifest field (`standalone` | `backend` | `exclusive`).
  Bundled `kind: backend` plugins auto-load — no `plugins.enabled`
  incantation. User-installed backends stay opt-in.
- Path-derived keys: `plugins/image_gen/openai/` gets key
  `image_gen/openai`, so a future `tts/openai` cannot collide.
- Depth-2 recursion into category namespaces (parent dirs without a
  `plugin.yaml` of their own).

Includes `OpenAIImageGenProvider` as the first consumer (gpt-image-1.5
default, plus gpt-image-1, gpt-image-1-mini, DALL-E 3/2). Base64
responses save to `$HERMES_HOME/cache/images/`; URL responses pass
through.

FAL stays in-tree for this PR — a follow-up ports it into
`plugins/image_gen/fal/` so the in-tree `image_generation_tool.py`
slims down. The dispatch shim in `_handle_image_generate` only fires
when `image_gen.provider` is explicitly set to a non-FAL value, so
existing FAL setups are untouched.

- 41 unit tests (scanner recursion, kind parsing, gate logic,
  registry, OpenAI payload shapes)
- E2E smoke verified: bundled plugin autoloads, registers, and
  `_handle_image_generate` routes to OpenAI when configured

* fix(image_gen/openai): don't send response_format to gpt-image-*

The live API rejects it: 'Unknown parameter: response_format'
(verified 2026-04-21 with gpt-image-1.5). gpt-image-* models return
b64_json unconditionally, so the parameter was both unnecessary and
actively broken.

* feat(image_gen/openai): gpt-image-2 only, drop legacy catalog

gpt-image-2 is the latest/best OpenAI image model (released 2026-04-21)
and there's no reason to expose the older gpt-image-1.5 / gpt-image-1 /
dall-e-3 / dall-e-2 alongside it — slower, lower quality, or awkward
(dall-e-2 squares only). Trim the catalog down to a single model.

Live-verified end-to-end: landscape 1536x1024 render of a Moog-style
synth matches prompt exactly, 2.4MB PNG saved to cache.

* feat(image_gen/openai): expose gpt-image-2 as three quality tiers

Users pick speed/fidelity via the normal model picker instead of a
hidden quality knob. All three tier IDs resolve to the single underlying
gpt-image-2 API model with a different quality parameter:

  gpt-image-2-low     ~15s   fast iteration
  gpt-image-2-medium  ~40s   default
  gpt-image-2-high    ~2min  highest fidelity

Live-measured on OpenAI's API today: 15.4s / 40.8s / 116.9s for the
same 1024x1024 prompt.

Config:
  image_gen.openai.model: gpt-image-2-high
  # or
  image_gen.model: gpt-image-2-low
  # or env var for scripts/tests
  OPENAI_IMAGE_MODEL=gpt-image-2-medium

Live-verified end-to-end with the low tier: 18.8s landscape render of a
golden retriever in wildflowers, vision-confirmed exact match.

* feat(tools_config): plugin image_gen providers inject themselves into picker

'hermes tools' → Image Generation now shows plugin-registered backends
alongside Nous Subscription and FAL.ai without tools_config.py needing
to know about them. OpenAI appears as a third option today; future
backends appear automatically as they're added.

Mechanism:
- ImageGenProvider gains an optional get_setup_schema() hook
  (name, badge, tag, env_vars). Default derived from display_name.
- tools_config._plugin_image_gen_providers() pulls the schemas from
  every registered non-FAL plugin provider.
- _visible_providers() appends those rows when rendering the Image
  Generation category.
- _configure_provider() handles the new image_gen_plugin_name marker:
  writes image_gen.provider and routes to the plugin's list_models()
  catalog for the model picker.
- _toolset_needs_configuration_prompt('image_gen') stops demanding a
  FAL key when any plugin provider reports is_available().

FAL is skipped in the plugin path because it already has hardcoded
TOOL_CATEGORIES rows — when it gets ported to a plugin in a follow-up
PR the hardcoded rows go away and it surfaces through the same path
as OpenAI.

Verified live: picker shows Nous Subscription / FAL.ai / OpenAI.
Picking OpenAI prompts for OPENAI_API_KEY, then shows the
gpt-image-2-low/medium/high model picker sourced from the plugin.

397 tests pass across plugins/, tools_config, registry, and picker.

* fix(image_gen): close final gaps for plugin-backend parity with FAL

Two small places that still hardcoded FAL:

- hermes_cli/setup.py status line: an OpenAI-only setup showed
  'Image Generation: missing FAL_KEY'. Now probes plugin providers
  and reports '(OpenAI)' when one is_available() — or falls back to
  'missing FAL_KEY or OPENAI_API_KEY' if nothing is configured.

- image_generate tool schema description: said 'using FAL.ai, default
  FLUX 2 Klein 9B'. Rewrote provider-neutral — 'backend and model are
  user-configured' — and notes the 'image' field can be a URL or an
  absolute path, which the gateway delivers either way via
  extract_local_files().
This commit is contained in:
Teknium 2026-04-21 21:30:10 -07:00 committed by GitHub
parent d1acf17773
commit ff9752410a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
13 changed files with 2122 additions and 67 deletions

View file

@ -133,6 +133,9 @@ def _get_enabled_plugins() -> Optional[set]:
# Data classes
# ---------------------------------------------------------------------------
_VALID_PLUGIN_KINDS: Set[str] = {"standalone", "backend", "exclusive"}
@dataclass
class PluginManifest:
"""Parsed representation of a plugin.yaml manifest."""
@ -146,6 +149,23 @@ class PluginManifest:
provides_hooks: List[str] = field(default_factory=list)
source: str = "" # "user", "project", or "entrypoint"
path: Optional[str] = None
# Plugin kind — see plugins.py module docstring for semantics.
# ``standalone`` (default): hooks/tools of its own; opt-in via
# ``plugins.enabled``.
# ``backend``: pluggable backend for an existing core tool (e.g.
# image_gen). Built-in (bundled) backends auto-load;
# user-installed still gated by ``plugins.enabled``.
# ``exclusive``: category with exactly one active provider (memory).
# Selection via ``<category>.provider`` config key; the
# category's own discovery system handles loading and the
# general scanner skips these.
kind: str = "standalone"
# Registry key — path-derived, used by ``plugins.enabled``/``disabled``
# lookups and by ``hermes plugins list``. For a flat plugin at
# ``plugins/disk-cleanup/`` the key is ``disk-cleanup``; for a nested
# category plugin at ``plugins/image_gen/openai/`` the key is
# ``image_gen/openai``. When empty, falls back to ``name``.
key: str = ""
@dataclass
@ -366,6 +386,33 @@ class PluginContext:
self.manifest.name, engine.name,
)
# -- image gen provider registration ------------------------------------
def register_image_gen_provider(self, provider) -> None:
"""Register an image generation backend.
``provider`` must be an instance of
:class:`agent.image_gen_provider.ImageGenProvider`. The
``provider.name`` attribute is what ``image_gen.provider`` in
``config.yaml`` matches against when routing ``image_generate``
tool calls.
"""
from agent.image_gen_provider import ImageGenProvider
from agent.image_gen_registry import register_provider
if not isinstance(provider, ImageGenProvider):
logger.warning(
"Plugin '%s' tried to register an image_gen provider that does "
"not inherit from ImageGenProvider. Ignoring.",
self.manifest.name,
)
return
register_provider(provider)
logger.info(
"Plugin '%s' registered image_gen provider: %s",
self.manifest.name, provider.name,
)
# -- hook registration --------------------------------------------------
def register_hook(self, hook_name: str, callback: Callable) -> None:
@ -465,11 +512,16 @@ class PluginManager:
manifests: List[PluginManifest] = []
# 1. Bundled plugins (<repo>/plugins/<name>/)
# Repo-shipped generic plugins live next to hermes_cli/. Memory and
# context_engine subdirs are handled by their own discovery paths, so
# skip those names here. Bundled plugins are discovered (so they
# show up in `hermes plugins`) but only loaded when added to
# `plugins.enabled` in config.yaml — opt-in like any other plugin.
#
# Repo-shipped plugins live next to hermes_cli/. Two layouts are
# supported (see ``_scan_directory`` for details):
#
# - flat: ``plugins/disk-cleanup/plugin.yaml`` (standalone)
# - category: ``plugins/image_gen/openai/plugin.yaml`` (backend)
#
# ``memory/`` and ``context_engine/`` are skipped at the top level —
# they have their own discovery systems. Porting those to the
# category-namespace ``kind: exclusive`` model is a future PR.
repo_plugins = Path(__file__).resolve().parent.parent / "plugins"
manifests.extend(
self._scan_directory(
@ -492,36 +544,69 @@ class PluginManager:
manifests.extend(self._scan_entry_points())
# Load each manifest (skip user-disabled plugins).
# Later sources override earlier ones on name collision — user plugins
# take precedence over bundled, project plugins take precedence over
# user. Dedup here so we only load the final winner.
# Later sources override earlier ones on key collision — user
# plugins take precedence over bundled, project plugins take
# precedence over user. Dedup here so we only load the final
# winner. Keys are path-derived (``image_gen/openai``,
# ``disk-cleanup``) so ``tts/openai`` and ``image_gen/openai``
# don't collide even when both manifests say ``name: openai``.
disabled = _get_disabled_plugins()
enabled = _get_enabled_plugins() # None = opt-in default (nothing enabled)
winners: Dict[str, PluginManifest] = {}
for manifest in manifests:
winners[manifest.name] = manifest
winners[manifest.key or manifest.name] = manifest
for manifest in winners.values():
# Explicit disable always wins.
if manifest.name in disabled:
lookup_key = manifest.key or manifest.name
# Explicit disable always wins (matches on key or on legacy
# bare name for back-compat with existing user configs).
if lookup_key in disabled or manifest.name in disabled:
loaded = LoadedPlugin(manifest=manifest, enabled=False)
loaded.error = "disabled via config"
self._plugins[manifest.name] = loaded
logger.debug("Skipping disabled plugin '%s'", manifest.name)
self._plugins[lookup_key] = loaded
logger.debug("Skipping disabled plugin '%s'", lookup_key)
continue
# Opt-in gate: plugins must be in the enabled allow-list.
# If the allow-list is missing (None), treat as "nothing enabled"
# — users have to explicitly enable plugins to load them.
# Memory and context_engine providers are excluded from this gate
# since they have their own single-select config (memory.provider
# / context.engine), not the enabled list.
if enabled is None or manifest.name not in enabled:
# Exclusive plugins (memory providers) have their own
# discovery/activation path. The general loader records the
# manifest for introspection but does not load the module.
if manifest.kind == "exclusive":
loaded = LoadedPlugin(manifest=manifest, enabled=False)
loaded.error = "not enabled in config (run `hermes plugins enable {}` to activate)".format(
manifest.name
loaded.error = (
"exclusive plugin — activate via <category>.provider config"
)
self._plugins[manifest.name] = loaded
self._plugins[lookup_key] = loaded
logger.debug(
"Skipping '%s' (not in plugins.enabled)", manifest.name
"Skipping '%s' (exclusive, handled by category discovery)",
lookup_key,
)
continue
# Built-in backends auto-load — they ship with hermes and must
# just work. Selection among them (e.g. which image_gen backend
# services calls) is driven by ``<category>.provider`` config,
# enforced by the tool wrapper.
if manifest.kind == "backend" and manifest.source == "bundled":
self._load_plugin(manifest)
continue
# Everything else (standalone, user-installed backends,
# entry-point plugins) is opt-in via plugins.enabled.
# Accept both the path-derived key and the legacy bare name
# so existing configs keep working.
is_enabled = (
enabled is not None
and (lookup_key in enabled or manifest.name in enabled)
)
if not is_enabled:
loaded = LoadedPlugin(manifest=manifest, enabled=False)
loaded.error = (
"not enabled in config (run `hermes plugins enable {}` to activate)"
.format(lookup_key)
)
self._plugins[lookup_key] = loaded
logger.debug(
"Skipping '%s' (not in plugins.enabled)", lookup_key
)
continue
self._load_plugin(manifest)
@ -545,9 +630,37 @@ class PluginManager:
) -> List[PluginManifest]:
"""Read ``plugin.yaml`` manifests from subdirectories of *path*.
*skip_names* is an optional allow-list of names to ignore (used
for the bundled scan to exclude ``memory`` / ``context_engine``
subdirs that have their own discovery path).
Supports two layouts, mixed freely:
* **Flat** ``<root>/<plugin-name>/plugin.yaml``. Key is
``<plugin-name>`` (e.g. ``disk-cleanup``).
* **Category** ``<root>/<category>/<plugin-name>/plugin.yaml``,
where the ``<category>`` directory itself has no ``plugin.yaml``.
Key is ``<category>/<plugin-name>`` (e.g. ``image_gen/openai``).
Depth is capped at two segments.
*skip_names* is an optional allow-list of names to ignore at the
top level (kept for back-compat; the current call sites no longer
pass it now that categories are first-class).
"""
return self._scan_directory_level(
path, source, skip_names=skip_names, prefix="", depth=0
)
def _scan_directory_level(
self,
path: Path,
source: str,
*,
skip_names: Optional[Set[str]],
prefix: str,
depth: int,
) -> List[PluginManifest]:
"""Recursive implementation of :meth:`_scan_directory`.
``prefix`` is the category path already accumulated ("" at root,
"image_gen" one level in). ``depth`` is the recursion depth; we
cap at 2 so ``<root>/a/b/c/`` is ignored.
"""
manifests: List[PluginManifest] = []
if not path.is_dir():
@ -556,37 +669,88 @@ class PluginManager:
for child in sorted(path.iterdir()):
if not child.is_dir():
continue
if skip_names and child.name in skip_names:
if depth == 0 and skip_names and child.name in skip_names:
continue
manifest_file = child / "plugin.yaml"
if not manifest_file.exists():
manifest_file = child / "plugin.yml"
if not manifest_file.exists():
logger.debug("Skipping %s (no plugin.yaml)", child)
if manifest_file.exists():
manifest = self._parse_manifest(
manifest_file, child, source, prefix
)
if manifest is not None:
manifests.append(manifest)
continue
try:
if yaml is None:
logger.warning("PyYAML not installed cannot load %s", manifest_file)
continue
data = yaml.safe_load(manifest_file.read_text()) or {}
manifest = PluginManifest(
name=data.get("name", child.name),
version=str(data.get("version", "")),
description=data.get("description", ""),
author=data.get("author", ""),
requires_env=data.get("requires_env", []),
provides_tools=data.get("provides_tools", []),
provides_hooks=data.get("provides_hooks", []),
source=source,
path=str(child),
# No manifest at this level. If we're still within the depth
# cap, treat this directory as a category namespace and recurse
# one level in looking for children with manifests.
if depth >= 1:
logger.debug("Skipping %s (no plugin.yaml, depth cap reached)", child)
continue
sub_prefix = f"{prefix}/{child.name}" if prefix else child.name
manifests.extend(
self._scan_directory_level(
child,
source,
skip_names=None,
prefix=sub_prefix,
depth=depth + 1,
)
manifests.append(manifest)
except Exception as exc:
logger.warning("Failed to parse %s: %s", manifest_file, exc)
)
return manifests
def _parse_manifest(
self,
manifest_file: Path,
plugin_dir: Path,
source: str,
prefix: str,
) -> Optional[PluginManifest]:
"""Parse a single ``plugin.yaml`` into a :class:`PluginManifest`.
Returns ``None`` on parse failure (logs a warning).
"""
try:
if yaml is None:
logger.warning("PyYAML not installed cannot load %s", manifest_file)
return None
data = yaml.safe_load(manifest_file.read_text()) or {}
name = data.get("name", plugin_dir.name)
key = f"{prefix}/{plugin_dir.name}" if prefix else name
raw_kind = data.get("kind", "standalone")
if not isinstance(raw_kind, str):
raw_kind = "standalone"
kind = raw_kind.strip().lower()
if kind not in _VALID_PLUGIN_KINDS:
logger.warning(
"Plugin %s: unknown kind '%s' (valid: %s); treating as 'standalone'",
key, raw_kind, ", ".join(sorted(_VALID_PLUGIN_KINDS)),
)
kind = "standalone"
return PluginManifest(
name=name,
version=str(data.get("version", "")),
description=data.get("description", ""),
author=data.get("author", ""),
requires_env=data.get("requires_env", []),
provides_tools=data.get("provides_tools", []),
provides_hooks=data.get("provides_hooks", []),
source=source,
path=str(plugin_dir),
kind=kind,
key=key,
)
except Exception as exc:
logger.warning("Failed to parse %s: %s", manifest_file, exc)
return None
# -----------------------------------------------------------------------
# Entry-point scanning
# -----------------------------------------------------------------------
@ -609,6 +773,7 @@ class PluginManager:
name=ep.name,
source="entrypoint",
path=ep.value,
key=ep.name,
)
manifests.append(manifest)
except Exception as exc:
@ -670,10 +835,16 @@ class PluginManager:
loaded.error = str(exc)
logger.warning("Failed to load plugin '%s': %s", manifest.name, exc)
self._plugins[manifest.name] = loaded
self._plugins[manifest.key or manifest.name] = loaded
def _load_directory_module(self, manifest: PluginManifest) -> types.ModuleType:
"""Import a directory-based plugin as ``hermes_plugins.<name>``."""
"""Import a directory-based plugin as ``hermes_plugins.<slug>``.
The module slug is derived from ``manifest.key`` so category-namespaced
plugins (``image_gen/openai``) import as
``hermes_plugins.image_gen__openai`` without colliding with any
future ``tts/openai``.
"""
plugin_dir = Path(manifest.path) # type: ignore[arg-type]
init_file = plugin_dir / "__init__.py"
if not init_file.exists():
@ -686,7 +857,9 @@ class PluginManager:
ns_pkg.__package__ = _NS_PARENT
sys.modules[_NS_PARENT] = ns_pkg
module_name = f"{_NS_PARENT}.{manifest.name.replace('-', '_')}"
key = manifest.key or manifest.name
slug = key.replace("/", "__").replace("-", "_")
module_name = f"{_NS_PARENT}.{slug}"
spec = importlib.util.spec_from_file_location(
module_name,
init_file,
@ -767,10 +940,12 @@ class PluginManager:
def list_plugins(self) -> List[Dict[str, Any]]:
"""Return a list of info dicts for all discovered plugins."""
result: List[Dict[str, Any]] = []
for name, loaded in sorted(self._plugins.items()):
for key, loaded in sorted(self._plugins.items()):
result.append(
{
"name": name,
"name": loaded.manifest.name,
"key": loaded.manifest.key or loaded.manifest.name,
"kind": loaded.manifest.kind,
"version": loaded.manifest.version,
"description": loaded.manifest.description,
"source": loaded.manifest.source,