Add GitHub Copilot CLI as a chat model provider

.apm/skills/benchmark-qed-setup/SKILL.md

@@ -144,7 +144,7 @@ Ask in **a single `ask_user` form** (split into two if the workflow is autoq, si
 
 | Field | Type | Options / examples | Notes |
 |-------|------|--------------------|-------|
-| `llm_provider` | enum | `openai.chat`, `azure.openai.chat`, `azure.inference.chat` | See provider table in [references/config-reference.md](references/config-reference.md). |
+| `llm_provider` | enum | `openai.chat`, `azure.openai.chat`, `azure.inference.chat`, `copilot.chat` | See provider table in [references/config-reference.md](references/config-reference.md). |
 | `model` | string | `gpt-4.1`, `gpt-4o`, `o3-mini`, an Azure deployment name | For Azure providers this is the **deployment name**, not the base model id. |
 | `auth_type` | enum | `api_key` (default), `azure_managed_identity` | `azure_managed_identity` is only valid for `azure.*` providers. |
 | `api_key_env_var` | string | `OPENAI_API_KEY` (default), `AZURE_OPENAI_API_KEY`, … | Only ask when `auth_type=api_key`. The skill writes `${VAR_NAME}` into YAML and adds the variable to `.env`. |

.apm/skills/benchmark-qed-setup/references/config-reference.md

@@ -481,6 +481,7 @@ significance_correction: holm
 | Azure OpenAI Embedding | `azure.openai.embedding` | Azure-hosted embeddings |
 | Azure Inference Chat | `azure.inference.chat` | Azure AI Inference chat |
 | Azure Inference Embedding | `azure.inference.embedding` | Azure AI Inference embeddings |
+| GitHub Copilot CLI | `copilot.chat` | Chat models via the Copilot CLI (`copilot login`; no `api_key`/`auth_type`; no embeddings, so pair with another embedding provider; `temperature`/`seed` ignored). Optional `init_args`: `cli_path`, `github_token`, `timeout`, `reasoning_effort`, `validate_model`, `delete_sessions`. |
 
 ## Custom LLM Providers
 

.semversioner/next-release/minor-20261023000000000000.json

@@ -0,0 +1,4 @@
+{
+  "type": "minor",
+  "description": "Add copilot.chat LLM provider to use models through the GitHub Copilot CLI"
+}

benchmark_qed/cli/init_config.py

@@ -61,7 +61,7 @@ class ConfigType(StrEnum):
   model: gpt-4.1
   auth_type: api_key # or azure_managed_identity
   api_key: ${OPENAI_API_KEY} # remove this if using azure_managed_identity
-  llm_provider: openai.chat # or azure.openai.chat | azure.inference.chat
+  llm_provider: openai.chat # or azure.openai.chat | azure.inference.chat | copilot.chat
   concurrent_requests: 4 # The number of concurrent requests to send to the model.
   # init_args:
   #   Additional initialization arguments for the LLM can be added here.

benchmark_qed/cli/interactive.py

@@ -88,6 +88,7 @@ class ProviderResult:
     ("openai.chat", "OpenAI", "OpenAI API (default)"),
     ("azure.openai.chat", "Azure OpenAI", "Azure-hosted OpenAI models"),
     ("azure.inference.chat", "Azure Inference", "Azure AI Inference endpoint"),
+    ("copilot.chat", "GitHub Copilot CLI", "Models available via GitHub Copilot"),
 ]
 
 EMBEDDING_PROVIDERS: list[tuple[str, str, str]] = [
@@ -106,6 +107,7 @@ class ProviderResult:
 ]
 
 DEFAULT_CHAT_MODEL = "gpt-4.1"
+DEFAULT_COPILOT_CHAT_MODEL = "gpt-6-luna"
 DEFAULT_EMBEDDING_MODEL = "text-embedding-3-large"
 DEFAULT_API_VERSION = "2024-12-01-preview"
 
@@ -255,6 +257,18 @@ def prompt_provider(
         [(val, label) for val, label, _desc in providers],
     )
 
+    if provider_value == "copilot.chat":
+        # The Copilot CLI handles authentication via the user's GitHub login.
+        model = typer.prompt(
+            "Model name",
+            default=default_model or DEFAULT_COPILOT_CHAT_MODEL,
+        )
+        return ProviderResult(
+            llm_provider=provider_value,
+            model=model,
+            auth_type="copilot_cli",
+        )
+
     # Auth type
     auth_type = select_option("Authentication type", AUTH_TYPES)
 
@@ -276,6 +290,12 @@ def prompt_embedding_provider(
     chat_result: ProviderResult,
 ) -> ProviderResult:
     """Ask whether to reuse the chat provider for embeddings, or configure separately."""
+    if chat_result.llm_provider == "copilot.chat":
+        rich_print(
+            "[yellow]GitHub Copilot CLI does not provide embeddings; "
+            "select a separate embedding provider.[/yellow]"
+        )
+        return prompt_provider("embedding")
     if typer.confirm("Use the same provider for embeddings?", default=True):
         # Derive the embedding provider from the chat provider
         mapping = {

benchmark_qed/cli/yaml_renderer.py

@@ -34,12 +34,17 @@ def _render_llm_section(provider_dict: dict[str, Any], indent: int = 2) -> str:
         provider_dict = dataclasses.asdict(provider_dict)
 
     pad = " " * indent
-    lines: list[str] = []
+    lines: list[str] = [f"{pad}model: {provider_dict['model']}"]
 
-    lines.extend([
-        f"{pad}model: {provider_dict['model']}",
-        f"{pad}auth_type: {provider_dict['auth_type']}",
-    ])
+    if provider_dict["llm_provider"] == "copilot.chat":
+        # Copilot CLI authenticates via the user's GitHub login; no API key.
+        lines.extend([
+            f"{pad}llm_provider: {provider_dict['llm_provider']}",
+            f"{pad}concurrent_requests: 4",
+        ])
+        return "\n".join(lines)
+
+    lines.append(f"{pad}auth_type: {provider_dict['auth_type']}")
 
     if provider_dict["auth_type"] == "api_key":
         lines.append(f"{pad}api_key: ${{OPENAI_API_KEY}}")

benchmark_qed/config/llm_config.py

@@ -22,6 +22,11 @@ class LLMProvider(StrEnum):
     AzureOpenAIEmbedding = "azure.openai.embedding"
     AzureInferenceChat = "azure.inference.chat"
     AzureInferenceEmbedding = "azure.inference.embedding"
+    CopilotChat = "copilot.chat"
+
+
+# graphrag-llm completion strategy name used for the GitHub Copilot CLI provider.
+COPILOT_COMPLETION_TYPE = "copilot_cli"
 
 
 class ModelType(StrEnum):
@@ -121,9 +126,17 @@ class LLMConfig(BaseModel):
         description="List of custom LLM providers to register.",
     )
 
+    @property
+    def is_copilot(self) -> bool:
+        """Whether this config targets the GitHub Copilot CLI provider."""
+        return str(self.llm_provider) == LLMProvider.CopilotChat
+
     @model_validator(mode="after")
     def check_api_key(self) -> Self:
         """Check if the API key is set."""
+        # The Copilot CLI authenticates with the user's GitHub login or token.
+        if self.is_copilot:
+            return self
         if self.auth_type == "api_key" and (
             self.api_key is None or self.api_key.get_secret_value().strip() == ""
         ):
@@ -138,8 +151,14 @@ def to_model_config(self) -> ModelConfig:
         ``azure``, ``azure_ai``). Custom providers are passed through with
         their registered name as the ``ModelConfig.type`` so that
         ``register_completion`` / ``register_embedding`` callers can resolve
-        them.
+        them. The GitHub Copilot CLI provider is mapped to the
+        ``copilot_cli`` completion type registered by
+        :class:`benchmark_qed.llm.factory.ModelFactory`; its ``init_args`` are
+        forwarded to :class:`benchmark_qed.llm.copilot_completion.CopilotCLICompletion`.
         """
+        if self.is_copilot:
+            return self._to_copilot_model_config()
+
         provider_key = str(self.llm_provider)
         is_builtin = provider_key in _PROVIDER_TO_MODEL_PROVIDER
         model_provider = _PROVIDER_TO_MODEL_PROVIDER.get(provider_key, provider_key)
@@ -161,17 +180,6 @@ def to_model_config(self) -> ModelConfig:
 
         azure_deployment_name = self.model if model_provider == "azure" else None
 
-        rate_limit: RateLimitConfig | None = None
-        if self.concurrent_requests and self.concurrent_requests > 0:
-            rate_limit = RateLimitConfig(
-                type=RateLimitType.SlidingWindow,
-                period_in_seconds=1,
-                requests_per_period=max(1, self.concurrent_requests),
-            )
-
-        # Track usage by enabling a metrics processor with no writer (silent).
-        metrics = MetricsConfig(writer=None)
-
         config_kwargs: dict[str, Any] = {
             "model_provider": model_provider,
             "model": self.model,
@@ -182,8 +190,9 @@ def to_model_config(self) -> ModelConfig:
             "azure_deployment_name": azure_deployment_name,
             "call_args": dict(self.call_args),
             "retry": self.retry,
-            "rate_limit": rate_limit,
-            "metrics": metrics,
+            "rate_limit": self._rate_limit_config(),
+            # Track usage by enabling a metrics processor with no writer (silent).
+            "metrics": MetricsConfig(writer=None),
             **init_args,
         }
 
@@ -194,3 +203,32 @@ def to_model_config(self) -> ModelConfig:
             config_kwargs["type"] = provider_key
 
         return ModelConfig(**config_kwargs)
+
+    def _rate_limit_config(self) -> RateLimitConfig | None:
+        if self.concurrent_requests and self.concurrent_requests > 0:
+            return RateLimitConfig(
+                type=RateLimitType.SlidingWindow,
+                period_in_seconds=1,
+                requests_per_period=max(1, self.concurrent_requests),
+            )
+        return None
+
+    def _to_copilot_model_config(self) -> ModelConfig:
+        """Build a ``ModelConfig`` for the GitHub Copilot CLI provider.
+
+        Authentication is handled by the Copilot CLI (logged-in user or a
+        ``github_token`` in ``init_args``), so ``api_key`` is not forwarded.
+        ``init_args`` become extra ``ModelConfig`` fields, which graphrag-llm
+        passes through to the completion constructor.
+        """
+        return ModelConfig(
+            type=COPILOT_COMPLETION_TYPE,
+            model_provider="copilot",
+            model=self.model,
+            auth_method=AuthMethod.ApiKey,
+            call_args=dict(self.call_args),
+            retry=self.retry,
+            rate_limit=self._rate_limit_config(),
+            metrics=MetricsConfig(writer=None),
+            **self.init_args,
+        )