feat(vwb): add dashboard competence testing and health tools

2026-06-02 16:27:19 +02:00
parent d38f0b0f2f
commit 18ed6cb751
23 changed files with 2769 additions and 27 deletions
--- a/tests/unit/test_vlm_grounding_profile.py
+++ b/tests/unit/test_vlm_grounding_profile.py
@@ -0,0 +1,310 @@
+"""Tests pour D5-v2 : profil grounding VLM centralisé + generate_grounding().
+
+Couvre :
+- vlm_config.get_grounding_profile() avec valeurs par défaut et overrides env
+- ollama_client.OllamaClient.generate_grounding() avec mocks requests.post
+- Parsing JSON prefill-aware (reconstitution {"x_pct": ...} → dict)
+- Pas d'appel Ollama live (tous les requests.post sont mockés)
+
+Référence : inbox_claude/2026-05-25_1620_codex-to-claude_GO-revue-strategique-D5v2-C2d.md
+Fix : core/detection/vlm_config.py (get_grounding_profile) +
+      core/detection/ollama_client.py (generate_grounding, _extract_first_json_object)
+"""
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+from unittest.mock import patch, MagicMock
+
+import pytest
+
+
+ROOT = Path(__file__).resolve().parents[2]
+if str(ROOT) not in sys.path:
+    sys.path.insert(0, str(ROOT))
+
+
+# ────────────────────────────────────────────────────────────────────────────
+# vlm_config.get_grounding_profile
+# ────────────────────────────────────────────────────────────────────────────
+
+
+@pytest.mark.unit
+def test_grounding_profile_defaults(monkeypatch):
+    """Sans env vars, valeurs par défaut D5-v2."""
+    for k in ("RPA_GROUNDING_MODEL", "RPA_GROUNDING_CTX", "RPA_GROUNDING_FALLBACK", "RPA_VLM_PREFILL"):
+        monkeypatch.delenv(k, raising=False)
+    from core.detection.vlm_config import get_grounding_profile
+
+    p = get_grounding_profile()
+    assert p["model"] == "qwen3.5:9b"
+    assert p["num_ctx"] == 4096
+    assert p["prefill"] == '{"x_pct":'
+    assert p["temperature"] == 0.0
+    assert p["num_predict"] == 96
+    assert p["fallback_model"] == "qwen2.5vl:7b-rpa"
+    assert p["keep_alive"] == "30m"
+    # qwen3.5 = thinking model → think doit être False côté payload
+    # Le profile expose think comme bool ; False signifie "envoyer think:false"
+    assert p["think"] is False
+
+
+@pytest.mark.unit
+def test_grounding_profile_env_override(monkeypatch):
+    """Env vars override modèle, ctx, fallback."""
+    monkeypatch.setenv("RPA_GROUNDING_MODEL", "qwen2.5vl:7b-rpa")
+    monkeypatch.setenv("RPA_GROUNDING_CTX", "8192")
+    monkeypatch.setenv("RPA_GROUNDING_FALLBACK", "gemma4:latest")
+    from core.detection.vlm_config import get_grounding_profile
+
+    p = get_grounding_profile()
+    assert p["model"] == "qwen2.5vl:7b-rpa"
+    assert p["num_ctx"] == 8192
+    assert p["fallback_model"] == "gemma4:latest"
+    # qwen2.5vl n'est PAS thinking et n'est PAS gemma4 → think=True (rien à envoyer)
+    assert p["think"] is True
+
+
+@pytest.mark.unit
+def test_grounding_profile_ctx_invalid_falls_back_to_default(monkeypatch):
+    """RPA_GROUNDING_CTX non-numeric → fallback 4096."""
+    monkeypatch.setenv("RPA_GROUNDING_CTX", "not_a_number")
+    from core.detection.vlm_config import get_grounding_profile
+
+    p = get_grounding_profile()
+    assert p["num_ctx"] == 4096
+
+
+@pytest.mark.unit
+def test_grounding_profile_prefill_disabled(monkeypatch):
+    """RPA_VLM_PREFILL=false → prefill None."""
+    monkeypatch.setenv("RPA_VLM_PREFILL", "false")
+    from core.detection.vlm_config import get_grounding_profile
+
+    p = get_grounding_profile()
+    assert p["prefill"] is None
+
+
+@pytest.mark.unit
+def test_grounding_profile_gemma4_triggers_think_false(monkeypatch):
+    """Si on remplace par gemma4, think_false doit être déclenché."""
+    monkeypatch.setenv("RPA_GROUNDING_MODEL", "gemma4:latest")
+    from core.detection.vlm_config import get_grounding_profile
+
+    p = get_grounding_profile()
+    assert p["think"] is False  # gemma4 needs think=false
+
+
+# ────────────────────────────────────────────────────────────────────────────
+# _extract_first_json_object
+# ────────────────────────────────────────────────────────────────────────────
+
+
+@pytest.mark.unit
+def test_extract_first_json_object_clean():
+    """JSON propre directement."""
+    from core.detection.ollama_client import _extract_first_json_object
+    obj = _extract_first_json_object('{"x_pct": 0.5, "y_pct": 0.3, "confidence": 0.95}')
+    assert obj == {"x_pct": 0.5, "y_pct": 0.3, "confidence": 0.95}
+
+
+@pytest.mark.unit
+def test_extract_first_json_object_with_trailing_text():
+    """JSON suivi de texte parasite (typique VLM)."""
+    from core.detection.ollama_client import _extract_first_json_object
+    text = '{"x_pct": 0.4, "y_pct": 0.6, "confidence": 0.88}\n\nThe button is located in the bottom-right area.'
+    obj = _extract_first_json_object(text)
+    assert obj["x_pct"] == 0.4
+    assert obj["confidence"] == 0.88
+
+
+@pytest.mark.unit
+def test_extract_first_json_object_with_nested():
+    """JSON avec objet imbriqué."""
+    from core.detection.ollama_client import _extract_first_json_object
+    text = '{"x_pct": 0.5, "meta": {"source": "qwen", "score": 0.9}}'
+    obj = _extract_first_json_object(text)
+    assert obj["meta"]["source"] == "qwen"
+
+
+@pytest.mark.unit
+def test_extract_first_json_object_with_braces_in_strings():
+    """Les accolades dans les strings ne doivent pas perturber le compteur."""
+    from core.detection.ollama_client import _extract_first_json_object
+    text = '{"x_pct": 0.5, "label": "Click {here}"}'
+    obj = _extract_first_json_object(text)
+    assert obj["label"] == "Click {here}"
+
+
+@pytest.mark.unit
+def test_extract_first_json_object_invalid():
+    """Texte sans JSON → None."""
+    from core.detection.ollama_client import _extract_first_json_object
+    assert _extract_first_json_object("no json here at all") is None
+
+
+@pytest.mark.unit
+def test_extract_first_json_object_empty():
+    """Texte vide → None."""
+    from core.detection.ollama_client import _extract_first_json_object
+    assert _extract_first_json_object("") is None
+    assert _extract_first_json_object(None) is None
+
+
+# ────────────────────────────────────────────────────────────────────────────
+# OllamaClient.generate_grounding (mocks requests.post)
+# ────────────────────────────────────────────────────────────────────────────
+
+
+@pytest.fixture
+def mock_client(monkeypatch):
+    """Construit un OllamaClient sans connexion réelle à Ollama."""
+    from core.detection import ollama_client as oc_module
+
+    # Bypass _check_connection
+    monkeypatch.setattr(oc_module.OllamaClient, "_check_connection", lambda self: True)
+    # Force le modèle pour éviter get_vlm_model() qui appelle Ollama
+    client = oc_module.OllamaClient(model="qwen2.5vl:7b-rpa")
+    return client
+
+
+@pytest.mark.unit
+def test_generate_grounding_payload_uses_profile(mock_client, monkeypatch):
+    """Le payload envoyé à Ollama utilise le profile (model, ctx, prefill, temp, etc.)."""
+    monkeypatch.delenv("RPA_VLM_PREFILL", raising=False)
+
+    captured = {}
+
+    def fake_post(url, json=None, timeout=None):
+        captured["url"] = url
+        captured["payload"] = json
+        # Réponse simulée : Ollama renvoie le contenu APRÈS le prefill
+        resp = MagicMock()
+        resp.status_code = 200
+        resp.json.return_value = {
+            "message": {
+                "content": ' 0.5, "y_pct": 0.3, "confidence": 0.95}'
+            }
+        }
+        return resp
+
+    monkeypatch.setattr("core.detection.ollama_client.requests.post", fake_post)
+
+    result = mock_client.generate_grounding(prompt="Find the Save button")
+
+    # Payload vérifié
+    assert captured["url"].endswith("/api/chat")
+    payload = captured["payload"]
+    assert payload["model"] == "qwen3.5:9b"  # défaut grounding
+    assert payload["options"]["num_ctx"] == 4096
+    assert payload["options"]["temperature"] == 0.0
+    assert payload["options"]["num_predict"] == 96
+    # qwen3.5 = thinking → think=false dans payload
+    assert payload.get("think") is False
+    # Le prefill doit être présent dans le dernier message (assistant)
+    last_msg = payload["messages"][-1]
+    assert last_msg["role"] == "assistant"
+    assert last_msg["content"] == '{"x_pct":'
+
+
+@pytest.mark.unit
+def test_generate_grounding_parses_prefilled_json(mock_client, monkeypatch):
+    """Le JSON est correctement reconstitué via prefill + parsé."""
+    monkeypatch.delenv("RPA_VLM_PREFILL", raising=False)
+
+    def fake_post(url, json=None, timeout=None):
+        resp = MagicMock()
+        resp.status_code = 200
+        # Ollama renvoie SANS le prefill (le client le rajoute)
+        resp.json.return_value = {
+            "message": {
+                "content": ' 0.42, "y_pct": 0.68, "confidence": 0.91}'
+            }
+        }
+        return resp
+
+    monkeypatch.setattr("core.detection.ollama_client.requests.post", fake_post)
+
+    result = mock_client.generate_grounding(prompt="Find OK button")
+
+    assert result["success"] is True
+    # response contient le JSON complet reconstitué
+    assert result["response"].startswith('{"x_pct":')
+    # parsed_json est le dict reconstruit
+    parsed = result["parsed_json"]
+    assert parsed is not None
+    assert parsed["x_pct"] == 0.42
+    assert parsed["y_pct"] == 0.68
+    assert parsed["confidence"] == 0.91
+    # profile_used est exposé
+    assert result["profile_used"]["model"] == "qwen3.5:9b"
+    assert result["profile_used"]["num_ctx"] == 4096
+
+
+@pytest.mark.unit
+def test_generate_grounding_restores_original_model(mock_client, monkeypatch):
+    """Après generate_grounding, self.model est restauré (pas de side-effect)."""
+    original = mock_client.model  # qwen2.5vl:7b-rpa
+
+    def fake_post(url, json=None, timeout=None):
+        resp = MagicMock()
+        resp.status_code = 200
+        resp.json.return_value = {"message": {"content": ' 0.5, "y_pct": 0.5, "confidence": 0.5}'}}
+        return resp
+
+    monkeypatch.setattr("core.detection.ollama_client.requests.post", fake_post)
+    mock_client.generate_grounding(prompt="test")
+    assert mock_client.model == original, (
+        f"self.model doit être restauré ({original}), trouvé : {mock_client.model}"
+    )
+
+
+@pytest.mark.unit
+def test_generate_grounding_handles_ollama_error(mock_client, monkeypatch):
+    """Si Ollama retourne 500, success=False, parsed_json=None, model restauré."""
+    def fake_post(url, json=None, timeout=None):
+        resp = MagicMock()
+        resp.status_code = 500
+        resp.text = "Internal error"
+        return resp
+
+    monkeypatch.setattr("core.detection.ollama_client.requests.post", fake_post)
+    result = mock_client.generate_grounding(prompt="test")
+    assert result["success"] is False
+    assert result["parsed_json"] is None
+
+
+@pytest.mark.unit
+def test_generate_grounding_profile_override(mock_client, monkeypatch):
+    """Override profile explicite dans l'appel (pour tests / contextes spéciaux)."""
+    captured = {}
+
+    def fake_post(url, json=None, timeout=None):
+        captured["payload"] = json
+        resp = MagicMock()
+        resp.status_code = 200
+        resp.json.return_value = {"message": {"content": ' 0.1, "y_pct": 0.1, "confidence": 0.5}'}}
+        return resp
+
+    monkeypatch.setattr("core.detection.ollama_client.requests.post", fake_post)
+
+    custom_profile = {
+        "model": "gemma4:latest",
+        "num_ctx": 2048,
+        "prefill": None,
+        "temperature": 0.2,
+        "num_predict": 50,
+        "think": False,
+        "keep_alive": "5m",
+        "fallback_model": "qwen2.5vl:7b-rpa",
+    }
+    result = mock_client.generate_grounding(prompt="test", profile=custom_profile)
+    payload = captured["payload"]
+    assert payload["model"] == "gemma4:latest"
+    assert payload["options"]["num_ctx"] == 2048
+    assert payload["options"]["temperature"] == 0.2
+    # Pas de prefill → pas de message assistant
+    assert payload["messages"][-1]["role"] == "user"
+    # gemma4 needs think=false → injecté
+    assert payload.get("think") is False