feat(providers): add aimlapi.com as a gateway

core/config.py

@@ -194,6 +194,7 @@ class ProvidersConfig(_Base):
     and adding the matching :class:`~core.providers.registry.ProviderSpec`.
     """
 
+    aimlapi: ProviderConfig = Field(default_factory=ProviderConfig)
     custom: ProviderConfig = Field(default_factory=ProviderConfig)
     openrouter: ProviderConfig = Field(default_factory=ProviderConfig)
     forge: ProviderConfig = Field(default_factory=ProviderConfig)

core/providers/catalog_service.py

@@ -409,6 +409,12 @@ def _catalog_request(connection: ResolvedConnection) -> tuple[str, dict[str, str
         base = connection.api_base.rstrip("/")
     if connection.api_key:
         headers["Authorization"] = f"Bearer {connection.api_key}"
+    if connection.provider_name == "aimlapi":
+        # AI/ML API serves one directory for every endpoint family it hosts:
+        # 936 rows, of which only 353 are chat models — the rest are image,
+        # video, speech and embedding endpoints that a chat request rejects,
+        # and ids repeat across families. Ask for the chat surface only.
+        return f"{base}/models?type=openai%2Fchat-completions", headers
     return f"{base}/models", headers
 
 
@@ -421,15 +427,21 @@ def _parse_model(value: dict[str, Any]) -> CatalogModel | None:
     top_provider = (
         value.get("top_provider") if isinstance(value.get("top_provider"), dict) else {}
     )
+    # Some gateways (AI/ML API) nest the descriptive fields one level down
+    # instead of publishing OpenRouter's flat keys. Read it after the flat
+    # names so a provider that has both keeps winning at the top level.
+    info = value.get("info") if isinstance(value.get("info"), dict) else {}
     context = (
         value.get("context_length")
         or value.get("context_window")
         or value.get("max_input_tokens")
+        or info.get("contextLength")
         or fallback.context_window
     )
     output = (
         value.get("max_output_tokens")
         or top_provider.get("max_completion_tokens")
+        or info.get("outputMax")
         or fallback.max_output_tokens
     )
     supported = value.get("supported_parameters", [])
@@ -446,7 +458,12 @@ def _parse_model(value: dict[str, Any]) -> CatalogModel | None:
         )
     return CatalogModel(
         id=model_id,
-        name=str(value.get("name") or value.get("display_name") or model_id),
+        name=str(
+            value.get("name")
+            or value.get("display_name")
+            or info.get("name")
+            or model_id
+        ),
         context_window=max(1, int(context)),
         max_output_tokens=max(1, int(output)),
         supported_parameters=supported_parameters,

core/providers/openai_compat.py

@@ -12,6 +12,7 @@
 import uuid
 from collections.abc import Awaitable, Callable
 from typing import TYPE_CHECKING, Any
+from urllib.parse import urlparse
 
 import json_repair
 from loguru import logger
@@ -79,6 +80,17 @@
     "HTTP-Referer": "https://github.com/HKUDS/DeepCode",
     "X-Title": "DeepCode",
 }
+_DEFAULT_AIMLAPI_HEADERS = {
+    "HTTP-Referer": "https://github.com/HKUDS/DeepCode",
+    "X-Title": "DeepCode",
+    "X-AIMLAPI-Source": "agent/deepcode",
+    "X-AIMLAPI-Partner-ID": "part_CxcOejScJ3hI0O2cExWchiBy",
+}
+# Hosts that are actually AI/ML API. Attribution is keyed on the resolved
+# request origin rather than on the selected template, so a user who repoints
+# the ``aimlapi`` template at a proxy of their own does not hand a third party
+# DeepCode's partner identity.
+_AIMLAPI_HOST_SUFFIX = ".aimlapi.com"
 # Per-model thinking / reasoning quirks now live declaratively in
 # ``core.providers.model_compat`` (resolved via ``resolve_model_compat``);
 # this module only assembles requests from the resolved value.
@@ -171,6 +183,22 @@ def _uses_requesty_attribution(
     return bool(api_base and "requesty" in api_base.lower())
 
 
+def _uses_aimlapi_attribution(
+    spec: "ProviderSpec | None", api_base: str | None
+) -> bool:
+    """Apply DeepCode attribution headers to AI/ML API requests by default.
+
+    Stricter than the two helpers above: a substring test would also fire for
+    a gateway that merely *fronts* AI/ML API, so the host of the resolved base
+    URL has to be ours. The template name only decides the case where no base
+    URL was resolved at all.
+    """
+    if not api_base:
+        return bool(spec and spec.name == "aimlapi")
+    host = (urlparse(api_base.strip()).hostname or "").lower()
+    return host == "aimlapi.com" or host.endswith(_AIMLAPI_HOST_SUFFIX)
+
+
 _RESPONSES_FAILURE_THRESHOLD = 3
 _RESPONSES_PROBE_INTERVAL_S = 300  # 5 minutes
 
@@ -250,6 +278,8 @@ def __init__(
             default_headers.update(_DEFAULT_OPENROUTER_HEADERS)
         if _uses_requesty_attribution(spec, effective_base):
             default_headers.update(_DEFAULT_REQUESTY_HEADERS)
+        if _uses_aimlapi_attribution(spec, effective_base):
+            default_headers.update(_DEFAULT_AIMLAPI_HEADERS)
         if extra_headers:
             default_headers.update(extra_headers)
 

core/providers/registry.py

@@ -54,6 +54,21 @@ def label(self) -> str:
 
 
 PROVIDERS: tuple[ProviderSpec, ...] = (
+    ProviderSpec(
+        name="aimlapi",
+        keywords=("aimlapi",),
+        env_key="AIMLAPI_API_KEY",
+        # The vendor writes its own name lowercase, with the TLD.
+        display_name="aimlapi.com",
+        backend="openai_compat",
+        is_gateway=True,
+        detect_by_base_keyword="aimlapi",
+        default_api_base="https://api.aimlapi.com/v1",
+        # ``vendor/model`` ids like the OpenRouter-style gateways below, not
+        # Forge's bare ids. ``cache_control`` markers are honoured and the
+        # gateway reports ``cached_tokens`` back in usage.
+        supports_prompt_caching=True,
+    ),
     ProviderSpec(
         name="custom",
         keywords=(),