Merge branch 'dev' of github.com:Significant-Gravitas/AutoGPT into fix/dry-run-special-blocks

autogpt_platform/backend/backend/blocks/io.py

@@ -2,6 +2,8 @@ import copy
 from datetime import date, time
 from typing import Any, Optional
 
+from pydantic import AliasChoices, Field
+
 from backend.blocks._base import (
     Block,
     BlockCategory,
@@ -467,7 +469,8 @@ class AgentFileInputBlock(AgentInputBlock):
 
 class AgentDropdownInputBlock(AgentInputBlock):
     """
-    A specialized text input block that relies on placeholder_values to present a dropdown.
+    A specialized text input block that presents a dropdown selector
+    restricted to a fixed set of values.
     """
 
     class Input(AgentInputBlock.Input):
@@ -477,16 +480,23 @@ class AgentDropdownInputBlock(AgentInputBlock):
             advanced=False,
             title="Default Value",
         )
-        placeholder_values: list = SchemaField(
-            description="Possible values for the dropdown.",
+        # Use Field() directly (not SchemaField) to pass validation_alias,
+        # which handles backward compat for legacy "placeholder_values" across
+        # all construction paths (model_construct, __init__, model_validate).
+        options: list = Field(
             default_factory=list,
-            advanced=False,
             title="Dropdown Options",
+            description=(
+                "If provided, renders the input as a dropdown selector "
+                "restricted to these values. Leave empty for free-text input."
+            ),
+            validation_alias=AliasChoices("options", "placeholder_values"),
+            json_schema_extra={"advanced": False, "secret": False},
         )
 
         def generate_schema(self):
             schema = super().generate_schema()
-            if possible_values := self.placeholder_values:
+            if possible_values := self.options:
                 schema["enum"] = possible_values
             return schema
 
@@ -504,13 +514,13 @@ class AgentDropdownInputBlock(AgentInputBlock):
                 {
                     "value": "Option A",
                     "name": "dropdown_1",
-                    "placeholder_values": ["Option A", "Option B", "Option C"],
+                    "options": ["Option A", "Option B", "Option C"],
                     "description": "Dropdown example 1",
                 },
                 {
                     "value": "Option C",
                     "name": "dropdown_2",
-                    "placeholder_values": ["Option A", "Option B", "Option C"],
+                    "options": ["Option A", "Option B", "Option C"],
                     "description": "Dropdown example 2",
                 },
             ],

autogpt_platform/backend/backend/blocks/test/test_block.py

@@ -300,13 +300,27 @@ def test_agent_input_block_ignores_legacy_placeholder_values():
 
 
 def test_dropdown_input_block_produces_enum():
-    """Verify AgentDropdownInputBlock.Input.generate_schema() produces enum."""
-    options = ["Option A", "Option B"]
+    """Verify AgentDropdownInputBlock.Input.generate_schema() produces enum
+    using the canonical 'options' field name."""
+    opts = ["Option A", "Option B"]
     instance = AgentDropdownInputBlock.Input.model_construct(
-        name="choice", value=None, placeholder_values=options
+        name="choice", value=None, options=opts
     )
     schema = instance.generate_schema()
-    assert schema.get("enum") == options
+    assert schema.get("enum") == opts
+
+
+def test_dropdown_input_block_legacy_placeholder_values_produces_enum():
+    """Verify backward compat: passing legacy 'placeholder_values' to
+    AgentDropdownInputBlock still produces enum via model_construct remap."""
+    opts = ["Option A", "Option B"]
+    instance = AgentDropdownInputBlock.Input.model_construct(
+        name="choice", value=None, placeholder_values=opts
+    )
+    schema = instance.generate_schema()
+    assert (
+        schema.get("enum") == opts
+    ), "Legacy placeholder_values should be remapped to options"
 
 
 def test_generate_schema_integration_legacy_placeholder_values():
@@ -329,11 +343,11 @@ def test_generate_schema_integration_legacy_placeholder_values():
 
 def test_generate_schema_integration_dropdown_produces_enum():
     """Test the full Graph._generate_schema path with AgentDropdownInputBlock
-    — verifies enum IS produced for dropdown blocks."""
+    — verifies enum IS produced for dropdown blocks using canonical field name."""
     dropdown_input_default = {
         "name": "color",
         "value": None,
-        "placeholder_values": ["Red", "Green", "Blue"],
+        "options": ["Red", "Green", "Blue"],
     }
     result = BaseGraph._generate_schema(
         (AgentDropdownInputBlock.Input, dropdown_input_default),
@@ -344,3 +358,36 @@ def test_generate_schema_integration_dropdown_produces_enum():
         "Green",
         "Blue",
     ], "Graph schema should contain enum from AgentDropdownInputBlock"
+
+
+def test_generate_schema_integration_dropdown_legacy_placeholder_values():
+    """Test the full Graph._generate_schema path with AgentDropdownInputBlock
+    using legacy 'placeholder_values' — verifies backward compat produces enum."""
+    legacy_dropdown_input_default = {
+        "name": "color",
+        "value": None,
+        "placeholder_values": ["Red", "Green", "Blue"],
+    }
+    result = BaseGraph._generate_schema(
+        (AgentDropdownInputBlock.Input, legacy_dropdown_input_default),
+    )
+    color_props = result["properties"]["color"]
+    assert color_props.get("enum") == [
+        "Red",
+        "Green",
+        "Blue",
+    ], "Legacy placeholder_values should still produce enum via model_construct remap"
+
+
+def test_dropdown_input_block_init_legacy_placeholder_values():
+    """Verify backward compat: constructing AgentDropdownInputBlock.Input via
+    model_validate with legacy 'placeholder_values' correctly maps to 'options'."""
+    opts = ["Option A", "Option B"]
+    instance = AgentDropdownInputBlock.Input.model_validate(
+        {"name": "choice", "value": None, "placeholder_values": opts}
+    )
+    assert (
+        instance.options == opts
+    ), "Legacy placeholder_values should be remapped to options via model_validate"
+    schema = instance.generate_schema()
+    assert schema.get("enum") == opts

autogpt_platform/backend/backend/copilot/sdk/agent_generation_guide.md

@@ -89,8 +89,8 @@ These define the agent's interface — what it accepts and what it produces.
 
 **AgentDropdownInputBlock** (ID: `655d6fdf-a334-421c-b733-520549c07cd1`):
 - Specialized input block that presents a dropdown/select to the user
-- Required `input_default` fields: `name` (str), `placeholder_values` (list of options, must have at least one)
-- Optional: `title`, `description`, `value` (default selection)
+- Required `input_default` fields: `name` (str)
+- Optional: `options` (list of dropdown values; when omitted/empty, input behaves as free-text), `title`, `description`, `value` (default selection)
 - Output: `result` — the user-selected value at runtime
 - Use this instead of AgentInputBlock when the user should pick from a fixed set of options
 

autogpt_platform/backend/backend/copilot/sdk/response_adapter_test.py

@@ -29,6 +29,7 @@ from backend.copilot.response_model import (
     StreamToolOutputAvailable,
 )
 
+from .compaction import compaction_events
 from .response_adapter import SDKResponseAdapter
 from .tool_adapter import MCP_TOOL_PREFIX
 from .tool_adapter import _pending_tool_outputs as _pto
@@ -689,3 +690,102 @@ def test_already_resolved_tool_skipped_in_user_message():
     assert (
         len(output_events) == 0
     ), "Already-resolved tool should not emit duplicate output"
+
+
+# -- _end_text_if_open before compaction -------------------------------------
+
+
+def test_end_text_if_open_emits_text_end_before_finish_step():
+    """StreamTextEnd must be emitted before StreamFinishStep during compaction.
+
+    When ``emit_end_if_ready`` fires compaction events while a text block is
+    still open, ``_end_text_if_open`` must close it first.  If StreamFinishStep
+    arrives before StreamTextEnd, the Vercel AI SDK clears ``activeTextParts``
+    and raises "Received text-end for missing text part".
+    """
+    adapter = _adapter()
+
+    # Open a text block by processing an AssistantMessage with text
+    msg = AssistantMessage(content=[TextBlock(text="partial response")], model="test")
+    adapter.convert_message(msg)
+    assert adapter.has_started_text
+    assert not adapter.has_ended_text
+
+    # Simulate what service.py does before yielding compaction events
+    pre_close: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(pre_close)
+    combined = pre_close + list(compaction_events("Compacted transcript"))
+
+    text_end_idx = next(
+        (i for i, e in enumerate(combined) if isinstance(e, StreamTextEnd)), None
+    )
+    finish_step_idx = next(
+        (i for i, e in enumerate(combined) if isinstance(e, StreamFinishStep)), None
+    )
+
+    assert text_end_idx is not None, "StreamTextEnd must be present"
+    assert finish_step_idx is not None, "StreamFinishStep must be present"
+    assert text_end_idx < finish_step_idx, (
+        f"StreamTextEnd (idx={text_end_idx}) must precede "
+        f"StreamFinishStep (idx={finish_step_idx}) — otherwise the Vercel AI SDK "
+        "clears activeTextParts before text-end arrives"
+    )
+
+
+def test_step_open_must_reset_after_compaction_finish_step():
+    """Adapter step_open must be reset when compaction emits StreamFinishStep.
+
+    Compaction events bypass the adapter, so service.py must explicitly clear
+    step_open after yielding a StreamFinishStep from compaction. Without this,
+    the next AssistantMessage skips StreamStartStep because the adapter still
+    thinks a step is open.
+    """
+    adapter = _adapter()
+
+    # Open a step + text block via an AssistantMessage
+    msg = AssistantMessage(content=[TextBlock(text="thinking...")], model="test")
+    adapter.convert_message(msg)
+    assert adapter.step_open is True
+
+    # Simulate what service.py does: close text, then check compaction events
+    pre_close: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(pre_close)
+
+    events = list(compaction_events("Compacted transcript"))
+    if any(isinstance(ev, StreamFinishStep) for ev in events):
+        adapter.step_open = False
+
+    assert (
+        adapter.step_open is False
+    ), "step_open must be False after compaction emits StreamFinishStep"
+
+    # Next AssistantMessage must open a new step
+    msg2 = AssistantMessage(content=[TextBlock(text="continued")], model="test")
+    results = adapter.convert_message(msg2)
+    assert any(
+        isinstance(r, StreamStartStep) for r in results
+    ), "A new StreamStartStep must be emitted after compaction closed the step"
+
+
+def test_end_text_if_open_no_op_when_no_text_open():
+    """_end_text_if_open emits nothing when no text block is open."""
+    adapter = _adapter()
+    results: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(results)
+    assert results == []
+
+
+def test_end_text_if_open_no_op_after_text_already_ended():
+    """_end_text_if_open emits nothing when the text block is already closed."""
+    adapter = _adapter()
+    msg = AssistantMessage(content=[TextBlock(text="hello")], model="test")
+    adapter.convert_message(msg)
+    # Close it once
+    first: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(first)
+    assert len(first) == 1
+    assert isinstance(first[0], StreamTextEnd)
+    # Second call must be a no-op
+    second: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(second)
+    assert second == []

autogpt_platform/backend/backend/copilot/sdk/retry_scenarios_test.py

@@ -1487,3 +1487,188 @@ class TestStreamChatCompletionRetryIntegration:
         errors = [e for e in events if isinstance(e, StreamError)]
         assert not errors, f"Unexpected StreamError: {errors}"
         assert any(isinstance(e, StreamStart) for e in events)
+
+    @pytest.mark.asyncio
+    async def test_result_message_success_subtype_prompt_too_long_triggers_compaction(
+        self,
+    ):
+        """CLI returns ResultMessage(subtype="success") with result="Prompt is too long".
+
+        The SDK internally compacts but the transcript is still too long.  It
+        returns subtype="success" (process completed) with result="Prompt is
+        too long" (the actual rejection message).  The retry loop must detect
+        this as a context-length error and trigger compaction — the subtype
+        "success" must not fool it into treating this as a real response.
+        """
+        import contextlib
+
+        from claude_agent_sdk import ResultMessage
+
+        from backend.copilot.response_model import StreamError, StreamStart
+        from backend.copilot.sdk.service import stream_chat_completion_sdk
+
+        session = self._make_session()
+        success_result = self._make_result_message()
+        attempt_count = [0]
+
+        error_result = ResultMessage(
+            subtype="success",
+            result="Prompt is too long",
+            duration_ms=100,
+            duration_api_ms=0,
+            is_error=False,
+            num_turns=1,
+            session_id="test-session-id",
+        )
+
+        def _client_factory(*args, **kwargs):
+            attempt_count[0] += 1
+
+            async def _receive_error():
+                yield error_result
+
+            async def _receive_success():
+                yield success_result
+
+            client = MagicMock()
+            client._transport = MagicMock()
+            client._transport.write = AsyncMock()
+            client.query = AsyncMock()
+            if attempt_count[0] == 1:
+                client.receive_response = _receive_error
+            else:
+                client.receive_response = _receive_success
+            cm = AsyncMock()
+            cm.__aenter__.return_value = client
+            cm.__aexit__.return_value = None
+            return cm
+
+        original_transcript = _build_transcript(
+            [("user", "prior question"), ("assistant", "prior answer")]
+        )
+        compacted_transcript = _build_transcript(
+            [("user", "[summary]"), ("assistant", "summary reply")]
+        )
+
+        patches = _make_sdk_patches(
+            session,
+            original_transcript=original_transcript,
+            compacted_transcript=compacted_transcript,
+            client_side_effect=_client_factory,
+        )
+
+        events = []
+        with contextlib.ExitStack() as stack:
+            for target, kwargs in patches:
+                stack.enter_context(patch(target, **kwargs))
+            async for event in stream_chat_completion_sdk(
+                session_id="test-session-id",
+                message="hello",
+                is_user_message=True,
+                user_id="test-user",
+                session=session,
+            ):
+                events.append(event)
+
+        assert attempt_count[0] == 2, (
+            f"Expected 2 SDK attempts (subtype='success' with 'Prompt is too long' "
+            f"result should trigger compaction retry), got {attempt_count[0]}"
+        )
+        errors = [e for e in events if isinstance(e, StreamError)]
+        assert not errors, f"Unexpected StreamError: {errors}"
+        assert any(isinstance(e, StreamStart) for e in events)
+
+    @pytest.mark.asyncio
+    async def test_assistant_message_error_content_prompt_too_long_triggers_compaction(
+        self,
+    ):
+        """AssistantMessage.error="invalid_request" with content "Prompt is too long".
+
+        The SDK returns error type "invalid_request" but puts the actual
+        rejection message ("Prompt is too long") in the content blocks.
+        The retry loop must detect this via content inspection (sdk_error
+        being set confirms it's an error message, not user content).
+        """
+        import contextlib
+
+        from claude_agent_sdk import AssistantMessage, ResultMessage, TextBlock
+
+        from backend.copilot.response_model import StreamError, StreamStart
+        from backend.copilot.sdk.service import stream_chat_completion_sdk
+
+        session = self._make_session()
+        success_result = self._make_result_message()
+        attempt_count = [0]
+
+        def _client_factory(*args, **kwargs):
+            attempt_count[0] += 1
+
+            async def _receive_error():
+                # SDK returns invalid_request with "Prompt is too long" in content.
+                # ResultMessage.result is a non-PTL value ("done") to isolate
+                # the AssistantMessage content detection path exclusively.
+                yield AssistantMessage(
+                    content=[TextBlock(text="Prompt is too long")],
+                    model="<synthetic>",
+                    error="invalid_request",
+                )
+                yield ResultMessage(
+                    subtype="success",
+                    result="done",
+                    duration_ms=100,
+                    duration_api_ms=0,
+                    is_error=False,
+                    num_turns=1,
+                    session_id="test-session-id",
+                )
+
+            async def _receive_success():
+                yield success_result
+
+            client = MagicMock()
+            client._transport = MagicMock()
+            client._transport.write = AsyncMock()
+            client.query = AsyncMock()
+            if attempt_count[0] == 1:
+                client.receive_response = _receive_error
+            else:
+                client.receive_response = _receive_success
+            cm = AsyncMock()
+            cm.__aenter__.return_value = client
+            cm.__aexit__.return_value = None
+            return cm
+
+        original_transcript = _build_transcript(
+            [("user", "prior question"), ("assistant", "prior answer")]
+        )
+        compacted_transcript = _build_transcript(
+            [("user", "[summary]"), ("assistant", "summary reply")]
+        )
+
+        patches = _make_sdk_patches(
+            session,
+            original_transcript=original_transcript,
+            compacted_transcript=compacted_transcript,
+            client_side_effect=_client_factory,
+        )
+
+        events = []
+        with contextlib.ExitStack() as stack:
+            for target, kwargs in patches:
+                stack.enter_context(patch(target, **kwargs))
+            async for event in stream_chat_completion_sdk(
+                session_id="test-session-id",
+                message="hello",
+                is_user_message=True,
+                user_id="test-user",
+                session=session,
+            ):
+                events.append(event)
+
+        assert attempt_count[0] == 2, (
+            f"Expected 2 SDK attempts (AssistantMessage error content 'Prompt is "
+            f"too long' should trigger compaction retry), got {attempt_count[0]}"
+        )
+        errors = [e for e in events if isinstance(e, StreamError)]
+        assert not errors, f"Unexpected StreamError: {errors}"
+        assert any(isinstance(e, StreamStart) for e in events)

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

@@ -1310,10 +1310,16 @@ async def _run_stream_attempt(
                 # AssistantMessage.error (not as a Python exception).
                 # Re-raise so the outer retry loop can compact the
                 # transcript and retry with reduced context.
-                # Only check error_text (the error field), not the
-                # content preview — content may contain arbitrary text
-                # that false-positives the pattern match.
-                if _is_prompt_too_long(Exception(error_text)):
+                # Check both error_text and error_preview: sdk_error
+                # being set confirms this is an error message (not user
+                # content), so checking content is safe. The actual
+                # error description (e.g. "Prompt is too long") may be
+                # in the content, not the error type field
+                # (e.g. error="invalid_request", content="Prompt is
+                # too long").
+                if _is_prompt_too_long(Exception(error_text)) or _is_prompt_too_long(
+                    Exception(error_preview)
+                ):
                     logger.warning(
                         "%s Prompt-too-long detected via AssistantMessage "
                         "error — raising for retry",
@@ -1414,13 +1420,16 @@ async def _run_stream_attempt(
                         ctx.log_prefix,
                         sdk_msg.result or "(no error message provided)",
                     )
-                    # If the CLI itself rejected the prompt as too long
-                    # (pre-API check, duration_api_ms=0), re-raise as an
-                    # exception so the retry loop can trigger compaction.
-                    # Without this, the ResultMessage is silently consumed
-                    # and the retry/compaction mechanism is never invoked.
-                    if _is_prompt_too_long(RuntimeError(sdk_msg.result or "")):
-                        raise RuntimeError("Prompt is too long")
+
+                # Check for prompt-too-long regardless of subtype — the
+                # SDK may return subtype="success" with result="Prompt is
+                # too long" when the CLI rejects the prompt before calling
+                # the API (cost_usd=0, no tokens consumed).  If we only
+                # check the "error" subtype path, the stream appears to
+                # complete normally, the synthetic error text is stored
+                # in the transcript, and the session grows without bound.
+                if _is_prompt_too_long(RuntimeError(sdk_msg.result or "")):
+                    raise RuntimeError("Prompt is too long")
 
                 # Capture token usage from ResultMessage.
                 # Anthropic reports cached tokens separately:
@@ -1453,6 +1462,23 @@ async def _run_stream_attempt(
             # Emit compaction end if SDK finished compacting.
             # Sync TranscriptBuilder with the CLI's active context.
             compact_result = await ctx.compaction.emit_end_if_ready(ctx.session)
+            if compact_result.events:
+                # Compaction events end with StreamFinishStep, which maps to
+                # Vercel AI SDK's "finish-step" — that clears activeTextParts.
+                # Close any open text block BEFORE the compaction events so
+                # the text-end arrives before finish-step, preventing
+                # "text-end for missing text part" errors on the frontend.
+                pre_close: list[StreamBaseResponse] = []
+                state.adapter._end_text_if_open(pre_close)
+                # Compaction events bypass the adapter, so sync step state
+                # when a StreamFinishStep is present — otherwise the adapter
+                # will skip StreamStartStep on the next AssistantMessage.
+                if any(
+                    isinstance(ev, StreamFinishStep) for ev in compact_result.events
+                ):
+                    state.adapter.step_open = False
+                for r in pre_close:
+                    yield r
             for ev in compact_result.events:
                 yield ev
             entries_replaced = False

autogpt_platform/backend/load-tests/tests/api/graph-execution-test.js

@@ -22,7 +22,6 @@ function generateTestGraph(name = null) {
           input_default: {
             name: "Load Test Input",
             description: "Test input for load testing",
-            placeholder_values: {},
           },
           input_nodes: [],
           output_nodes: ["output_node"],
@@ -59,11 +58,7 @@ function generateExecutionInputs() {
     "Load Test Input": {
       name: "Load Test Input",
       description: "Test input for load testing",
-      placeholder_values: {
-        test_data: `Test execution at ${new Date().toISOString()}`,
-        test_parameter: Math.random().toString(36).substr(2, 9),
-        numeric_value: Math.floor(Math.random() * 1000),
-      },
+      value: `Test execution at ${new Date().toISOString()}`,
     },
   };
 }

docs/integrations/block-integrations/basic.md

@@ -169,7 +169,7 @@ Block for dropdown text selection.
 
 ### How it works
 <!-- MANUAL: how_it_works -->
-This block provides a dropdown selection input for users interacting with your agent. You define the available options using placeholder_values, and users select one option from the list at runtime.
+This block provides a dropdown selection input for users interacting with your agent. You define the available options using the `options` field, and users select one option from the list at runtime.
 
 This is ideal when you want to constrain user input to a predefined set of choices, ensuring valid input and simplifying the user experience. The selected value is passed to downstream blocks in your workflow.
 <!-- END MANUAL -->
@@ -184,7 +184,7 @@ This is ideal when you want to constrain user input to a predefined set of choic
 | description | The description of the input. | str | No |
 | advanced | Whether to show the input in the advanced section, if the field is not required. | bool | No |
 | secret | Whether the input should be treated as a secret. | bool | No |
-| placeholder_values | Possible values for the dropdown. | List[Any] | No |
+| options | If provided, renders the input as a dropdown selector restricted to these values. Leave empty for free-text input. | List[Any] | No |
 
 ### Outputs