feat(providers): Support DeepInfra as an LLM provider

.env.example

@@ -108,6 +108,10 @@
 # HF_BASE_URL=https://router.huggingface.co/v1  # Override default base URL
 # OPENCODE_GO_BASE_URL=https://opencode.ai/zen/go/v1  # Override default base URL
 
+# DeepInfra — 100+ top open models, pay-per-use.
+# Get your key at: https://deepinfra.com/dash/api_keys
+# DEEPINFRA_API_KEY=
+
 # =============================================================================
 # LLM PROVIDER (Qwen OAuth)
 # =============================================================================

agent/auxiliary_client.py

@@ -426,6 +426,10 @@ def _get_aux_model_for_provider(provider_id: str) -> str:
     "kilocode": "google/gemini-3-flash-preview",
     "ollama-cloud": "nemotron-3-nano:30b",
     "tencent-tokenhub": "hy3-preview",
+    # NB: no "deepinfra" entry — its aux model lives on the ProviderProfile
+    # (plugins/model-providers/deepinfra: default_aux_model), which
+    # _get_aux_model_for_provider() reads first. Duplicating it here would be
+    # dead data that drifts when the profile's value is bumped.
 }
 
 # Legacy alias — callers that haven't been updated to _get_aux_model_for_provider()
@@ -441,6 +445,33 @@ def _get_aux_model_for_provider(provider_id: str) -> str:
     "zai": "glm-5v-turbo",
 }
 
+
+def _resolve_provider_vision_default(provider: str) -> Optional[str]:
+    """Return the provider's preferred default vision model id, or None.
+
+    Static entries in :data:`_PROVIDER_VISION_MODELS` win first (xiaomi /
+    zai have dedicated vision-only model names that don't live in any
+    discoverable catalog). Otherwise the provider's :class:`ProviderProfile`
+    gets a chance to supply one via its ``default_vision_model()`` hook —
+    that's where catalog-backed providers (DeepInfra) resolve a live default,
+    keeping the discovery logic inside their plugin instead of a name-check
+    branch here.
+    """
+    static = _PROVIDER_VISION_MODELS.get(provider)
+    if static:
+        return static
+    try:
+        from providers import get_provider_profile
+        profile = get_provider_profile(provider)
+    except Exception:
+        return None
+    if profile is None:
+        return None
+    try:
+        return profile.default_vision_model()
+    except Exception:
+        return None
+
 # Providers whose endpoint does not accept image input, even though the
 # provider's broader ecosystem has vision models available elsewhere.  When
 # `auxiliary.vision.provider: auto` sees one of these as the main provider,
@@ -4842,6 +4873,7 @@ def get_async_text_auxiliary_client(task: str = "", *, main_runtime: Optional[Di
 _VISION_AUTO_PROVIDER_ORDER = (
     "openrouter",
     "nous",
+    "deepinfra",
 )
 
 
@@ -4898,6 +4930,21 @@ def _resolve_strict_vision_backend(
         return resolve_provider_client("openai-codex", model, is_vision=True)
     if provider == "anthropic":
         return _try_anthropic()
+    if provider == "deepinfra":
+        # DeepInfra exposes vision-capable models (Llama-4 Scout/Maverick,
+        # Qwen3-VL, Gemma 3, Gemini) on the same OpenAI-compatible endpoint
+        # as its chat models. The default is discovered live via the profile's
+        # default_vision_model() hook (key-gated, chat-surface + vision tag) so
+        # we don't pin a hardcoded id that may rot when DeepInfra retires a
+        # model, and this module stays provider-agnostic.
+        vision_model = model or _resolve_provider_vision_default("deepinfra")
+        if not vision_model:
+            logger.debug(
+                "Vision auto-detect: deepinfra catalog unreachable or "
+                "returned no vision-tagged models — skipping"
+            )
+            return None, None
+        return resolve_provider_client("deepinfra", vision_model, is_vision=True)
     if provider == "custom":
         return _try_custom_endpoint()
     return None, None
@@ -4983,16 +5030,29 @@ def _finalize(resolved_provider: str, sync_client: Any, default_model: Optional[
         #      _PROVIDER_VISION_MODELS provides per-provider vision model
         #      overrides when the provider has a dedicated multimodal model
         #      that differs from the chat model (e.g. xiaomi → mimo-v2-omni,
-        #      zai → glm-5v-turbo). Nous is the exception: it has a dedicated
-        #      strict vision backend with tier-aware defaults, so it must not
-        #      fall through to the user's text chat model here.
-        #   2. OpenRouter  (vision-capable aggregator fallback)
+        #      zai → glm-5v-turbo). DeepInfra is similar but resolves its
+        #      default vision model live from the catalog (see
+        #      :func:`_resolve_provider_vision_default`). Nous is the
+        #      exception: it has a dedicated strict vision backend with
+        #      tier-aware defaults, so it must not fall through to the
+        #      user's text chat model here.
+        #   2. OpenRouter (vision-capable aggregator fallback)
         #   3. Nous Portal (vision-capable aggregator fallback)
-        #   4. Stop
+        #   4. DeepInfra   (OpenAI-compatible; vision model discovered
+        #                   live from the catalog — tried when
+        #                   DEEPINFRA_API_KEY is set)
+        #   5. Stop
         main_provider = _read_main_provider()
         main_model = _read_main_model()
         if main_provider and main_provider not in {"auto", ""}:
-            vision_model = _PROVIDER_VISION_MODELS.get(main_provider, main_model)
+            # A provider-specific vision default wins over the user's chat model:
+            # static overrides (xiaomi/zai) and catalog-backed discovery (the
+            # DeepInfra profile hook) both yield a *known* vision-capable model,
+            # whereas the pinned chat model is usually NOT multimodal (e.g. the
+            # DeepSeek-V4-Flash default) and _main_model_supports_vision can't be
+            # trusted to catch that. Only fall back to the chat model when no
+            # provider default is available (catalog unreachable).
+            vision_model = _resolve_provider_vision_default(main_provider) or main_model
             if main_provider == "nous":
                 sync_client, default_model = _resolve_strict_vision_backend(
                     main_provider, vision_model

agent/model_metadata.py

@@ -47,7 +47,7 @@ def _resolve_requests_verify() -> bool | str:
 # are preserved so the full model name reaches cache lookups and server queries.
 _PROVIDER_PREFIXES: frozenset[str] = frozenset({
     "openrouter", "nous", "openai-codex", "copilot", "copilot-acp",
-    "gemini", "ollama-cloud", "zai", "kimi-coding", "kimi-coding-cn", "stepfun", "minimax", "minimax-oauth", "minimax-cn", "anthropic", "deepseek",
+    "gemini", "ollama-cloud", "zai", "kimi-coding", "kimi-coding-cn", "stepfun", "minimax", "minimax-oauth", "minimax-cn", "anthropic", "deepseek", "deepinfra",
     "opencode-zen", "opencode-go", "kilocode", "alibaba", "novita",
     "qwen-oauth",
     "xiaomi",
@@ -58,7 +58,7 @@ def _resolve_requests_verify() -> bool | str:
     # Common aliases
     "google", "google-gemini", "google-ai-studio",
     "glm", "z-ai", "z.ai", "zhipu", "github", "github-copilot",
-    "github-models", "kimi", "moonshot", "kimi-cn", "moonshot-cn", "claude", "deep-seek",
+    "github-models", "kimi", "moonshot", "kimi-cn", "moonshot-cn", "claude", "deep-seek", "deep-infra",
     "ollama",
     "stepfun", "opencode", "zen", "go", "kilo", "dashscope", "aliyun", "qwen",
     "mimo", "xiaomi-mimo",
@@ -437,6 +437,7 @@ def _is_custom_endpoint(base_url: str) -> bool:
     "generativelanguage.googleapis.com": "gemini",
     "inference-api.nousresearch.com": "nous",
     "api.deepseek.com": "deepseek",
+    "api.deepinfra.com": "deepinfra",
     "api.githubcopilot.com": "copilot",
     # Enterprise Copilot endpoints look like api.enterprise.githubcopilot.com,
     # api.business.githubcopilot.com, etc.  Match the suffix so context-window
@@ -740,6 +741,24 @@ def _extract_pricing(payload: Dict[str, Any]) -> Dict[str, Any]:
             pricing["completion"] = str(float(novita_output) / 10_000 / 1_000_000)
         return pricing
 
+    # DeepInfra ships pricing under ``metadata.pricing`` with $/MTok values:
+    # ``input_tokens``, ``output_tokens``, ``cache_read_tokens``. Convert to
+    # per-token strings so the generic cost machinery (usage_pricing.py)
+    # consumes them through the same path as OpenRouter / OpenAI.
+    metadata = payload.get("metadata") if isinstance(payload.get("metadata"), dict) else None
+    deepinfra_pricing = metadata.get("pricing") if metadata else None
+    if isinstance(deepinfra_pricing, dict) and any(
+        k in deepinfra_pricing for k in ("input_tokens", "output_tokens", "cache_read_tokens")
+    ):
+        result: Dict[str, Any] = {}
+        if deepinfra_pricing.get("input_tokens") is not None:
+            result["prompt"] = str(float(deepinfra_pricing["input_tokens"]) / 1_000_000)
+        if deepinfra_pricing.get("output_tokens") is not None:
+            result["completion"] = str(float(deepinfra_pricing["output_tokens"]) / 1_000_000)
+        if deepinfra_pricing.get("cache_read_tokens") is not None:
+            result["cache_read"] = str(float(deepinfra_pricing["cache_read_tokens"]) / 1_000_000)
+        return result
+
     alias_map = {
         "prompt": ("prompt", "input", "input_cost_per_token", "prompt_token_cost"),
         "completion": ("completion", "output", "output_cost_per_token", "completion_token_cost"),

agent/tts_registry.py

@@ -56,6 +56,7 @@
     "neutts",
     "kittentts",
     "piper",
+    "deepinfra",
 })