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

artemis/memory/transcript.py

@@ -52,7 +52,7 @@ import json
 import time
 from typing import Any, Callable
 
-from langchain_core.messages import AIMessage, BaseMessage, HumanMessage
+from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, ToolMessage
 
 from artemis.memory.step_memory import StepMemoryService
 from artemis.utils.logger import get_logger
@@ -288,6 +288,7 @@ class TranscriptLedger:
         self._restored: list[BaseMessage] = []
         self._active: list[BaseMessage] = []
         self._staged: list[BaseMessage] | None = None
+        self._fresh_tool_messages: set[int] = set()
         self._turn_count = 0
 
         # L2/L3 chunk compression. ``_turns`` records each committed
@@ -623,6 +624,7 @@ class TranscriptLedger:
         step_key: str | None = None,
         validator_result: Any | None = None,
         extra_step_keys: tuple[str, ...] | list[str] = (),
+        tool_results: dict[str, list[BaseMessage]] | None = None,
     ) -> None:
         """Move the staged turn into the active region.
 
@@ -642,12 +644,34 @@ class TranscriptLedger:
             return
         staged = self._staged
         self._staged = None
+        if tool_results:
+            self._fresh_tool_messages.update(
+                id(message) for messages in tool_results.values() for message in messages
+            )
+            # Replace declaration acknowledgements before committing the turn.
+            # Each result retains the model's original call id, including bursts
+            # containing several calls to the same plugin tool.
+            staged = [
+                replacement
+                for message in staged
+                for replacement in (
+                    tool_results.get(message.tool_call_id, [message])
+                    if isinstance(message, ToolMessage) and message.content == "Action Recorded"
+                    else [message]
+                )
+            ]
         self._last_turn_silent = self._turn_is_silent(staged)
 
         if step_key is not None:
-            observation = self._first_image_message(staged)
-            if observation is not None:
-                self._step_keys[id(observation)] = str(step_key)
+            if tool_results:
+                # The turn's initial phone image is not evidence of a plugin
+                # action. Associate the summary only with the returned images.
+                observations = [m for result in tool_results.values() for m in result]
+            else:
+                observations = [self._first_image_message(staged)]
+            for observation in observations:
+                if observation is not None and self._first_image_message([observation]) is not None:
+                    self._step_keys[id(observation)] = str(step_key)
 
         span_start = len(self._active)
         self._active.extend(staged)
@@ -783,7 +807,12 @@ class TranscriptLedger:
         appending to it (the operator's in-turn tool loop) never mutates the
         ledger regions.
         """
-        self._compressor.compress(self._active, image_scrub_depth=self.effective_image_scrub_depth)
+        self._compressor.compress(
+            self._active,
+            image_scrub_depth=self.effective_image_scrub_depth,
+            protected_message_ids=self._fresh_tool_messages,
+        )
+        self._fresh_tool_messages.clear()
         if self._chunker is not None:
             try:
                 self._chunker.on_render(self)