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

artemis/agents/operator/operator.py

@@ -295,7 +295,16 @@ class OperatorNode:
             kwargs_to_pass.update(extra_kwargs)
             await component(builder, state, self.ctx, **kwargs_to_pass)
 
-        return builder.build()
+        messages = builder.build()
+        if getattr(state, "structured_decisions", None):
+            from langchain_core.messages import HumanMessage
+
+            # The legacy prompt has no preceding AI tool calls. Deliver the
+            # same evidence as observations, without orphan function outputs.
+            for results in self._plugin_result_messages(state.last_execution_result).values():
+                for result in results:
+                    messages.append(HumanMessage(content=result.content))
+        return messages
 
     # ------------------------------------------------------------------
     # Transcript prompt path (M2, flag agent.memory.transcript.enabled)
@@ -358,6 +367,11 @@ class OperatorNode:
             pass
         return ledger
 
+    def _plugin_result_messages(self, report: Any) -> dict[str, list[BaseMessage]]:
+        from artemis.plugins.results import plugin_result_messages
+
+        return plugin_result_messages(report, self.ctx.data_engine)
+
     async def _build_prompt_transcript(
         self,
         state: State,
@@ -385,7 +399,15 @@ class OperatorNode:
         # 1. S region: rendered exactly once per session.
         if not ledger.has_static_prefix:
             static_text = render_transcript_static_system(self.prompts, self.ctx, state)
-            ledger.set_static_prefix([SystemMessage(content=static_text)])
+            prefix = []
+            if getattr(self.ctx, "exploration_id", None):
+                from artemis.exploration.runtime import exploration_prompt
+
+                prefix.append(exploration_prompt(self.ctx))
+            if getattr(self.ctx, "plugins", None):
+                prefix.append(self.ctx.plugins.prompt())
+            prefix.append(static_text)
+            ledger.set_static_prefix([SystemMessage(content="\n".join(prefix))])
 
         # 2. F region cold start: an empty ledger over an existing step record
         # trail means the process restarted — freeze the compiled history once.
@@ -423,10 +445,21 @@ class OperatorNode:
         ledger.commit_staged(
             step_key=getattr(state, "current_step_id", None),
             validator_result=validator_result,
+            tool_results=self._plugin_result_messages(validator_result),
         )
 
         # 4. Current tail: observation + plan recitation + injected components.
         builder = PromptBuilder()
+        guidance = None
+        if getattr(self.ctx, "exploration_id", None):
+            from artemis.exploration.runtime import exploration_messages
+
+            # Guidance belongs at the append-only tail, never before the
+            # reusable rules/history. Assistant progress does not change it.
+            guidance = exploration_messages(self.ctx)[1].content
+            if guidance != getattr(self, "_last_exploration_guidance", None):
+                builder.add_human_content(guidance)
+                self._pending_exploration_guidance = guidance
         builder.add_human_content(f"# CURRENT OBSERVATION [{ledger.elapsed_label()}]")
         components = [
             (PlanRecitationPromptComponent(), {}),
@@ -468,6 +501,24 @@ class OperatorNode:
             mark_ephemeral(tail, builder.ephemeral_indices)
 
         messages = ledger.render([tail])
+        # Compaction can remove the original guidance message. Restore the
+        # exact current instructions once at the new tail, without changing
+        # the reusable prefix or relying on a lossy history capsule.
+        if guidance is not None and not any(
+            guidance
+            in (
+                message.content
+                if isinstance(message.content, str)
+                else "\n".join(
+                    block.get("text", "") for block in message.content if isinstance(block, dict)
+                )
+            )
+            for message in messages
+            if isinstance(message, HumanMessage)
+        ):
+            tail_content.append({"type": "text", "text": guidance})
+            tail.content = tail_content
+            self._pending_exploration_guidance = guidance
         self._transcript_turn_tail = tail
         return messages
 
@@ -482,6 +533,9 @@ class OperatorNode:
             if ledger is not None:
                 base = next(i for i, message in enumerate(messages) if message is tail)
                 ledger.stage_turn(messages[base:])
+                if hasattr(self, "_pending_exploration_guidance"):
+                    self._last_exploration_guidance = self._pending_exploration_guidance
+                    del self._pending_exploration_guidance
         except Exception as e:
             logger.error(f"Failed to stage transcript turn: {e}")
 
@@ -547,7 +601,7 @@ class OperatorNode:
                     t
                     for t in traced_tools
                     if not t.name.startswith("plug_")
-                    and t.name not in {"plugin_load", "plugin_read_file"}
+                    and t.name not in {"plugin_load", "plugin_read_file", "plugin_read_result"}
                 ]
                 traced_tools.extend(plugins.langchain_tools())
             bound_llm = active_llm.bind_tools(tools=traced_tools)
@@ -903,6 +957,9 @@ class OperatorNode:
                     not plan_gate_bounced
                     and iteration < max_iterations - 1
                     and len(action_calls) == 1
+                    and not (
+                        plugins and plugins.is_read_only(normalize_name(action_calls[0]["name"]))
+                    )
                     and not getattr(state, "open_incident", None)
                     and self._plan_changed_this_turn() is False
                 ):
@@ -1151,14 +1208,16 @@ class OperatorNode:
             active_background_tasks=active_background_tasks,
             newly_finished_tasks=newly_finished_tasks,
         )
-        if getattr(self.ctx, "plugins", None):
+        if self._transcript_turn_tail is None and getattr(self.ctx, "plugins", None):
             from langchain_core.messages import SystemMessage
 
             messages.insert(0, SystemMessage(content=self.ctx.plugins.prompt()))
-        if getattr(self.ctx, "exploration_id", None):
+        if self._transcript_turn_tail is None and getattr(self.ctx, "exploration_id", None):
             from artemis.exploration.runtime import exploration_messages
 
-            messages[0:0] = exploration_messages(self.ctx)
+            rules, guidance = exploration_messages(self.ctx)
+            messages.insert(0, rules)
+            messages.append(guidance)
 
         # 8. Invoke LLM Loop
         new_subagent_calls = []
@@ -1310,7 +1369,7 @@ class OperatorNode:
 
         plugins = get_plugin_runtime(self.ctx)
         if plugins and tool_name in plugins.bindings:
-            return [{"action": tool_name, "args": args}], None
+            return [{"action": tool_name, "args": args, "tool_call_id": tc["id"]}], None
 
         # Get device resolution fallbacks
         width = 1080