fix(backend/copilot): set transcript_content on baseline restore, fix relative import in transcript.py

- _restore_cli_session_for_turn now sets result.transcript_content when loading
  baseline content into the TranscriptBuilder, preventing the _seed_transcript
  guard in stream_chat_completion_sdk from overwriting the builder with a full
  DB reconstruction (which would duplicate entries since load_previous appends).
- Change transcript.py TYPE_CHECKING and runtime ChatMessage import from
  absolute (backend.copilot.model) to relative (.model) to match service.py's
  import style and eliminate Pyright type-identity collisions.
- Unpack _load_prior_transcript tuple return in mode_switch_context_test.py
  and assert dl is not None.
- Add assert result.transcript_content != "" in service_helpers_test.py.

autogpt_platform/backend/backend/copilot/sdk/mode_switch_context_test.py

@@ -277,7 +277,7 @@ class TestSdkToFastModeSwitch:
             "backend.copilot.baseline.service.download_transcript",
             new=AsyncMock(return_value=restore),
         ):
-            covers = await _load_prior_transcript(
+            covers, dl = await _load_prior_transcript(
                 user_id="user-1",
                 session_id="session-1",
                 session_messages=[
@@ -290,6 +290,7 @@ class TestSdkToFastModeSwitch:
 
         # CLI session is valid and covers the prefix.
         assert covers is True
+        assert dl is not None
         assert baseline_builder.entry_count == 2
 
     @pytest.mark.asyncio
@@ -332,7 +333,7 @@ class TestSdkToFastModeSwitch:
             "backend.copilot.baseline.service.download_transcript",
             new=AsyncMock(return_value=restore),
         ):
-            covers = await _load_prior_transcript(
+            covers, dl = await _load_prior_transcript(
                 user_id="user-1",
                 session_id="session-1",
                 session_messages=session_messages,
@@ -341,5 +342,6 @@ class TestSdkToFastModeSwitch:
 
         # With gap filling, covers is True and gap messages are appended.
         assert covers is True
+        assert dl is not None
         # 2 from transcript + 7 gap messages (positions 2..8, excluding last user turn)
         assert baseline_builder.entry_count == 9

autogpt_platform/backend/backend/copilot/sdk/service.py

@@ -2538,6 +2538,9 @@ async def _restore_cli_session_for_turn(
     )
 
     # Load baseline transcript content into builder so the upload path has accurate state.
+    # Also sets result.transcript_content so the _seed_transcript guard in the caller
+    # (``not transcript_content``) does not overwrite this builder state with a DB
+    # reconstruction — which would duplicate entries since load_previous appends.
     if result.baseline_download is not None:
         try:
             raw_for_builder = result.baseline_download.content
@@ -2546,6 +2549,7 @@ async def _restore_cli_session_for_turn(
             stripped = strip_for_upload(raw_for_builder)
             if validate_transcript(stripped):
                 transcript_builder.load_previous(stripped, log_prefix=log_prefix)
+                result.transcript_content = stripped
         except Exception as _load_err:
             logger.debug(
                 "%s Could not load baseline transcript into builder: %s",

autogpt_platform/backend/backend/copilot/sdk/service_helpers_test.py

@@ -859,6 +859,10 @@ class TestRestoreCliSessionModeCheck:
         assert result.context_messages[0].role == "user"
         assert result.context_messages[1].role == "assistant"
         assert "TRANSCRIPT_ASSISTANT" in (result.context_messages[1].content or "")
+        # transcript_content must be non-empty so the _seed_transcript guard in
+        # stream_chat_completion_sdk skips DB reconstruction (which would duplicate
+        # builder entries since load_previous appends).
+        assert result.transcript_content != ""
 
     @pytest.mark.asyncio
     async def test_baseline_mode_gap_present_context_includes_gap(self, tmp_path):

autogpt_platform/backend/backend/copilot/transcript.py

@@ -29,7 +29,7 @@ from backend.util.prompt import CompressResult, compress_context
 from backend.util.workspace_storage import GCSWorkspaceStorage, get_workspace_storage
 
 if TYPE_CHECKING:
-    from backend.copilot.model import ChatMessage
+    from .model import ChatMessage
 
 logger = logging.getLogger(__name__)
 
@@ -858,7 +858,7 @@ def extract_context_messages(
         A list of ``ChatMessage`` objects covering the prior conversation
         context, suitable for injection as conversation history.
     """
-    from backend.copilot.model import ChatMessage as _ChatMessage  # runtime import
+    from .model import ChatMessage as _ChatMessage  # runtime import
 
     prior = session_messages[:-1]