feat: support non-Gemini LLM providers via LiteLLM model prefixes (#7)

INSTALLATION.md

@@ -103,6 +103,23 @@ GITHUB_TOKEN=your_github_token_here  # Optional for npx, required for web UI
     - Select `repo` scope for private repositories
     - Select `public_repo` for public repositories only
 
+### Using non-Gemini providers
+
+AsyncReview uses DSPy/LiteLLM model prefixes, so provider-specific environment
+variables are picked up automatically when you choose a matching model.
+
+```bash
+# OpenAI
+export OPENAI_API_KEY=your_openai_api_key
+asyncreview review --url https://github.com/org/repo/pull/123 \
+  -q "Review this" --model openai/gpt-4o-mini
+
+# Local Ollama
+ollama serve
+asyncreview review --url https://github.com/org/repo/pull/123 \
+  -q "Review this" --model ollama_chat/qwen3:4b
+```
+
 ## Running AsyncReview Locally
 
 ### Option 1: Using the API Server + Web UI

README.md

@@ -64,13 +64,30 @@ npx skills add AsyncFuncAI/AsyncReview
 
 ### Public Repositories
 
-For public repos, you only need a Gemini API key.
+For public repos, you only need a Gemini API key by default.
 
 ```bash
 export GEMINI_API_KEY="your-key"
 npx asyncreview review --url https://github.com/org/repo/pull/123 -q "Review this"
 ```
 
+### Using non-Gemini providers
+
+AsyncReview passes model names directly to DSPy/LiteLLM, so you can use any
+LiteLLM-compatible provider prefix. Gemini remains the default.
+
+```bash
+# OpenAI
+export OPENAI_API_KEY="your-key"
+npx asyncreview review --url https://github.com/org/repo/pull/123 \
+  -q "Review this" --model openai/gpt-4o-mini
+
+# Local Ollama
+ollama serve
+npx asyncreview review --url https://github.com/org/repo/pull/123 \
+  -q "Review this" --model ollama_chat/qwen3:4b
+```
+
 ### Private Repositories
 
 For private repos, you also need a GitHub token.
@@ -92,7 +109,7 @@ For private repos, you also need a GitHub token.
 ## Configuration
 
 **Required:**
-- **Gemini API Key:** Get one from Google AI Studio. Set as `GEMINI_API_KEY`.
+- **LLM API Key:** Gemini uses `GEMINI_API_KEY`; other LiteLLM providers use their own env vars, such as `OPENAI_API_KEY` or `ANTHROPIC_API_KEY`.
 
 **Optional:**
 - **GitHub Token:** Required for private repositories to access file contents. Set as `GITHUB_TOKEN`.

cli/main.py

@@ -160,7 +160,10 @@ def main():
         "--model", "-m",
         type=str,
         default=None,
-        help="Model to use (e.g. gemini-3.0-pro-preview)",
+        help=(
+            "Model to use (e.g. gemini/gemini-3-pro-preview, "
+            "openai/gpt-4o, anthropic/claude-3-5-sonnet, ollama_chat/qwen3:4b)"
+        ),
     )
 
     args = parser.parse_args()

cli/virtual_runner.py

@@ -16,6 +16,13 @@
 )
 
 
+def _normalize_model_name(model_name: str) -> str:
+    """Keep legacy bare Gemini names working while allowing LiteLLM prefixes."""
+    if model_name.startswith("gemini-"):
+        return f"gemini/{model_name}"
+    return model_name
+
+
 class VirtualReviewRunner:
     """Run RLM code reviews on GitHub PRs without a local repository.
 
@@ -31,7 +38,7 @@ def __init__(
         """Initialize the virtual runner.
 
         Args:
-            model: Override model (e.g. "gemini-3.0-pro-preview")
+            model: Override model (e.g. "gemini/gemini-3-pro-preview" or "openai/gpt-4o")
             quiet: If True, suppress progress output
             on_step: Optional callback for RLM step updates
         """
@@ -59,11 +66,7 @@ def _ensure_configured(self):
             logging.getLogger(name).setLevel(logging.WARNING)
 
         # Configure DSPy with specified model
-        model_name = self.model
-        if not model_name.startswith("gemini/"):
-            model_name = f"gemini/{model_name}"
-
-        dspy.configure(lm=dspy.LM(model_name))
+        dspy.configure(lm=dspy.LM(_normalize_model_name(self.model)))
 
         # Create RLM with custom interpreter that has Deno 2.x fix
         from dspy.primitives.python_interpreter import PythonInterpreter
@@ -76,7 +79,7 @@ def _ensure_configured(self):
             signature="context, question -> answer, sources",
             max_iterations=MAX_ITERATIONS,
             max_llm_calls=MAX_LLM_CALLS,
-            sub_lm=dspy.LM(f"gemini/{SUB_MODEL}" if not SUB_MODEL.startswith("gemini/") else SUB_MODEL),
+            sub_lm=dspy.LM(_normalize_model_name(SUB_MODEL)),
             verbose=not self.quiet,
             interpreter=interpreter,
         )

cr/config.py

@@ -7,6 +7,9 @@
 load_dotenv()
 
 # LLM Configuration
+# LiteLLM reads provider-specific keys such as OPENAI_API_KEY, ANTHROPIC_API_KEY,
+# GROQ_API_KEY, and OLLAMA_API_BASE directly based on the model prefix.
+# GEMINI_API_KEY remains supported for the default Gemini models.
 GEMINI_API_KEY = os.getenv("GEMINI_API_KEY", "")
 MAIN_MODEL = os.getenv("MAIN_MODEL", "gemini/gemini-3-pro-preview")
 SUB_MODEL = os.getenv("SUB_MODEL", "gemini/gemini-3-flash-preview")
@@ -137,4 +140,3 @@
     "test/**",
     "spec/**",
 ]
-

npx/python/cli/main.py

@@ -292,7 +292,10 @@ def main():
         "--model", "-m",
         type=str,
         default=None,
-        help="Model to use (e.g. gemini-3.0-pro-preview)",
+        help=(
+            "Model to use (e.g. gemini/gemini-3-pro-preview, "
+            "openai/gpt-4o, anthropic/claude-3-5-sonnet, ollama_chat/qwen3:4b)"
+        ),
     )
     review_parser.add_argument(
         "--submit",

npx/python/cli/virtual_runner.py

@@ -25,6 +25,13 @@
 
 
 
+def _normalize_model_name(model_name: str) -> str:
+    """Keep legacy bare Gemini names working while allowing LiteLLM prefixes."""
+    if model_name.startswith("gemini-"):
+        return f"gemini/{model_name}"
+    return model_name
+
+
 class VirtualReviewRunner:
     """Run RLM code reviews on GitHub PRs and local directories.
 
@@ -41,7 +48,7 @@ def __init__(
         """Initialize the virtual runner.
 
         Args:
-            model: Override model (e.g. "gemini-3.0-pro-preview")
+            model: Override model (e.g. "gemini/gemini-3-pro-preview" or "openai/gpt-4o")
             quiet: If True, suppress progress output
             on_step: Optional callback for RLM step updates
         """
@@ -199,23 +206,18 @@ def _ensure_configured(self):
             logging.getLogger(name).setLevel(logging.WARNING)
 
         # Configure DSPy with specified model (cache=False to prevent disk caching)
-        model_name = self.model
-        if not model_name.startswith("gemini/"):
-            model_name = f"gemini/{model_name}"
-
-        self._lm = dspy.LM(model_name, cache=False)
+        self._lm = dspy.LM(_normalize_model_name(self.model), cache=False)
 
         # Create RLM with custom interpreter that has Deno 2.x fix
         deno_command = build_deno_command()
         interpreter = PythonInterpreter(deno_command=deno_command)
 
         # Standard signature
-        sub_model = f"gemini/{SUB_MODEL}" if not SUB_MODEL.startswith("gemini/") else SUB_MODEL
         self._rlm = dspy.RLM(
             signature="context, question -> answer, sources",
             max_iterations=MAX_ITERATIONS,
             max_llm_calls=MAX_LLM_CALLS,
-            sub_lm=dspy.LM(sub_model, cache=False),
+            sub_lm=dspy.LM(_normalize_model_name(SUB_MODEL), cache=False),
             verbose=not self.quiet,
             interpreter=interpreter,
             tools=self._create_tool_functions(),

npx/python/cr/config.py

@@ -7,6 +7,9 @@
 load_dotenv()
 
 # LLM Configuration
+# LiteLLM reads provider-specific keys such as OPENAI_API_KEY, ANTHROPIC_API_KEY,
+# GROQ_API_KEY, and OLLAMA_API_BASE directly based on the model prefix.
+# GEMINI_API_KEY remains supported for the default Gemini models.
 GEMINI_API_KEY = os.getenv("GEMINI_API_KEY", "")
 MAIN_MODEL = os.getenv("MAIN_MODEL", "gemini/gemini-3-pro-preview")
 SUB_MODEL = os.getenv("SUB_MODEL", "gemini/gemini-3-flash-preview")
@@ -137,4 +140,3 @@
     "test/**",
     "spec/**",
 ]
-

npx/python/tests/test_e2e_virtual_runner.py

@@ -1,4 +1,4 @@
-"""E2E tests for VirtualReviewRunner with real Gemini API and GitHub.
+"""E2E tests for VirtualReviewRunner with real LLM API and GitHub.
 
 These tests verify:
 1. FETCH_FILE tool interception works across iterations
@@ -7,8 +7,8 @@
 4. Multi-turn RLM conversations handle state correctly
 
 Requirements:
-- GEMINI_API_KEY environment variable must be set
-- Internet connection for GitHub and Gemini API
+- GEMINI_API_KEY, OPENAI_API_KEY, or ANTHROPIC_API_KEY environment variable must be set
+- Internet connection for GitHub and the selected LLM provider
 - Deno must be installed and in PATH
 """
 
@@ -17,19 +17,23 @@
 import pytest
 from cli.virtual_runner import VirtualReviewRunner
 
+LLM_API_KEY_VARS = ("GEMINI_API_KEY", "OPENAI_API_KEY", "ANTHROPIC_API_KEY")
+TEST_MODEL = os.getenv("ASYNCREVIEW_TEST_MODEL", "gemini-2.0-flash-exp")
+
 
 # Require explicit API key for E2E tests
 @pytest.fixture(scope="module")
-def gemini_api_key():
-    """Ensure GEMINI_API_KEY is set for E2E tests."""
-    key = os.getenv("GEMINI_API_KEY")
-    if not key:
-        pytest.skip("GEMINI_API_KEY not set, skipping E2E tests")
-    return key
+def llm_api_key():
+    """Ensure an LLM provider API key is set for E2E tests."""
+    for name in LLM_API_KEY_VARS:
+        key = os.getenv(name)
+        if key:
+            return key
+    pytest.skip("No LLM API key set, skipping E2E tests")
 
 
 @pytest.mark.asyncio
-async def test_fetch_file_interception(gemini_api_key):
+async def test_fetch_file_interception(llm_api_key):
     """Test that FETCH_FILE reliably populates repo_files across iterations.
 
     This is the core test for the variable rebuild fix. It verifies that when
@@ -40,7 +44,7 @@ async def test_fetch_file_interception(gemini_api_key):
     url = "https://github.com/stanfordnlp/dspy/pull/9240"
     question = "What is in dspy/predict/rlm.py? Please fetch and analyze the complete contents of this file."
 
-    runner = VirtualReviewRunner(model="gemini-3-flash-preview", quiet=False)
+    runner = VirtualReviewRunner(model=TEST_MODEL, quiet=False)
 
     # Intercept to verify state propagation
     original_acall = None
@@ -81,12 +85,12 @@ async def intercepted_acall(*args, **kwargs):
 
 
 @pytest.mark.asyncio
-async def test_search_code_tool(gemini_api_key):
+async def test_search_code_tool(llm_api_key):
     """Test SEARCH_CODE tool integration."""
     url = "https://github.com/stanfordnlp/dspy/pull/9240"
     question = "Use SEARCH_CODE to find all files related to 'DataFrame'. List the paths you find."
 
-    runner = VirtualReviewRunner(model="gemini-2.0-flash-exp", quiet=True)
+    runner = VirtualReviewRunner(model=TEST_MODEL, quiet=True)
 
     answer, sources, metadata = await runner.review(url, question)
 
@@ -97,12 +101,12 @@ async def test_search_code_tool(gemini_api_key):
 
 
 @pytest.mark.asyncio
-async def test_list_directory_tool(gemini_api_key):
+async def test_list_directory_tool(llm_api_key):
     """Test LIST_DIR tool integration."""
     url = "https://github.com/stanfordnlp/dspy/pull/9240"
     question = "Use LIST_DIR to list the contents of the 'dspy/predict/' directory."
 
-    runner = VirtualReviewRunner(model="gemini-2.0-flash-exp", quiet=True)
+    runner = VirtualReviewRunner(model=TEST_MODEL, quiet=True)
 
     answer, sources, metadata = await runner.review(url, question)
 
@@ -113,13 +117,13 @@ async def test_list_directory_tool(gemini_api_key):
 
 
 @pytest.mark.asyncio
-async def test_multi_file_fetch(gemini_api_key):
+async def test_multi_file_fetch(llm_api_key):
     """Test fetching multiple files in sequence."""
     url = "https://github.com/stanfordnlp/dspy/pull/9240"
     question = ("Find and fetch both dspy/predict/rlm.py and any test file related to RLM. "
                 "Compare their contents briefly.")
 
-    runner = VirtualReviewRunner(model="gemini-2.0-flash-exp", quiet=True)
+    runner = VirtualReviewRunner(model=TEST_MODEL, quiet=True)
 
     answer, sources, metadata = await runner.review(url, question)
 
@@ -132,12 +136,12 @@ async def test_multi_file_fetch(gemini_api_key):
 
 
 @pytest.mark.asyncio
-async def test_error_handling_invalid_path(gemini_api_key):
+async def test_error_handling_invalid_path(llm_api_key):
     """Test that invalid file paths are handled gracefully."""
     url = "https://github.com/stanfordnlp/dspy/pull/9240"
     question = "Try to fetch the file 'nonexistent/fake/path.py' and report what happens."
 
-    runner = VirtualReviewRunner(model="gemini-2.0-flash-exp", quiet=True)
+    runner = VirtualReviewRunner(model=TEST_MODEL, quiet=True)
 
     # Should not raise, even with invalid path
     answer, sources, metadata = await runner.review(url, question)
@@ -150,28 +154,28 @@ async def test_error_handling_invalid_path(gemini_api_key):
 
 
 @pytest.mark.asyncio
-async def test_issue_review(gemini_api_key):
+async def test_issue_review(llm_api_key):
     """Test reviewing a GitHub issue (not just PRs)."""
     # Use a known issue
     url = "https://github.com/stanfordnlp/dspy/issues/100"
     question = "Summarize what this issue is about."
 
-    runner = VirtualReviewRunner(model="gemini-2.0-flash-exp", quiet=True)
+    runner = VirtualReviewRunner(model=TEST_MODEL, quiet=True)
 
     answer, sources, metadata = await runner.review(url, question)
 
     assert answer, "Should return an answer for issues"
     assert metadata.get("type") == "issue", "Should identify as issue type"
 
 
-@pytest.mark.asyncio 
-async def test_context_preservation(gemini_api_key):
+@pytest.mark.asyncio
+async def test_context_preservation(llm_api_key):
     """Test that PR context (diff, description) is preserved alongside tool results."""
     url = "https://github.com/stanfordnlp/dspy/pull/9240"
     question = ("Based on the PR description and the actual code in dspy/predict/rlm.py, "
                 "explain how the DataFrame feature is implemented.")
 
-    runner = VirtualReviewRunner(model="gemini-2.0-flash-exp", quiet=True)
+    runner = VirtualReviewRunner(model=TEST_MODEL, quiet=True)
 
     answer, sources, metadata = await runner.review(url, question)
 
@@ -187,9 +191,9 @@ async def test_context_preservation(gemini_api_key):
     # Allow running tests directly with: python test_e2e_virtual_runner.py
     import sys
 
-    api_key = os.getenv("GEMINI_API_KEY")
+    api_key = next((os.getenv(name) for name in LLM_API_KEY_VARS if os.getenv(name)), None)
     if not api_key:
-        print("ERROR: GEMINI_API_KEY not set")
+        print("ERROR: no LLM API key set")
         sys.exit(1)
 
     print("Running E2E tests...")