diff --git a/tests/test_image_chat_cli.py b/tests/test_image_chat_cli.py
new file mode 100644
index 000000000..899866150
--- /dev/null
+++ b/tests/test_image_chat_cli.py
@@ -0,0 +1,285 @@
+#!/usr/bin/env python3
+"""
+Chat interactif en ligne de commande avec gemma4:26b via Ollama.
+
+Usage interactif :
+    python tests/test_image_chat_cli.py
+    # puis taper des questions sur l'image fournie
+
+Usage one-shot :
+    python tests/test_image_chat_cli.py /chemin/vers/image.png "Que vois-tu ?"
+
+Usage avec modèle différent :
+    python tests/test_image_chat_cli.py --model qwen3-vl:8b image.png
+
+Le script utilise l'API Ollama directement (via la lib `ollama` du projet,
+`ollama==0.6.1` dans requirements.txt).
+"""
+
+import argparse
+import base64
+import sys
+from pathlib import Path
+
+try:
+    import ollama
+except ImportError:
+    print("ERREUR : la librairie 'ollama' n'est pas installée.")
+    print("Installez-la avec : pip install ollama")
+    sys.exit(1)
+
+
+DEFAULT_MODEL = "gemma4:26b"
+
+
+def encode_image(image_path: str) -> str:
+    """Encode une image en base64 pour l'API Ollama."""
+    path = Path(image_path)
+    if not path.exists():
+        print(f"ERREUR : le fichier '{image_path}' n'existe pas.")
+        sys.exit(1)
+    if not path.is_file():
+        print(f"ERREUR : '{image_path}' n'est pas un fichier.")
+        sys.exit(1)
+    with open(path, "rb") as f:
+        return base64.b64encode(f.read()).decode("utf-8")
+
+
+def get_client(host: str):
+    """Renvoie un client Ollama configuré pour l'hôte donné."""
+    return ollama.Client(host=host)
+
+
+def check_ollama_running(host: str = "http://localhost:11434") -> bool:
+    """Vérifie que le serveur Ollama est accessible."""
+    try:
+        client = get_client(host)
+        client.list()
+        return True
+    except Exception as e:
+        print(f"ERREUR : impossible de joindre Ollama sur {host}")
+        print(f"Détail : {e}")
+        print()
+        print("Assurez-vous qu'Ollama est lancé :")
+        print("  ollama serve")
+        return False
+
+
+def check_model_available(model: str, host: str = "http://localhost:11434") -> bool:
+    """Vérifie que le modèle est disponible dans Ollama."""
+    try:
+        client = get_client(host)
+        tags = client.list()
+        # ollama.list() retourne un ListResponse avec un attribut 'models'
+        models = getattr(tags, "models", [])
+
+        model_names = []
+        for m in models:
+            if isinstance(m, dict):
+                model_names.append(m.get("name", ""))
+            else:
+                model_names.append(getattr(m, "name", str(m)))
+
+        # Correspondance exacte ou préfixe
+        matched = [name for name in model_names if model in name]
+        if matched:
+            return True
+        else:
+            print(f"AVERTISSEMENT : modèle '{model}' non trouvé dans Ollama.")
+            print(f"Modèles disponibles : {', '.join(model_names) or '(aucun)'}")
+            print()
+            print(f"Pour le télécharger :")
+            print(f"  ollama pull {model}")
+            return False
+    except Exception as e:
+        print(f"ERREUR : impossible de lister les modèles : {e}")
+        return False
+
+
+def chat_with_image(image_path: str, model: str, host: str = "http://localhost:11434") -> None:
+    """Mode interactif : charge l'image une fois, puis pose des questions."""
+    client = get_client(host)
+    image_b64 = encode_image(image_path)
+    print(f"🖼️  Image chargée : {image_path}")
+    print(f"🤖 Modèle       : {model}")
+    print(f"🔗 Ollama       : {host}")
+    print()
+    print("Mode interactif — tapez vos questions (ou 'exit'/'quit' pour sortir)")
+    print("Tapez '/image /chemin/nouvelle.png' pour changer d'image")
+    print("-" * 60)
+
+    # Historique de conversation (sans l'image à chaque fois pour économiser la mémoire)
+    messages = []
+
+    while True:
+        try:
+            question = input("\nVous > ").strip()
+        except (EOFError, KeyboardInterrupt):
+            print("\n👋 Au revoir !")
+            break
+
+        if not question:
+            continue
+
+        if question.lower() in ("exit", "quit", "q"):
+            print("👋 Au revoir !")
+            break
+
+        # Changement d'image
+        if question.startswith("/image "):
+            new_path = question[len("/image "):].strip()
+            try:
+                image_b64 = encode_image(new_path)
+                image_path = new_path
+                # Réinitialiser l'historique car image différente
+                messages = []
+                print(f"🖼️  Nouvelle image : {new_path}")
+            except SystemExit:
+                pass
+            continue
+
+        # Construire le message user avec l'image au premier tour
+        # Ensuite, l'image n'est ré-envoyée que si l'historique est vide
+        has_image_in_context = any(
+            isinstance(m.get("images"), list) and len(m["images"]) > 0
+            for m in messages
+        )
+
+        user_msg = {"role": "user", "content": question}
+        if not has_image_in_context:
+            # Première question ou image changée — inclure l'image
+            user_msg["images"] = [image_b64]
+        messages.append(user_msg)
+
+        print(f"🤖 Réponse ({model})...", end=" ", flush=True)
+
+        try:
+            response = client.chat(
+                model=model,
+                messages=messages,
+                stream=True,
+                options={
+                    "temperature": 0.2,
+                    "num_predict": 2048,
+                },
+            )
+
+            full_response = ""
+            print()  # nouvelle ligne après le "..."
+            for chunk in response:
+                content = chunk.get("message", {}).get("content", "")
+                if content:
+                    print(content, end="", flush=True)
+                    full_response += content
+
+            print()  # retour à la ligne après la réponse
+            messages.append({"role": "assistant", "content": full_response})
+
+        except Exception as e:
+            print(f"\n❌ Erreur : {e}")
+            # Retirer le dernier message user en cas d'erreur
+            messages.pop()
+
+
+def one_shot(image_path: str, question: str, model: str, host: str = "http://localhost:11434") -> None:
+    """Mode one-shot : une question, une réponse."""
+    client = get_client(host)
+    image_b64 = encode_image(image_path)
+
+    messages = [
+        {"role": "user", "content": question, "images": [image_b64]},
+    ]
+
+    try:
+        response = client.chat(
+            model=model,
+            messages=messages,
+            stream=True,
+            options={
+                "temperature": 0.2,
+                "num_predict": 2048,
+            },
+        )
+
+        print(f"🤖 {model} — '{question}'\n")
+        for chunk in response:
+            content = chunk.get("message", {}).get("content", "")
+            if content:
+                print(content, end="", flush=True)
+        print()
+
+    except Exception as e:
+        print(f"❌ Erreur : {e}")
+        sys.exit(1)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Chat interactif avec une image via Ollama (gemma4:26b par défaut)",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Exemples :
+  # Mode interactif avec une image
+  python tests/test_image_chat_cli.py screenshot.png
+
+  # Mode one-shot (question directe)
+  python tests/test_image_chat_cli.py screenshot.png "Quels boutons vois-tu ?"
+
+  # Avec un autre modèle
+  python tests/test_image_chat_cli.py --model qwen3-vl:8b screenshot.png
+
+  # Ollama sur une machine distante
+  python tests/test_image_chat_cli.py --host http://dgx:11434 screenshot.png
+        """,
+    )
+    parser.add_argument(
+        "image",
+        nargs="?",
+        help="Chemin vers l'image à analyser",
+    )
+    parser.add_argument(
+        "question",
+        nargs="?",
+        default=None,
+        help="Question one-shot (si absent → mode interactif)",
+    )
+    parser.add_argument(
+        "--model",
+        default=DEFAULT_MODEL,
+        help=f"Modèle Ollama à utiliser (défaut: {DEFAULT_MODEL})",
+    )
+    parser.add_argument(
+        "--host",
+        default="http://localhost:11434",
+        help="URL du serveur Ollama (défaut: http://localhost:11434)",
+    )
+
+    args = parser.parse_args()
+
+    # Vérifications préalables
+    if not check_ollama_running(args.host):
+        sys.exit(1)
+
+    if not check_model_available(args.model, args.host):
+        sys.exit(1)
+
+    if not args.image:
+        print("Utilisation interactive — veuillez fournir le chemin d'une image.")
+        print()
+        print("Usage :")
+        print(f"  python {sys.argv[0]} /chemin/vers/image.png")
+        print(f"  python {sys.argv[0]} /chemin/vers/image.png \"Votre question\"")
+        print()
+        parser.print_help()
+        sys.exit(1)
+
+    if args.question:
+        # Mode one-shot
+        one_shot(args.image, args.question, args.model, args.host)
+    else:
+        # Mode interactif
+        chat_with_image(args.image, args.model, args.host)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/unit/test_capture_io.py b/tests/unit/test_capture_io.py
new file mode 100644
index 000000000..f9a266d8c
--- /dev/null
+++ b/tests/unit/test_capture_io.py
@@ -0,0 +1,155 @@
+"""Tests unitaires de la politique de sauvegarde des captures (agent_v1).
+
+Objectif : réduire le poids disque des captures (90 Go / 13 sessions = trop)
+sans casser la précision du grounding. La politique distingue le *type* de
+shot :
+
+- ``crop``      → PNG lossless (cible de grounding qwen3-vl, précision pixel) ;
+- ``full`` / ``window`` / ``context`` → JPEG ``optimize=True`` (vue humaine /
+  contexte, compression ~5-10x acceptable) ;
+- ``heartbeat`` → JPEG **downscalé** (liveness, pas de grounding → on peut
+  réduire la résolution).
+
+La fonction ``save_capture`` retourne le chemin RÉELLEMENT écrit (extension
+ajustée selon le format), pour que l'appelant streame le bon fichier.
+
+Branche feat/push-log-dgx — réduction du poids de capture (unité testée,
+non encore câblée dans capturer.py).
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+from pathlib import Path
+
+from PIL import Image
+
+_ROOT = str(Path(__file__).resolve().parents[2])
+if _ROOT not in sys.path:
+    sys.path.insert(0, _ROOT)
+
+
+def _noisy_image(width: int, height: int) -> Image.Image:
+    """Image RGB avec du bruit réel.
+
+    Un aplat uni se compresse à quasi-zéro en PNG comme en JPEG : la
+    comparaison de poids serait truquée. On injecte du bruit pour que la
+    différence PNG/JPEG soit représentative d'un vrai screenshot.
+    """
+    return Image.frombytes("RGB", (width, height), os.urandom(width * height * 3))
+
+
+def test_crop_reste_png_et_dimensions_identiques(tmp_path):
+    """Un crop est sauvé en PNG lossless, dimensions inchangées."""
+    from agent_v0.agent_v1.vision.capture_io import save_capture
+
+    img = _noisy_image(80, 80)
+    base = str(tmp_path / "shot_0001_crop")
+
+    out_path = save_capture(img, base, kind="crop")
+
+    assert out_path.endswith(".png"), f"crop doit rester PNG, obtenu {out_path}"
+    assert os.path.exists(out_path)
+    reread = Image.open(out_path)
+    assert reread.size == (80, 80)
+    # PNG lossless : les pixels doivent être identiques au bruit d'origine.
+    assert list(reread.convert("RGB").getdata()) == list(img.getdata())
+
+
+def test_full_est_jpeg(tmp_path):
+    """Un full est sauvé en JPEG (.jpg)."""
+    from agent_v0.agent_v1.vision.capture_io import save_capture
+
+    img = _noisy_image(640, 480)
+    base = str(tmp_path / "shot_0001_full")
+
+    out_path = save_capture(img, base, kind="full")
+
+    assert out_path.endswith(".jpg"), f"full doit être JPEG, obtenu {out_path}"
+    assert os.path.exists(out_path)
+
+
+def test_full_jpeg_significativement_plus_leger_que_png(tmp_path):
+    """Le JPEG full doit peser nettement moins que le PNG équivalent.
+
+    On génère une image bruitée plein écran (2560×1600) et on compare le
+    poids du JPEG produit par la politique au poids d'un PNG lossless du
+    même contenu. Le gain doit être substantiel (au moins 2x plus léger).
+    """
+    from agent_v0.agent_v1.vision.capture_io import save_capture
+
+    img = _noisy_image(2560, 1600)
+
+    jpeg_path = save_capture(img, str(tmp_path / "full_jpeg"), kind="full")
+    png_ref = tmp_path / "full_ref.png"
+    img.save(png_ref, "PNG")
+
+    jpeg_size = os.path.getsize(jpeg_path)
+    png_size = os.path.getsize(png_ref)
+
+    assert jpeg_size < png_size / 2, (
+        f"JPEG ({jpeg_size}o) doit peser < moitié du PNG ({png_size}o)"
+    )
+
+
+def test_context_et_window_sont_jpeg(tmp_path):
+    """context et window suivent la même politique JPEG que full."""
+    from agent_v0.agent_v1.vision.capture_io import save_capture
+
+    img = _noisy_image(320, 240)
+    for kind in ("context", "window"):
+        out_path = save_capture(img, str(tmp_path / f"x_{kind}"), kind=kind)
+        assert out_path.endswith(".jpg"), f"{kind} doit être JPEG, obtenu {out_path}"
+        assert os.path.exists(out_path)
+
+
+def test_heartbeat_est_downscale(tmp_path):
+    """Un heartbeat est downscalé (largeur réduite) et reste JPEG."""
+    from agent_v0.agent_v1.vision.capture_io import save_capture
+
+    img = _noisy_image(2560, 1600)
+    out_path = save_capture(img, str(tmp_path / "heartbeat_1234"), kind="heartbeat")
+
+    assert out_path.endswith(".jpg"), f"heartbeat doit être JPEG, obtenu {out_path}"
+    reread = Image.open(out_path)
+    assert reread.width < 2560, "heartbeat doit être downscalé en largeur"
+    # Ratio préservé (16:10 → la hauteur doit suivre la largeur réduite).
+    ratio_src = 2560 / 1600
+    ratio_out = reread.width / reread.height
+    assert abs(ratio_src - ratio_out) < 0.02, "le ratio doit être préservé"
+
+
+def test_heartbeat_plus_leger_que_full_jpeg(tmp_path):
+    """Le downscale du heartbeat le rend plus léger que le full JPEG plein res."""
+    from agent_v0.agent_v1.vision.capture_io import save_capture
+
+    img = _noisy_image(2560, 1600)
+    hb = save_capture(img, str(tmp_path / "heartbeat_5678"), kind="heartbeat")
+    full = save_capture(img, str(tmp_path / "shot_9999_full"), kind="full")
+
+    assert os.path.getsize(hb) < os.path.getsize(full), (
+        "le heartbeat downscalé doit peser moins que le full JPEG plein res"
+    )
+
+
+def test_kind_inconnu_leve_erreur(tmp_path):
+    """Un kind non reconnu doit échouer explicitement (fail-closed)."""
+    from agent_v0.agent_v1.vision.capture_io import save_capture
+
+    img = _noisy_image(40, 40)
+    try:
+        save_capture(img, str(tmp_path / "x"), kind="inexistant")
+    except ValueError:
+        return
+    raise AssertionError("un kind inconnu doit lever ValueError")
+
+
+def test_rgba_converti_pour_jpeg(tmp_path):
+    """Une image RGBA doit être convertie avant l'encodage JPEG (pas d'alpha)."""
+    from agent_v0.agent_v1.vision.capture_io import save_capture
+
+    img = Image.new("RGBA", (64, 64), (10, 20, 30, 128))
+    out_path = save_capture(img, str(tmp_path / "shot_rgba_full"), kind="full")
+    assert out_path.endswith(".jpg")
+    assert os.path.exists(out_path)
diff --git a/tests/unit/test_coords_consumption_gap.py b/tests/unit/test_coords_consumption_gap.py
new file mode 100644
index 000000000..4adaa6e75
--- /dev/null
+++ b/tests/unit/test_coords_consumption_gap.py
@@ -0,0 +1,202 @@
+"""Tests documenting the coords consumption gap: write-only navigate coords.
+
+Test 1 (POSITIVE): _resolve_runtime_vars mechanism works — template strings
+like {{navigate_login_coords.x_pct}} resolve correctly when variables dict
+contains the stored coords.
+
+Test 2 (NEGATIVE): _edge_to_normalized_actions bakes coords as literal floats,
+never producing template strings — so runtime variable resolution is never
+triggered for navigate coords, proving the write-only gap.
+
+These tests are evidence, not regression guards. Test 2 documents a known
+structural gap; when the gap is fixed, Test 2 should be updated to assert
+templates ARE produced.
+"""
+
+import os
+import re
+from types import SimpleNamespace
+
+os.environ.setdefault("RPA_AUTH_DISABLED", "true")
+
+from agent_v0.server_v1.replay_engine import (
+    _edge_to_normalized_actions,
+    _resolve_runtime_vars,
+    _resolve_runtime_vars_in_str,
+)
+
+
+# ── Fake fixtures (minimal, per test_visual_anchor_semantics.py pattern) ──
+
+
+class _FakeAction:
+    def __init__(self, type_, target=None, parameters=None):
+        self.type = type_
+        self.target = target
+        self.parameters = parameters or {}
+
+
+class _FakeEdge:
+    def __init__(self, action):
+        self.edge_id = "edge_coords_gap"
+        self.from_node = "node_src"
+        self.to_node = "node_dst"
+        self.action = action
+
+
+# ── Test 1: resolve mechanism is viable ──────────────────────────────────
+
+
+class TestResolveRuntimeVarsViable:
+    """Prove _resolve_runtime_vars infrastructure works with template strings."""
+
+    VARIABLES = {
+        "navigate_login_coords": {
+            "x_pct": 0.15,
+            "y_pct": 0.07,
+            "method": "ocr_anchor",
+        }
+    }
+
+    def test_resolve_in_str_dot_path(self):
+        """{{navigate_login_coords.x_pct}} → "0.15" (string, not float)."""
+        result = _resolve_runtime_vars_in_str(
+            "{{navigate_login_coords.x_pct}}", self.VARIABLES
+        )
+        assert result == "0.15"
+
+    def test_resolve_in_str_y_pct(self):
+        """{{navigate_login_coords.y_pct}} → "0.07"."""
+        result = _resolve_runtime_vars_in_str(
+            "{{navigate_login_coords.y_pct}}", self.VARIABLES
+        )
+        assert result == "0.07"
+
+    def test_resolve_dict_with_templates(self):
+        """_resolve_runtime_vars substitutes templates inside dict values."""
+        action = {
+            "type": "click",
+            "x_pct": "{{navigate_login_coords.x_pct}}",
+            "y_pct": "{{navigate_login_coords.y_pct}}",
+        }
+        resolved = _resolve_runtime_vars(action, self.VARIABLES)
+        assert resolved["x_pct"] == "0.15"
+        assert resolved["y_pct"] == "0.07"
+        assert resolved["type"] == "click"  # no-template strings unchanged
+
+    def test_resolve_nested_dict(self):
+        """_resolve_runtime_vars handles nested dicts with templates."""
+        action = {
+            "parameters": {
+                "coords": "{{navigate_login_coords.x_pct}}",
+            },
+        }
+        resolved = _resolve_runtime_vars(action, self.VARIABLES)
+        assert resolved["parameters"]["coords"] == "0.15"
+
+    def test_resolve_missing_var_leaves_template_intact(self):
+        """Missing variable: template string stays unchanged."""
+        result = _resolve_runtime_vars_in_str(
+            "{{navigate_password_coords.x_pct}}", self.VARIABLES
+        )
+        assert "{{navigate_password_coords.x_pct}}" in result
+
+    def test_resolve_float_passthrough(self):
+        """_resolve_runtime_vars returns non-str values unchanged — floats pass through."""
+        action = {"x_pct": 0.15, "y_pct": 0.07}
+        resolved = _resolve_runtime_vars(action, self.VARIABLES)
+        # Floats are NOT substituted — they're not strings containing {{...}}
+        assert resolved["x_pct"] == 0.15  # literal float, unchanged
+        assert resolved["y_pct"] == 0.07
+
+
+# ── Test 2: compiler gap — literals not templates ────────────────────────
+
+
+class TestCompilerGapLiteralFloats:
+    """Document that _edge_to_normalized_actions produces literal floats,
+    never template strings — so navigate coords are write-only.
+
+    This is the STRUCTURAL GAP: the compiler bakes coords as floats,
+    _resolve_runtime_vars only operates on strings, so stored navigate
+    variables are never consumed downstream.
+    """
+
+    def test_mouse_click_produces_literal_floats(self):
+        """mouse_click edge: x_pct/y_pct are literal floats, not templates."""
+        target = SimpleNamespace(
+            by_position=(0.15, 0.07),
+            by_role=None,
+            by_text=None,
+            context_hints={},
+        )
+        edge = _FakeEdge(
+            _FakeAction("mouse_click", target=target, parameters={"button": "left"})
+        )
+        actions = _edge_to_normalized_actions(edge, params={})
+        assert len(actions) == 1
+        action = actions[0]
+
+        # GAP: coords are literal floats, not template strings
+        assert isinstance(action["x_pct"], float)
+        assert isinstance(action["y_pct"], float)
+        assert action["x_pct"] == 0.15
+        assert action["y_pct"] == 0.07
+
+        # Proof: no template string is ever produced by the compiler
+        assert not isinstance(action["x_pct"], str)
+        assert not isinstance(action["y_pct"], str)
+
+    def test_literal_floats_not_resolved(self):
+        """Literal floats pass through _resolve_runtime_vars unchanged —
+        proving navigate coords stored in variables are NEVER consumed."""
+        target = SimpleNamespace(
+            by_position=(0.15, 0.07),
+            by_role=None,
+            by_text=None,
+            context_hints={},
+        )
+        edge = _FakeEdge(
+            _FakeAction("mouse_click", target=target, parameters={"button": "left"})
+        )
+        actions = _edge_to_normalized_actions(edge, params={})
+        action = actions[0]
+
+        # Simulate variables from a prior navigate_login step
+        different_coords = {
+            "navigate_login_coords": {"x_pct": 0.20, "y_pct": 0.10}
+        }
+        resolved = _resolve_runtime_vars(action, different_coords)
+
+        # Coords REMAIN the original literal floats — no substitution
+        assert resolved["x_pct"] == 0.15  # NOT 0.20 (no substitution)
+        assert resolved["y_pct"] == 0.07  # NOT 0.10 (no substitution)
+
+    def test_text_input_produces_literal_floats(self):
+        """text_input edge: same literal float pattern for click target."""
+        target = SimpleNamespace(
+            by_position=(0.30, 0.50),
+            by_role=None,
+            by_text=None,
+            context_hints={},
+        )
+        edge = _FakeEdge(
+            _FakeAction("text_input", target=target, parameters={"text": "admin"})
+        )
+        actions = _edge_to_normalized_actions(edge, params={})
+        assert len(actions) == 1
+        action = actions[0]
+
+        assert isinstance(action["x_pct"], float)
+        assert isinstance(action["y_pct"], float)
+        assert action["x_pct"] == 0.30
+        assert action["y_pct"] == 0.50
+
+    def test_navigate_action_type_unknown(self):
+        """navigate action type is NOT handled by _edge_to_normalized_actions —
+        falls into the else branch logging "Type d'action inconnu"."""
+        edge = _FakeEdge(_FakeAction("navigate", parameters={"target": "login"}))
+        actions = _edge_to_normalized_actions(edge, params={})
+
+        # navigate produces empty actions — not compiled at all
+        assert actions == []