fix: improve cross-device evidence, request diagnostics and prompt efficiency

artemis/data_engine/engine.py

@@ -1314,10 +1314,12 @@ class DataEngine:
 
     def get_step_image_path(self, step_number: int, which: str = "pre") -> Path | None:
         """Read-only path of a step's pre/post screenshot, if recorded."""
+        from artemis.data_engine.step_evidence import step_image_name
+
         step = self.get_step_record(step_number)
         if step is None:
             return None
-        image_name = step.pre_image_name if which == "pre" else step.post_image_name
+        image_name = step_image_name(step, which)
         if not image_name:
             return None
         path = self.get_image_path(image_name)

artemis/data_engine/history_reader.py

@@ -85,10 +85,12 @@ class OfflineHistoryReader:
         return self.global_base_dir / "images" / f"{image_name}.jpg"
 
     def get_step_image_path(self, step_number: int, which: str = "pre") -> Path | None:
+        from artemis.data_engine.step_evidence import step_image_name
+
         step = self.get_step_record(step_number)
         if step is None:
             return None
-        image_name = step.pre_image_name if which == "pre" else step.post_image_name
+        image_name = step_image_name(step, which)
         if not image_name:
             return None
         path = self.get_image_path(image_name)

artemis/data_engine/step_evidence.py

@@ -0,0 +1,34 @@
+"""Resolve a step's images from the resource that actually produced them."""
+
+from typing import Any
+
+
+def step_image_name(step: Any, which: str) -> str | None:
+    def field(name):
+        return step.get(name) if isinstance(step, dict) else getattr(step, name, None)
+
+    result = field("last_execution_result")
+    execution = result.get("execution") if isinstance(result, dict) else None
+    actions = execution if isinstance(execution, list) else field("action_taken")
+    if isinstance(actions, dict):
+        actions = [actions]
+    actions = actions if isinstance(actions, list) else []
+    plugins = [
+        action
+        for action in actions
+        if isinstance(action, dict)
+        and (action.get("plugin_id") or str(action.get("action", "")).startswith("plug_"))
+    ]
+    if not plugins:
+        return field(f"{which}_image_name")
+    # A burst may span several resources. Its per-call evidence must not be
+    # presented as a before/after pair on one screen.
+    if len(actions) != 1:
+        return None
+    action = plugins[0]
+    if isinstance(execution, list) or f"{which}_image_name" in action:
+        if which == "pre" and not action.get("pre_image_valid"):
+            return None
+        return action.get(f"{which}_image_name")
+    # Older Flash records already stored source-specific images at step level.
+    return field(f"{which}_image_name")

artemis/data_engine/storage.py

@@ -990,6 +990,9 @@ class StorageManager:
                         WHEN excluded.status = 'running' AND traces.status IN ('success', 'failed') THEN traces.payload
                         ELSE excluded.payload
                     END
+                WHERE traces.type != 'llm_diagnostic'
+                   OR COALESCE(json_extract(excluded.payload, '$.sequence'), 0)
+                      > COALESCE(json_extract(traces.payload, '$.sequence'), -1)
                 """,
                 (
                     str(trace.trace_id),