fix: support llama.cpp as direct backend without Lemonade

[stampby]

Summary

• LemonadeClient: Don't append /api/v1 when base_url already has a path (e.g. http://localhost:8081/v1 stays as-is). Bare URLs without a path still get /api/v1 appended for Lemonade compatibility.
• LemonadeClient: Skip _ensure_model_loaded for non-Lemonade backends, detected by absence of all_models_loaded/model_loaded fields in /health response.
• System status: Fallback to /v1/models for llama.cpp-style model detection when /health doesn't report model_loaded. Checks both {"models": [...]} (llama.cpp) and {"data": [...]} (OpenAI) formats.

Problem

When using gaia chat --ui --base-url http://localhost:8081/v1 with llama.cpp directly (no Lemonade), three issues caused 404 errors:

1. base_url normalization unconditionally appended /api/v1, turning http://localhost:8081/v1 into http://localhost:8081/v1/api/v1 — all API calls hit wrong paths
2. _ensure_model_loaded called Lemonade-specific load_model endpoint which doesn't exist on llama.cpp
3. System status couldn't detect loaded models because llama.cpp's /health returns {"status":"ok"} without model_loaded field

Test plan

• gaia chat --ui --base-url http://localhost:8081/v1 --no-lemonade-check works with llama.cpp
• gaia chat --ui with default Lemonade backend still works (no regression)
• Bare --base-url http://localhost:8080 still gets /api/v1 appended
• System status correctly shows model_loaded from llama.cpp /v1/models
• Existing unit tests pass

🤖 Generated with Claude Code

[kovtcharov]

@claude please review this PR and make sure it's backwards compatible.

[kovtcharov]

@claude what changes are needed in this PR to get all tests in CI to pass?

[github-actions]

Claude finished @kovtcharov's task —— View job

---

CI Test Failure Analysis

• Read PR diff to understand changes
• Check PR branch code
• Run lint checks on PR code
• Analyze unit tests
• Identify failing tests and root causes
• Provide specific fixes needed

---

Summary

The PR introduces two changes to lemonade_client.py:

1. _ensure_model_loaded now calls health_check() first to detect non-Lemonade backends
2. URL path handling - only appends /api/v1 to bare URLs without a path

Required Test Fixes

The existing unit tests in tests/unit/test_lemonade_model_loading.py will fail because they mock get_status and load_model but don't mock health_check(). The new code calls health_check() before get_status().

Tests that need updating:

Test	File	Issue
test_calls_load_when_model_not_loaded	test_lemonade_model_loading.py:15	Missing health_check mock
test_skips_load_when_model_already_loaded	test_lemonade_model_loading.py:33	Missing health_check mock
test_handles_status_check_error_gracefully	test_lemonade_model_loading.py:65	Missing health_check mock
test_ensure_model_loaded_passes_prompt_false	test_lemonade_model_loading.py:175	Missing health_check mock
test_model_loaded_when_not_present	test_lemonade_model_loading.py:201	Missing health_check mock
test_model_not_loaded_when_already_present	test_lemonade_model_loading.py:245	Missing health_check mock

Fix Required:

Add @patch.object(LemonadeClient, "health_check") to each test and configure the mock to return a Lemonade-style response:

@patch.object(LemonadeClient, "health_check")
@patch.object(LemonadeClient, "get_status")
@patch.object(LemonadeClient, "load_model")
def test_calls_load_when_model_not_loaded(self, mock_load, mock_status, mock_health):
    # Mock health_check to return Lemonade-style response (so it continues to get_status)
    mock_health.return_value = {"status": "ok", "model_loaded": "some-model"}
    # ... rest of test unchanged

Additional Test Needed:

Add a new test to verify the non-Lemonade backend detection works:

@patch.object(LemonadeClient, "health_check")
@patch.object(LemonadeClient, "get_status")
@patch.object(LemonadeClient, "load_model")
def test_skips_load_for_non_lemonade_backend(self, mock_load, mock_status, mock_health):
    """Verify _ensure_model_loaded skips load for non-Lemonade backends (llama.cpp)."""
    client = LemonadeClient(host="localhost", port=8000)
    # Return llama.cpp-style health response (no model_loaded field)
    mock_health.return_value = {"status": "ok"}
    
    client._ensure_model_loaded("model-a", auto_download=True)
    
    # Should NOT call get_status or load_model for non-Lemonade backends
    mock_status.assert_not_called()
    mock_load.assert_not_called()

Backwards Compatibility ✅

The changes are backwards compatible:

• URLs with paths (e.g., http://localhost:8081/v1) are preserved as-is
• Bare URLs (e.g., http://localhost:8080) still get /api/v1 appended
• Lemonade backends continue to work normally (health check returns model_loaded field)

---

[itomek]

Thanks for the PR, @stampby! One ask before merge: a parametrized test for the __init__ URL normalization. Something like:

  @pytest.mark.parametrize("base_url,expected", [
    ("http://localhost:8000", "http://localhost:8000/api/v1"), # bare → appended
    ("http://localhost:8081/v1", "http://localhost:8081/v1"), # llama.cpp — the fix
    ("http://localhost:8000/api/v1", "http://localhost:8000/api/v1"), # already normalized
    ("http://proxy.example.com/lemonade", ???), # see question below
  ])

  def test_base_url_normalization(base_url, expected):

  assert LemonadeClient(base_url=base_url).base_url == expected