ci: sync dashboard from 14f3545

Source: github.com/CouLiBaLy-B/gh-deepagents@14f3545190f0c5123a074d14a34077067ec20760

src/gh_deepagent/dashboard/pages/8_⚙️_LLM_Settings.py

@@ -87,6 +87,25 @@ PROVIDERS = {
         "spec_prefix": "groq:",
         "doc": "https://console.groq.com/keys",
     },
+    "openrouter": {
+        "label": "OpenRouter (unified — Anthropic, OpenAI, Llama, Mistral, …)",
+        "env_key": "OPENROUTER_API_KEY",
+        "models": [
+            "anthropic/claude-sonnet-4-5",
+            "anthropic/claude-haiku-4",
+            "openai/gpt-4o-mini",
+            "openai/gpt-4o",
+            "openai/gpt-4.1-mini",
+            "google/gemini-2.5-flash",
+            "google/gemini-2.5-pro",
+            "meta-llama/llama-3.3-70b-instruct",
+            "mistralai/mistral-large-latest",
+            "qwen/qwen-2.5-coder-32b-instruct",
+            "deepseek/deepseek-chat",
+        ],
+        "spec_prefix": "openrouter:",
+        "doc": "https://openrouter.ai/keys",
+    },
     "ollama": {
         "label": "Ollama (local, no API key)",
         "env_key": "",  # no key

src/gh_deepagent/models.py

@@ -2,9 +2,19 @@
 
 Every model returned is wired with the ``CostCallback`` so token/cost metrics
 flow through the observability stack automatically.
+
+Supported model specs:
+    ollama:<model>              → ChatOllama, OLLAMA_BASE_URL
+    openrouter:<model>          → ChatOpenAI with base_url=openrouter.ai
+    anthropic:<model>           → langchain init_chat_model
+    openai:<model>              → langchain init_chat_model
+    google_genai:<model>        → langchain init_chat_model
+    groq:<model>                → langchain init_chat_model
+    <anything else with ":"> → forwarded to init_chat_model verbatim
 """
 from __future__ import annotations
 
+import os
 from typing import Any
 
 from langchain.chat_models import init_chat_model
@@ -12,6 +22,9 @@ from langchain.chat_models import init_chat_model
 from .config import get_settings
 
 
+OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1"
+
+
 def _attach_callbacks(model):
     """Attach the cost callback to a chat model. Idempotent."""
     try:
@@ -30,6 +43,38 @@ def _attach_callbacks(model):
         return model.with_config({"callbacks": existing})
 
 
+def _build_openrouter(model_name: str, **overrides: Any):
+    """OpenRouter speaks the OpenAI HTTP API. Use ChatOpenAI with a custom base_url.
+
+    Recommended HTTP headers (`HTTP-Referer` + `X-Title`) help OpenRouter rank
+    your traffic; we set them from env vars when present.
+    """
+    from langchain_openai import ChatOpenAI
+
+    api_key = os.getenv("OPENROUTER_API_KEY", "")
+    if not api_key:
+        raise RuntimeError(
+            "OPENROUTER_API_KEY is not set. Configure it via the dashboard's "
+            "⚙️ LLM Settings page or in the Space's Settings → Secrets."
+        )
+    extra_headers = {}
+    if site := os.getenv("OPENROUTER_HTTP_REFERER"):
+        extra_headers["HTTP-Referer"] = site
+    if title := os.getenv("OPENROUTER_X_TITLE", "gh-deepagent"):
+        extra_headers["X-Title"] = title
+
+    params: dict[str, Any] = dict(
+        model=model_name,
+        base_url=OPENROUTER_BASE_URL,
+        api_key=api_key,
+        temperature=0.0,
+    )
+    if extra_headers:
+        params["default_headers"] = extra_headers
+    params.update(overrides)
+    return ChatOpenAI(**params)
+
+
 def build_model(model_spec: str | None = None, **overrides: Any):
     """Return a LangChain chat model with observability callbacks attached."""
     settings = get_settings()
@@ -49,4 +94,9 @@ def build_model(model_spec: str | None = None, **overrides: Any):
         params.update(overrides)
         return _attach_callbacks(ChatOllama(**params))
 
+    if spec.startswith("openrouter:"):
+        _, model_name = spec.split("openrouter:", 1)
+        return _attach_callbacks(_build_openrouter(model_name, **overrides))
+
+    # Everything else → LangChain's init_chat_model (handles anthropic, openai, …).
     return _attach_callbacks(init_chat_model(spec, temperature=0.0, **overrides))

src/gh_deepagent/observability/cost.py

@@ -47,6 +47,19 @@ PRICE_CATALOG: dict[str, dict[str, float]] = {
     # Google
     "google_genai:gemini-2.5-pro": {"input": 1.25,  "output": 10.00},
     "google_genai:gemini-2.5-flash": {"input": 0.075, "output": 0.30},
+    # OpenRouter — they expose hundreds of models; we ship the most common ones
+    # at upstream prices. Anything else falls through to the openrouter:* wildcard
+    # at $0 (tokens still counted; check openrouter.ai/usage for the real bill).
+    "openrouter:anthropic/claude-sonnet-4-5":  {"input": 3.00,  "output": 15.00},
+    "openrouter:anthropic/claude-haiku-4":     {"input": 0.80,  "output": 4.00},
+    "openrouter:openai/gpt-4o":                {"input": 2.50,  "output": 10.00},
+    "openrouter:openai/gpt-4o-mini":           {"input": 0.15,  "output": 0.60},
+    "openrouter:openai/gpt-4.1-mini":          {"input": 0.40,  "output": 1.60},
+    "openrouter:google/gemini-2.5-flash":      {"input": 0.075, "output": 0.30},
+    "openrouter:google/gemini-2.5-pro":        {"input": 1.25,  "output": 10.00},
+    "openrouter:meta-llama/llama-3.3-70b-instruct": {"input": 0.59, "output": 0.79},
+    "openrouter:mistralai/mistral-large-latest":    {"input": 2.00, "output": 6.00},
+    "openrouter:*": {"input": 0.0, "output": 0.0},
     # Local — counted, not billed
     "ollama:*":   {"input": 0.0, "output": 0.0},
     "vllm:*":     {"input": 0.0, "output": 0.0},