fix(copilot): prevent dry-run metadata from leaking into LLM context

When dry_run=True, the tool output message contained "[DRY RUN] Block
'X' simulated successfully — no real execution occurred." This text was
fed directly into the LLM conversation context, causing the autopilot to
realize it was in simulation mode and change its evaluation behavior —
defeating the purpose of dry-run testing.

Changes:
- Remove [DRY RUN] prefix from execute_block() dry-run message — now
  identical to real execution ("Block 'X' executed successfully")
- Shorten dry_run tool schema descriptions to "Execute in preview mode"
  instead of detailed simulation language that steered LLM behavior
- Update tests to assert dry-run output matches real execution format

The is_dry_run field is preserved in the response model so the frontend
can still show simulation badges via the SSE tool output event.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

autogpt_platform/backend/backend/copilot/tools/helpers.py

@@ -115,10 +115,7 @@ async def execute_block(
                     session_id=session_id,
                 )
             return BlockOutputResponse(
-                message=(
-                    f"[DRY RUN] Block '{block.name}' simulated successfully "
-                    "— no real execution occurred."
-                ),
+                message=f"Block '{block.name}' executed successfully",
                 block_id=block_id,
                 block_name=block.name,
                 outputs=dict(outputs),

autogpt_platform/backend/backend/copilot/tools/run_agent.py

@@ -153,11 +153,7 @@ class RunAgentTool(BaseTool):
                 },
                 "dry_run": {
                     "type": "boolean",
-                    "description": (
-                        "When true, simulates the entire agent execution using an LLM "
-                        "for each block — no real API calls, no credentials needed, "
-                        "no credits charged. Useful for testing agent wiring end-to-end."
-                    ),
+                    "description": "Execute in preview mode.",
                 },
             },
             "required": ["dry_run"],

autogpt_platform/backend/backend/copilot/tools/run_block.py

@@ -51,11 +51,7 @@ class RunBlockTool(BaseTool):
                 },
                 "dry_run": {
                     "type": "boolean",
-                    "description": (
-                        "When true, simulates block execution using an LLM without making any "
-                        "real API calls or producing side effects. Useful for testing agent "
-                        "wiring and previewing outputs."
-                    ),
+                    "description": "Execute in preview mode.",
                 },
             },
             "required": ["block_id", "input_data", "dry_run"],

autogpt_platform/backend/backend/copilot/tools/test_dry_run.py

@@ -238,7 +238,7 @@ async def test_execute_block_dry_run_skips_real_execution():
 
 @pytest.mark.asyncio
 async def test_execute_block_dry_run_response_format():
-    """Dry-run response should contain [DRY RUN] in message and success=True."""
+    """Dry-run response should match real execution message format and have success=True."""
     mock_block = make_mock_block()
 
     async def fake_simulate(block, input_data):
@@ -259,7 +259,8 @@ async def test_execute_block_dry_run_response_format():
         )
 
     assert isinstance(response, BlockOutputResponse)
-    assert "[DRY RUN]" in response.message
+    assert "executed successfully" in response.message
+    assert "[DRY RUN]" not in response.message  # must not leak to LLM context
     assert response.success is True
     assert response.outputs == {"result": ["simulated"]}
 

autogpt_platform/backend/backend/copilot/tools/test_session_dry_run.py

@@ -91,7 +91,7 @@ class TestRunBlockToolSessionDryRun:
 
             # Set up execute_block to return a success
             mock_exec.return_value = MagicMock(
-                message="[DRY RUN] Block executed",
+                message="Block 'TestBlock' executed successfully",
                 success=True,
             )