Update frontend build based on commit 4878665c66

## Background

This PR consolidates and unifies context window management for the
CoPilot backend.

### Problem
The CoPilot backend had **two separate implementations** of context
window management:

1. **`service.py` → `_manage_context_window()`** - Chat service
streaming/continuation
2. **`prompt.py` → `compress_prompt()`** - Sync LLM blocks

This duplication led to inconsistent behavior, maintenance burden, and
duplicate code.

---

## Solution: Unified `compress_context()`

A single async function that handles both use cases:

| Caller | Usage | Behavior |
|--------|-------|----------|
| **Chat service** | `compress_context(msgs, client=openai_client)` |
Summarization → Truncation |
| **LLM blocks** | `compress_context(msgs, client=None)` | Truncation
only (no API call) |

---

## Strategy Order

| Step | Description | Runs When |
|------|-------------|-----------|
| **1. LLM Summarization** | Summarize old messages into single context
message, keep recent 15 | Only if `client` provided |
| **2. Content Truncation** | Progressively truncate message content
(8192→4096→...→128 tokens) | If still over limit |
| **3. Middle-out Deletion** | Delete messages one at a time from center
outward | If still over limit |
| **4. First/Last Trim** | Truncate system prompt and last message
content | Last resort |

### Why This Order?

1. **Summarization first** (if available) - Preserves semantic meaning
of old messages
2. **Content truncation before deletion** - Keeps all conversation
turns, just shorter
3. **Middle-out deletion** - More granular than dropping all old
messages at once
4. **First/last trim** - Only touch system prompt as last resort

---

## Key Fixes

| Issue | Before | After |
|-------|--------|-------|
| **Socket leak** | `AsyncOpenAI` client never closed | `async with`
context manager |
| **Timeout ignored** | `timeout=30` passed to `create()` (invalid) |
`client.with_options(timeout=30)` |
| **OpenAI tool messages** | Not truncated | Properly truncated |
| **Tool pair integrity** | OpenAI format only | Both OpenAI + Anthropic
formats |

---

## Tool Format Support

`_ensure_tool_pairs_intact()` now supports both formats:

### OpenAI Format
```python
# Assistant with tool_calls
{"role": "assistant", "tool_calls": [{"id": "call_1", ...}]}
# Tool response
{"role": "tool", "tool_call_id": "call_1", "content": "result"}
```

### Anthropic Format
```python
# Assistant with tool_use
{"role": "assistant", "content": [{"type": "tool_use", "id": "toolu_1", ...}]}
# Tool result
{"role": "user", "content": [{"type": "tool_result", "tool_use_id": "toolu_1", ...}]}
```

---

## Files Changed

| File | Change |
|------|--------|
| `backend/util/prompt.py` | +450 lines: Add `CompressResult`,
`compress_context()`, helpers |
| `backend/api/features/chat/service.py` | -380 lines: Remove duplicate,
use thin wrapper |
| `backend/blocks/llm.py` | Migrate `llm_call()` to use
`compress_context(client=None)` |
| `backend/util/prompt_test.py` | +400 lines: Comprehensive tests
(OpenAI + Anthropic) |

### Removed
- `compress_prompt()` - Replaced by `compress_context(client=None)`
- `_manage_context_window()` - Replaced by
`compress_context(client=openai_client)`

---

## API

```python
async def compress_context(
    messages: list[dict],
    target_tokens: int = 120_000,
    *,
    model: str = "gpt-4o",
    client: AsyncOpenAI | None = None,  # None = truncation only
    keep_recent: int = 15,
    reserve: int = 2_048,
    start_cap: int = 8_192,
    floor_cap: int = 128,
) -> CompressResult:
    ...

@dataclass
class CompressResult:
    messages: list[dict]
    token_count: int
    was_compacted: bool
    error: str | None = None
    original_token_count: int = 0
    messages_summarized: int = 0
    messages_dropped: int = 0
```

---

## Tests Added

| Test Class | Coverage |
|------------|----------|
| `TestMsgTokens` | Token counting for regular messages, OpenAI tool
calls, Anthropic tool_use |
| `TestTruncateToolMessageContent` | OpenAI + Anthropic tool message
truncation |
| `TestEnsureToolPairsIntact` | OpenAI format (3 tests), Anthropic
format (3 tests), edge cases (3 tests) |
| `TestCompressContext` | No compression, truncation-only, tool pair
preservation, error handling |

---

## Checklist

- [x] Code follows project conventions
- [x] Linting passes (`poetry run format`)
- [x] Type checking passes (`pyright`)
- [x] Tests added for all new functions
- [x] Both OpenAI and Anthropic tool formats supported
- [x] Backward compatible behavior preserved
- [x] All review comments addressed

.gitignore

@@ -180,3 +180,4 @@ autogpt_platform/backend/settings.py
 .claude/settings.local.json
 CLAUDE.local.md
 /autogpt_platform/backend/logs
+.next

autogpt_platform/backend/backend/api/features/chat/service.py

@@ -3,10 +3,13 @@ import logging
 import time
 from asyncio import CancelledError
 from collections.abc import AsyncGenerator
-from dataclasses import dataclass
-from typing import Any, cast
+from typing import TYPE_CHECKING, Any, cast
 
 import openai
+
+if TYPE_CHECKING:
+    from backend.util.prompt import CompressResult
+
 import orjson
 from langfuse import get_client
 from openai import (
@@ -17,7 +20,6 @@ from openai import (
     RateLimitError,
 )
 from openai.types.chat import (
-    ChatCompletionAssistantMessageParam,
     ChatCompletionChunk,
     ChatCompletionMessageParam,
     ChatCompletionStreamOptionsParam,
@@ -31,7 +33,6 @@ from backend.data.understanding import (
     get_business_understanding,
 )
 from backend.util.exceptions import NotFoundError
-from backend.util.prompt import estimate_token_count
 from backend.util.settings import Settings
 
 from . import db as chat_db
@@ -803,402 +804,58 @@ def _is_region_blocked_error(error: Exception) -> bool:
     return "not available in your region" in str(error).lower()
 
 
-# Context window management constants
-TOKEN_THRESHOLD = 120_000
-KEEP_RECENT_MESSAGES = 15
-
-
-@dataclass
-class ContextWindowResult:
-    """Result of context window management."""
-
-    messages: list[dict[str, Any]]
-    token_count: int
-    was_compacted: bool
-    error: str | None = None
-
-
-def _messages_to_dicts(messages: list) -> list[dict[str, Any]]:
-    """Convert message objects to dicts, filtering None values.
-
-    Handles both TypedDict (dict-like) and other message formats.
-    """
-    result = []
-    for msg in messages:
-        if msg is None:
-            continue
-        if isinstance(msg, dict):
-            msg_dict = {k: v for k, v in msg.items() if v is not None}
-        else:
-            msg_dict = dict(msg)
-        result.append(msg_dict)
-    return result
-
-
 async def _manage_context_window(
     messages: list,
     model: str,
     api_key: str | None = None,
     base_url: str | None = None,
-) -> ContextWindowResult:
+) -> "CompressResult":
     """
-    Manage context window by summarizing old messages if token count exceeds threshold.
+    Manage context window using the unified compress_context function.
 
-    This function handles context compaction for LLM calls by:
-    1. Counting tokens in the message list
-    2. If over threshold, summarizing old messages while keeping recent ones
-    3. Ensuring tool_call/tool_response pairs stay intact
-    4. Progressively reducing message count if still over limit
+    This is a thin wrapper that creates an OpenAI client for summarization
+    and delegates to the shared compression logic in prompt.py.
 
     Args:
-        messages: List of messages in OpenAI format (with system prompt if present)
-        model: Model name for token counting
+        messages: List of messages in OpenAI format
+        model: Model name for token counting and summarization
         api_key: API key for summarization calls
         base_url: Base URL for summarization calls
 
     Returns:
-        ContextWindowResult with compacted messages and metadata
+        CompressResult with compacted messages and metadata
     """
-    if not messages:
-        return ContextWindowResult([], 0, False, "No messages to compact")
-
-    messages_dict = _messages_to_dicts(messages)
-
-    # Normalize model name for token counting (tiktoken only supports OpenAI models)
-    token_count_model = model.split("/")[-1] if "/" in model else model
-    if "claude" in token_count_model.lower() or not any(
-        known in token_count_model.lower()
-        for known in ["gpt", "o1", "chatgpt", "text-"]
-    ):
-        token_count_model = "gpt-4o"
-
-    try:
-        token_count = estimate_token_count(messages_dict, model=token_count_model)
-    except Exception as e:
-        logger.warning(f"Token counting failed: {e}. Using gpt-4o approximation.")
-        token_count_model = "gpt-4o"
-        token_count = estimate_token_count(messages_dict, model=token_count_model)
-
-    if token_count <= TOKEN_THRESHOLD:
-        return ContextWindowResult(messages, token_count, False)
-
-    has_system_prompt = messages[0].get("role") == "system"
-    slice_start = max(0, len(messages_dict) - KEEP_RECENT_MESSAGES)
-    recent_messages = _ensure_tool_pairs_intact(
-        messages_dict[-KEEP_RECENT_MESSAGES:], messages_dict, slice_start
-    )
-
-    # Determine old messages to summarize (explicit bounds to avoid slice edge cases)
-    system_msg = messages[0] if has_system_prompt else None
-    if has_system_prompt:
-        old_messages_dict = (
-            messages_dict[1:-KEEP_RECENT_MESSAGES]
-            if len(messages_dict) > KEEP_RECENT_MESSAGES + 1
-            else []
-        )
-    else:
-        old_messages_dict = (
-            messages_dict[:-KEEP_RECENT_MESSAGES]
-            if len(messages_dict) > KEEP_RECENT_MESSAGES
-            else []
-        )
-
-    # Try to summarize old messages, fall back to truncation on failure
-    summary_msg = None
-    if old_messages_dict:
-        try:
-            summary_text = await _summarize_messages(
-                old_messages_dict, model=model, api_key=api_key, base_url=base_url
-            )
-            summary_msg = ChatCompletionAssistantMessageParam(
-                role="assistant",
-                content=f"[Previous conversation summary — for context only]: {summary_text}",
-            )
-            base = [system_msg, summary_msg] if has_system_prompt else [summary_msg]
-            messages = base + recent_messages
-            logger.info(
-                f"Context summarized: {token_count} tokens, "
-                f"summarized {len(old_messages_dict)} msgs, kept {KEEP_RECENT_MESSAGES}"
-            )
-        except Exception as e:
-            logger.warning(f"Summarization failed, falling back to truncation: {e}")
-            messages = (
-                [system_msg] + recent_messages if has_system_prompt else recent_messages
-            )
-    else:
-        logger.warning(
-            f"Token count {token_count} exceeds threshold but no old messages to summarize"
-        )
-
-    new_token_count = estimate_token_count(
-        _messages_to_dicts(messages), model=token_count_model
-    )
-
-    # Progressive truncation if still over limit
-    if new_token_count > TOKEN_THRESHOLD:
-        logger.warning(
-            f"Still over limit: {new_token_count} tokens. Reducing messages."
-        )
-        base_msgs = (
-            recent_messages
-            if old_messages_dict
-            else (messages_dict[1:] if has_system_prompt else messages_dict)
-        )
-
-        def build_messages(recent: list) -> list:
-            """Build message list with optional system prompt and summary."""
-            prefix = []
-            if has_system_prompt and system_msg:
-                prefix.append(system_msg)
-            if summary_msg:
-                prefix.append(summary_msg)
-            return prefix + recent
-
-        for keep_count in [12, 10, 8, 5, 3, 2, 1, 0]:
-            if keep_count == 0:
-                messages = build_messages([])
-                if not messages:
-                    continue
-            elif len(base_msgs) < keep_count:
-                continue
-            else:
-                reduced = _ensure_tool_pairs_intact(
-                    base_msgs[-keep_count:],
-                    base_msgs,
-                    max(0, len(base_msgs) - keep_count),
-                )
-                messages = build_messages(reduced)
-
-            new_token_count = estimate_token_count(
-                _messages_to_dicts(messages), model=token_count_model
-            )
-            if new_token_count <= TOKEN_THRESHOLD:
-                logger.info(
-                    f"Reduced to {keep_count} messages, {new_token_count} tokens"
-                )
-                break
-        else:
-            logger.error(
-                f"Cannot reduce below threshold. Final: {new_token_count} tokens"
-            )
-            if has_system_prompt and len(messages) > 1:
-                messages = messages[1:]
-                logger.critical("Dropped system prompt as last resort")
-                return ContextWindowResult(
-                    messages, new_token_count, True, "System prompt dropped"
-                )
-            # No system prompt to drop - return error so callers don't proceed with oversized context
-            return ContextWindowResult(
-                messages,
-                new_token_count,
-                True,
-                "Unable to reduce context below token limit",
-            )
-
-    return ContextWindowResult(messages, new_token_count, True)
-
-
-async def _summarize_messages(
-    messages: list,
-    model: str,
-    api_key: str | None = None,
-    base_url: str | None = None,
-    timeout: float = 30.0,
-) -> str:
-    """Summarize a list of messages into concise context.
-
-    Uses the same model as the chat for higher quality summaries.
-
-    Args:
-        messages: List of message dicts to summarize
-        model: Model to use for summarization (same as chat model)
-        api_key: API key for OpenAI client
-        base_url: Base URL for OpenAI client
-        timeout: Request timeout in seconds (default: 30.0)
-
-    Returns:
-        Summarized text
-    """
-    # Format messages for summarization
-    conversation = []
-    for msg in messages:
-        role = msg.get("role", "")
-        content = msg.get("content", "")
-        # Include user, assistant, and tool messages (tool outputs are important context)
-        if content and role in ("user", "assistant", "tool"):
-            conversation.append(f"{role.upper()}: {content}")
-
-    conversation_text = "\n\n".join(conversation)
-
-    # Handle empty conversation
-    if not conversation_text:
-        return "No conversation history available."
-
-    # Truncate conversation to fit within summarization model's context
-    # gpt-4o-mini has 128k context, but we limit to ~25k tokens (~100k chars) for safety
-    MAX_CHARS = 100_000
-    if len(conversation_text) > MAX_CHARS:
-        conversation_text = conversation_text[:MAX_CHARS] + "\n\n[truncated]"
-
-    # Call LLM to summarize
     import openai
 
-    summarization_client = openai.AsyncOpenAI(
-        api_key=api_key, base_url=base_url, timeout=timeout
-    )
+    from backend.util.prompt import compress_context
 
-    response = await summarization_client.chat.completions.create(
-        model=model,
-        messages=[
-            {
-                "role": "system",
-                "content": (
-                    "Create a detailed summary of the conversation so far. "
-                    "This summary will be used as context when continuing the conversation.\n\n"
-                    "Before writing the summary, analyze each message chronologically to identify:\n"
-                    "- User requests and their explicit goals\n"
-                    "- Your approach and key decisions made\n"
-                    "- Technical specifics (file names, tool outputs, function signatures)\n"
-                    "- Errors encountered and resolutions applied\n\n"
-                    "You MUST include ALL of the following sections:\n\n"
-                    "## 1. Primary Request and Intent\n"
-                    "The user's explicit goals and what they are trying to accomplish.\n\n"
-                    "## 2. Key Technical Concepts\n"
-                    "Technologies, frameworks, tools, and patterns being used or discussed.\n\n"
-                    "## 3. Files and Resources Involved\n"
-                    "Specific files examined or modified, with relevant snippets and identifiers.\n\n"
-                    "## 4. Errors and Fixes\n"
-                    "Problems encountered, error messages, and their resolutions. "
-                    "Include any user feedback on fixes.\n\n"
-                    "## 5. Problem Solving\n"
-                    "Issues that have been resolved and how they were addressed.\n\n"
-                    "## 6. All User Messages\n"
-                    "A complete list of all user inputs (excluding tool outputs) to preserve their exact requests.\n\n"
-                    "## 7. Pending Tasks\n"
-                    "Work items the user explicitly requested that have not yet been completed.\n\n"
-                    "## 8. Current Work\n"
-                    "Precise description of what was being worked on most recently, including relevant context.\n\n"
-                    "## 9. Next Steps\n"
-                    "What should happen next, aligned with the user's most recent requests. "
-                    "Include verbatim quotes of recent instructions if relevant."
-                ),
-            },
-            {"role": "user", "content": f"Summarize:\n\n{conversation_text}"},
-        ],
-        max_tokens=1500,
-        temperature=0.3,
-    )
+    # Convert messages to dict format
+    messages_dict = []
+    for msg in messages:
+        if isinstance(msg, dict):
+            msg_dict = {k: v for k, v in msg.items() if v is not None}
+        else:
+            msg_dict = dict(msg)
+        messages_dict.append(msg_dict)
 
-    summary = response.choices[0].message.content
-    return summary or "No summary available."
-
-
-def _ensure_tool_pairs_intact(
-    recent_messages: list[dict],
-    all_messages: list[dict],
-    start_index: int,
-) -> list[dict]:
-    """
-    Ensure tool_call/tool_response pairs stay together after slicing.
-
-    When slicing messages for context compaction, a naive slice can separate
-    an assistant message containing tool_calls from its corresponding tool
-    response messages. This causes API validation errors (e.g., Anthropic's
-    "unexpected tool_use_id found in tool_result blocks").
-
-    This function checks for orphan tool responses in the slice and extends
-    backwards to include their corresponding assistant messages.
-
-    Args:
-        recent_messages: The sliced messages to validate
-        all_messages: The complete message list (for looking up missing assistants)
-        start_index: The index in all_messages where recent_messages begins
-
-    Returns:
-        A potentially extended list of messages with tool pairs intact
-    """
-    if not recent_messages:
-        return recent_messages
-
-    # Collect all tool_call_ids from assistant messages in the slice
-    available_tool_call_ids: set[str] = set()
-    for msg in recent_messages:
-        if msg.get("role") == "assistant" and msg.get("tool_calls"):
-            for tc in msg["tool_calls"]:
-                tc_id = tc.get("id")
-                if tc_id:
-                    available_tool_call_ids.add(tc_id)
-
-    # Find orphan tool responses (tool messages whose tool_call_id is missing)
-    orphan_tool_call_ids: set[str] = set()
-    for msg in recent_messages:
-        if msg.get("role") == "tool":
-            tc_id = msg.get("tool_call_id")
-            if tc_id and tc_id not in available_tool_call_ids:
-                orphan_tool_call_ids.add(tc_id)
-
-    if not orphan_tool_call_ids:
-        # No orphans, slice is valid
-        return recent_messages
-
-    # Find the assistant messages that contain the orphan tool_call_ids
-    # Search backwards from start_index in all_messages
-    messages_to_prepend: list[dict] = []
-    for i in range(start_index - 1, -1, -1):
-        msg = all_messages[i]
-        if msg.get("role") == "assistant" and msg.get("tool_calls"):
-            msg_tool_ids = {tc.get("id") for tc in msg["tool_calls"] if tc.get("id")}
-            if msg_tool_ids & orphan_tool_call_ids:
-                # This assistant message has tool_calls we need
-                # Also collect its contiguous tool responses that follow it
-                assistant_and_responses: list[dict] = [msg]
-
-                # Scan forward from this assistant to collect tool responses
-                for j in range(i + 1, start_index):
-                    following_msg = all_messages[j]
-                    if following_msg.get("role") == "tool":
-                        tool_id = following_msg.get("tool_call_id")
-                        if tool_id and tool_id in msg_tool_ids:
-                            assistant_and_responses.append(following_msg)
-                    else:
-                        # Stop at first non-tool message
-                        break
-
-                # Prepend the assistant and its tool responses (maintain order)
-                messages_to_prepend = assistant_and_responses + messages_to_prepend
-                # Mark these as found
-                orphan_tool_call_ids -= msg_tool_ids
-                # Also add this assistant's tool_call_ids to available set
-                available_tool_call_ids |= msg_tool_ids
-
-        if not orphan_tool_call_ids:
-            # Found all missing assistants
-            break
-
-    if orphan_tool_call_ids:
-        # Some tool_call_ids couldn't be resolved - remove those tool responses
-        # This shouldn't happen in normal operation but handles edge cases
-        logger.warning(
-            f"Could not find assistant messages for tool_call_ids: {orphan_tool_call_ids}. "
-            "Removing orphan tool responses."
-        )
-        recent_messages = [
-            msg
-            for msg in recent_messages
-            if not (
-                msg.get("role") == "tool"
-                and msg.get("tool_call_id") in orphan_tool_call_ids
+    # Only create client if api_key is provided (enables summarization)
+    # Use context manager to avoid socket leaks
+    if api_key:
+        async with openai.AsyncOpenAI(
+            api_key=api_key, base_url=base_url, timeout=30.0
+        ) as client:
+            return await compress_context(
+                messages=messages_dict,
+                model=model,
+                client=client,
             )
-        ]
-
-    if messages_to_prepend:
-        logger.info(
-            f"Extended recent messages by {len(messages_to_prepend)} to preserve "
-            f"tool_call/tool_response pairs"
+    else:
+        # No API key - use truncation-only mode
+        return await compress_context(
+            messages=messages_dict,
+            model=model,
+            client=None,
         )
-        return messages_to_prepend + recent_messages
-
-    return recent_messages
 
 
 async def _stream_chat_chunks(

autogpt_platform/backend/backend/api/features/chat/tools/models.py

@@ -38,6 +38,8 @@ class ResponseType(str, Enum):
     OPERATION_STARTED = "operation_started"
     OPERATION_PENDING = "operation_pending"
     OPERATION_IN_PROGRESS = "operation_in_progress"
+    # Input validation
+    INPUT_VALIDATION_ERROR = "input_validation_error"
 
 
 # Base response model
@@ -68,6 +70,10 @@ class AgentInfo(BaseModel):
     has_external_trigger: bool | None = None
     new_output: bool | None = None
     graph_id: str | None = None
+    inputs: dict[str, Any] | None = Field(
+        default=None,
+        description="Input schema for the agent, including field names, types, and defaults",
+    )
 
 
 class AgentsFoundResponse(ToolResponseBase):
@@ -194,6 +200,20 @@ class ErrorResponse(ToolResponseBase):
     details: dict[str, Any] | None = None
 
 
+class InputValidationErrorResponse(ToolResponseBase):
+    """Response when run_agent receives unknown input fields."""
+
+    type: ResponseType = ResponseType.INPUT_VALIDATION_ERROR
+    unrecognized_fields: list[str] = Field(
+        description="List of input field names that were not recognized"
+    )
+    inputs: dict[str, Any] = Field(
+        description="The agent's valid input schema for reference"
+    )
+    graph_id: str | None = None
+    graph_version: int | None = None
+
+
 # Agent output models
 class ExecutionOutputInfo(BaseModel):
     """Summary of a single execution's outputs."""

autogpt_platform/backend/backend/api/features/chat/tools/run_agent.py

@@ -30,6 +30,7 @@ from .models import (
     ErrorResponse,
     ExecutionOptions,
     ExecutionStartedResponse,
+    InputValidationErrorResponse,
     SetupInfo,
     SetupRequirementsResponse,
     ToolResponseBase,
@@ -273,6 +274,22 @@ class RunAgentTool(BaseTool):
             input_properties = graph.input_schema.get("properties", {})
             required_fields = set(graph.input_schema.get("required", []))
             provided_inputs = set(params.inputs.keys())
+            valid_fields = set(input_properties.keys())
+
+            # Check for unknown input fields
+            unrecognized_fields = provided_inputs - valid_fields
+            if unrecognized_fields:
+                return InputValidationErrorResponse(
+                    message=(
+                        f"Unknown input field(s) provided: {', '.join(sorted(unrecognized_fields))}. "
+                        f"Agent was not executed. Please use the correct field names from the schema."
+                    ),
+                    session_id=session_id,
+                    unrecognized_fields=sorted(unrecognized_fields),
+                    inputs=graph.input_schema,
+                    graph_id=graph.id,
+                    graph_version=graph.version,
+                )
 
             # If agent has inputs but none were provided AND use_defaults is not set,
             # always show what's available first so user can decide

autogpt_platform/backend/backend/api/features/chat/tools/run_agent_test.py

@@ -402,3 +402,42 @@ async def test_run_agent_schedule_without_name(setup_test_data):
     # Should return error about missing schedule_name
     assert result_data.get("type") == "error"
     assert "schedule_name" in result_data["message"].lower()
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_run_agent_rejects_unknown_input_fields(setup_test_data):
+    """Test that run_agent returns input_validation_error for unknown input fields."""
+    user = setup_test_data["user"]
+    store_submission = setup_test_data["store_submission"]
+
+    tool = RunAgentTool()
+    agent_marketplace_id = f"{user.email.split('@')[0]}/{store_submission.slug}"
+    session = make_session(user_id=user.id)
+
+    # Execute with unknown input field names
+    response = await tool.execute(
+        user_id=user.id,
+        session_id=str(uuid.uuid4()),
+        tool_call_id=str(uuid.uuid4()),
+        username_agent_slug=agent_marketplace_id,
+        inputs={
+            "unknown_field": "some value",
+            "another_unknown": "another value",
+        },
+        session=session,
+    )
+
+    assert response is not None
+    assert hasattr(response, "output")
+    assert isinstance(response.output, str)
+    result_data = orjson.loads(response.output)
+
+    # Should return input_validation_error type with unrecognized fields
+    assert result_data.get("type") == "input_validation_error"
+    assert "unrecognized_fields" in result_data
+    assert set(result_data["unrecognized_fields"]) == {
+        "another_unknown",
+        "unknown_field",
+    }
+    assert "inputs" in result_data  # Contains the valid schema
+    assert "Agent was not executed" in result_data["message"]

autogpt_platform/backend/backend/api/features/chat/tools/run_block.py

@@ -5,6 +5,8 @@ import uuid
 from collections import defaultdict
 from typing import Any
 
+from pydantic_core import PydanticUndefined
+
 from backend.api.features.chat.model import ChatSession
 from backend.data.block import get_block
 from backend.data.execution import ExecutionContext
@@ -75,15 +77,22 @@ class RunBlockTool(BaseTool):
         self,
         user_id: str,
         block: Any,
+        input_data: dict[str, Any] | None = None,
     ) -> tuple[dict[str, CredentialsMetaInput], list[CredentialsMetaInput]]:
         """
         Check if user has required credentials for a block.
 
+        Args:
+            user_id: User ID
+            block: Block to check credentials for
+            input_data: Input data for the block (used to determine provider via discriminator)
+
         Returns:
             tuple[matched_credentials, missing_credentials]
         """
         matched_credentials: dict[str, CredentialsMetaInput] = {}
         missing_credentials: list[CredentialsMetaInput] = []
+        input_data = input_data or {}
 
         # Get credential field info from block's input schema
         credentials_fields_info = block.input_schema.get_credentials_fields_info()
@@ -96,14 +105,33 @@ class RunBlockTool(BaseTool):
         available_creds = await creds_manager.store.get_all_creds(user_id)
 
         for field_name, field_info in credentials_fields_info.items():
-            # field_info.provider is a frozenset of acceptable providers
-            # field_info.supported_types is a frozenset of acceptable types
+            effective_field_info = field_info
+            if field_info.discriminator and field_info.discriminator_mapping:
+                # Get discriminator from input, falling back to schema default
+                discriminator_value = input_data.get(field_info.discriminator)
+                if discriminator_value is None:
+                    field = block.input_schema.model_fields.get(
+                        field_info.discriminator
+                    )
+                    if field and field.default is not PydanticUndefined:
+                        discriminator_value = field.default
+
+                if (
+                    discriminator_value
+                    and discriminator_value in field_info.discriminator_mapping
+                ):
+                    effective_field_info = field_info.discriminate(discriminator_value)
+                    logger.debug(
+                        f"Discriminated provider for {field_name}: "
+                        f"{discriminator_value} -> {effective_field_info.provider}"
+                    )
+
             matching_cred = next(
                 (
                     cred
                     for cred in available_creds
-                    if cred.provider in field_info.provider
-                    and cred.type in field_info.supported_types
+                    if cred.provider in effective_field_info.provider
+                    and cred.type in effective_field_info.supported_types
                 ),
                 None,
             )
@@ -117,8 +145,8 @@ class RunBlockTool(BaseTool):
                 )
             else:
                 # Create a placeholder for the missing credential
-                provider = next(iter(field_info.provider), "unknown")
-                cred_type = next(iter(field_info.supported_types), "api_key")
+                provider = next(iter(effective_field_info.provider), "unknown")
+                cred_type = next(iter(effective_field_info.supported_types), "api_key")
                 missing_credentials.append(
                     CredentialsMetaInput(
                         id=field_name,
@@ -186,10 +214,9 @@ class RunBlockTool(BaseTool):
 
         logger.info(f"Executing block {block.name} ({block_id}) for user {user_id}")
 
-        # Check credentials
         creds_manager = IntegrationCredentialsManager()
         matched_credentials, missing_credentials = await self._check_block_credentials(
-            user_id, block
+            user_id, block, input_data
         )
 
         if missing_credentials:

autogpt_platform/backend/backend/blocks/llm.py

@@ -32,7 +32,7 @@ from backend.data.model import (
 from backend.integrations.providers import ProviderName
 from backend.util import json
 from backend.util.logging import TruncatedLogger
-from backend.util.prompt import compress_prompt, estimate_token_count
+from backend.util.prompt import compress_context, estimate_token_count
 from backend.util.text import TextFormatter
 
 logger = TruncatedLogger(logging.getLogger(__name__), "[LLM-Block]")
@@ -634,11 +634,18 @@ async def llm_call(
     context_window = llm_model.context_window
 
     if compress_prompt_to_fit:
-        prompt = compress_prompt(
+        result = await compress_context(
             messages=prompt,
             target_tokens=llm_model.context_window // 2,
-            lossy_ok=True,
+            client=None,  # Truncation-only, no LLM summarization
+            reserve=0,  # Caller handles response token budget separately
         )
+        if result.error:
+            logger.warning(
+                f"Prompt compression did not meet target: {result.error}. "
+                f"Proceeding with {result.token_count} tokens."
+            )
+        prompt = result.messages
 
     # Calculate available tokens based on context window and input length
     estimated_input_tokens = estimate_token_count(prompt)

autogpt_platform/backend/backend/executor/database.py

@@ -17,6 +17,7 @@ from backend.data.analytics import (
     get_accuracy_trends_and_alerts,
     get_marketplace_graphs_for_monitoring,
 )
+from backend.data.auth.oauth import cleanup_expired_oauth_tokens
 from backend.data.credit import UsageTransactionMetadata, get_user_credit_model
 from backend.data.execution import (
     create_graph_execution,
@@ -219,6 +220,9 @@ class DatabaseManager(AppService):
     # Onboarding
     increment_onboarding_runs = _(increment_onboarding_runs)
 
+    # OAuth
+    cleanup_expired_oauth_tokens = _(cleanup_expired_oauth_tokens)
+
     # Store
     get_store_agents = _(get_store_agents)
     get_store_agent_details = _(get_store_agent_details)
@@ -349,6 +353,9 @@ class DatabaseManagerAsyncClient(AppServiceClient):
     # Onboarding
     increment_onboarding_runs = d.increment_onboarding_runs
 
+    # OAuth
+    cleanup_expired_oauth_tokens = d.cleanup_expired_oauth_tokens
+
     # Store
     get_store_agents = d.get_store_agents
     get_store_agent_details = d.get_store_agent_details

autogpt_platform/backend/backend/executor/scheduler.py

@@ -24,11 +24,9 @@ from dotenv import load_dotenv
 from pydantic import BaseModel, Field, ValidationError
 from sqlalchemy import MetaData, create_engine
 
-from backend.data.auth.oauth import cleanup_expired_oauth_tokens
 from backend.data.block import BlockInput
 from backend.data.execution import GraphExecutionWithNodes
 from backend.data.model import CredentialsMetaInput
-from backend.data.onboarding import increment_onboarding_runs
 from backend.executor import utils as execution_utils
 from backend.monitoring import (
     NotificationJobArgs,
@@ -38,7 +36,11 @@ from backend.monitoring import (
     report_execution_accuracy_alerts,
     report_late_executions,
 )
-from backend.util.clients import get_database_manager_client, get_scheduler_client
+from backend.util.clients import (
+    get_database_manager_async_client,
+    get_database_manager_client,
+    get_scheduler_client,
+)
 from backend.util.cloud_storage import cleanup_expired_files_async
 from backend.util.exceptions import (
     GraphNotFoundError,
@@ -148,6 +150,7 @@ def execute_graph(**kwargs):
 async def _execute_graph(**kwargs):
     args = GraphExecutionJobArgs(**kwargs)
     start_time = asyncio.get_event_loop().time()
+    db = get_database_manager_async_client()
     try:
         logger.info(f"Executing recurring job for graph #{args.graph_id}")
         graph_exec: GraphExecutionWithNodes = await execution_utils.add_graph_execution(
@@ -157,7 +160,7 @@ async def _execute_graph(**kwargs):
             inputs=args.input_data,
             graph_credentials_inputs=args.input_credentials,
         )
-        await increment_onboarding_runs(args.user_id)
+        await db.increment_onboarding_runs(args.user_id)
         elapsed = asyncio.get_event_loop().time() - start_time
         logger.info(
             f"Graph execution started with ID {graph_exec.id} for graph {args.graph_id} "
@@ -246,8 +249,13 @@ def cleanup_expired_files():
 
 def cleanup_oauth_tokens():
     """Clean up expired OAuth tokens from the database."""
+
     # Wait for completion
-    run_async(cleanup_expired_oauth_tokens())
+    async def _cleanup():
+        db = get_database_manager_async_client()
+        return await db.cleanup_expired_oauth_tokens()
+
+    run_async(_cleanup())
 
 
 def execution_accuracy_alerts():

autogpt_platform/backend/backend/util/prompt.py

@@ -1,10 +1,19 @@
+from __future__ import annotations
+
+import logging
 from copy import deepcopy
-from typing import Any
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any
 
 from tiktoken import encoding_for_model
 
 from backend.util import json
 
+if TYPE_CHECKING:
+    from openai import AsyncOpenAI
+
+logger = logging.getLogger(__name__)
+
 # ---------------------------------------------------------------------------#
 #  CONSTANTS                                                                 #
 # ---------------------------------------------------------------------------#
@@ -100,9 +109,17 @@ def _is_objective_message(msg: dict) -> bool:
 def _truncate_tool_message_content(msg: dict, enc, max_tokens: int) -> None:
     """
     Carefully truncate tool message content while preserving tool structure.
-    Only truncates tool_result content, leaves tool_use intact.
+    Handles both Anthropic-style (list content) and OpenAI-style (string content) tool messages.
     """
     content = msg.get("content")
+
+    # OpenAI-style tool message: role="tool" with string content
+    if msg.get("role") == "tool" and isinstance(content, str):
+        if _tok_len(content, enc) > max_tokens:
+            msg["content"] = _truncate_middle_tokens(content, enc, max_tokens)
+        return
+
+    # Anthropic-style: list content with tool_result items
     if not isinstance(content, list):
         return
 
@@ -140,141 +157,6 @@ def _truncate_middle_tokens(text: str, enc, max_tok: int) -> str:
 # ---------------------------------------------------------------------------#
 
 
-def compress_prompt(
-    messages: list[dict],
-    target_tokens: int,
-    *,
-    model: str = "gpt-4o",
-    reserve: int = 2_048,
-    start_cap: int = 8_192,
-    floor_cap: int = 128,
-    lossy_ok: bool = True,
-) -> list[dict]:
-    """
-    Shrink *messages* so that::
-
-        token_count(prompt) + reserve  ≤  target_tokens
-
-    Strategy
-    --------
-    1. **Token-aware truncation** – progressively halve a per-message cap
-       (`start_cap`, `start_cap/2`, … `floor_cap`) and apply it to the
-       *content* of every message except the first and last.  Tool shells
-       are included: we keep the envelope but shorten huge payloads.
-    2. **Middle-out deletion** – if still over the limit, delete whole
-       messages working outward from the centre, **skipping** any message
-       that contains ``tool_calls`` or has ``role == "tool"``.
-    3. **Last-chance trim** – if still too big, truncate the *first* and
-       *last* message bodies down to `floor_cap` tokens.
-    4. If the prompt is *still* too large:
-         • raise ``ValueError``      when ``lossy_ok == False`` (default)
-         • return the partially-trimmed prompt when ``lossy_ok == True``
-
-    Parameters
-    ----------
-    messages        Complete chat history (will be deep-copied).
-    model           Model name; passed to tiktoken to pick the right
-                    tokenizer (gpt-4o → 'o200k_base', others fallback).
-    target_tokens   Hard ceiling for prompt size **excluding** the model's
-                    forthcoming answer.
-    reserve         How many tokens you want to leave available for that
-                    answer (`max_tokens` in your subsequent completion call).
-    start_cap       Initial per-message truncation ceiling (tokens).
-    floor_cap       Lowest cap we'll accept before moving to deletions.
-    lossy_ok        If *True* return best-effort prompt instead of raising
-                    after all trim passes have been exhausted.
-
-    Returns
-    -------
-    list[dict]  – A *new* messages list that abides by the rules above.
-    """
-    enc = encoding_for_model(model)  # best-match tokenizer
-    msgs = deepcopy(messages)  # never mutate caller
-
-    def total_tokens() -> int:
-        """Current size of *msgs* in tokens."""
-        return sum(_msg_tokens(m, enc) for m in msgs)
-
-    original_token_count = total_tokens()
-
-    if original_token_count + reserve <= target_tokens:
-        return msgs
-
-    # ---- STEP 0 : normalise content --------------------------------------
-    # Convert non-string payloads to strings so token counting is coherent.
-    for i, m in enumerate(msgs):
-        if not isinstance(m.get("content"), str) and m.get("content") is not None:
-            if _is_tool_message(m):
-                continue
-
-            # Keep first and last messages intact (unless they're tool messages)
-            if i == 0 or i == len(msgs) - 1:
-                continue
-
-            # Reasonable 20k-char ceiling prevents pathological blobs
-            content_str = json.dumps(m["content"], separators=(",", ":"))
-            if len(content_str) > 20_000:
-                content_str = _truncate_middle_tokens(content_str, enc, 20_000)
-            m["content"] = content_str
-
-    # ---- STEP 1 : token-aware truncation ---------------------------------
-    cap = start_cap
-    while total_tokens() + reserve > target_tokens and cap >= floor_cap:
-        for m in msgs[1:-1]:  # keep first & last intact
-            if _is_tool_message(m):
-                # For tool messages, only truncate tool result content, preserve structure
-                _truncate_tool_message_content(m, enc, cap)
-                continue
-
-            if _is_objective_message(m):
-                # Never truncate objective messages - they contain the core task
-                continue
-
-            content = m.get("content") or ""
-            if _tok_len(content, enc) > cap:
-                m["content"] = _truncate_middle_tokens(content, enc, cap)
-        cap //= 2  # tighten the screw
-
-    # ---- STEP 2 : middle-out deletion -----------------------------------
-    while total_tokens() + reserve > target_tokens and len(msgs) > 2:
-        # Identify all deletable messages (not first/last, not tool messages, not objective messages)
-        deletable_indices = []
-        for i in range(1, len(msgs) - 1):  # Skip first and last
-            if not _is_tool_message(msgs[i]) and not _is_objective_message(msgs[i]):
-                deletable_indices.append(i)
-
-        if not deletable_indices:
-            break  # nothing more we can drop
-
-        # Delete from center outward - find the index closest to center
-        centre = len(msgs) // 2
-        to_delete = min(deletable_indices, key=lambda i: abs(i - centre))
-        del msgs[to_delete]
-
-    # ---- STEP 3 : final safety-net trim on first & last ------------------
-    cap = start_cap
-    while total_tokens() + reserve > target_tokens and cap >= floor_cap:
-        for idx in (0, -1):  # first and last
-            if _is_tool_message(msgs[idx]):
-                # For tool messages at first/last position, truncate tool result content only
-                _truncate_tool_message_content(msgs[idx], enc, cap)
-                continue
-
-            text = msgs[idx].get("content") or ""
-            if _tok_len(text, enc) > cap:
-                msgs[idx]["content"] = _truncate_middle_tokens(text, enc, cap)
-        cap //= 2  # tighten the screw
-
-    # ---- STEP 4 : success or fail-gracefully -----------------------------
-    if total_tokens() + reserve > target_tokens and not lossy_ok:
-        raise ValueError(
-            "compress_prompt: prompt still exceeds budget "
-            f"({total_tokens() + reserve} > {target_tokens})."
-        )
-
-    return msgs
-
-
 def estimate_token_count(
     messages: list[dict],
     *,
@@ -293,7 +175,8 @@ def estimate_token_count(
     -------
     int  – Token count.
     """
-    enc = encoding_for_model(model)  # best-match tokenizer
+    token_model = _normalize_model_for_tokenizer(model)
+    enc = encoding_for_model(token_model)
     return sum(_msg_tokens(m, enc) for m in messages)
 
 
@@ -315,6 +198,543 @@ def estimate_token_count_str(
     -------
     int  – Token count.
     """
-    enc = encoding_for_model(model)  # best-match tokenizer
+    token_model = _normalize_model_for_tokenizer(model)
+    enc = encoding_for_model(token_model)
     text = json.dumps(text) if not isinstance(text, str) else text
     return _tok_len(text, enc)
+
+
+# ---------------------------------------------------------------------------#
+#  UNIFIED CONTEXT COMPRESSION                                               #
+# ---------------------------------------------------------------------------#
+
+# Default thresholds
+DEFAULT_TOKEN_THRESHOLD = 120_000
+DEFAULT_KEEP_RECENT = 15
+
+
+@dataclass
+class CompressResult:
+    """Result of context compression."""
+
+    messages: list[dict]
+    token_count: int
+    was_compacted: bool
+    error: str | None = None
+    original_token_count: int = 0
+    messages_summarized: int = 0
+    messages_dropped: int = 0
+
+
+def _normalize_model_for_tokenizer(model: str) -> str:
+    """Normalize model name for tiktoken tokenizer selection."""
+    if "/" in model:
+        model = model.split("/")[-1]
+    if "claude" in model.lower() or not any(
+        known in model.lower() for known in ["gpt", "o1", "chatgpt", "text-"]
+    ):
+        return "gpt-4o"
+    return model
+
+
+def _extract_tool_call_ids_from_message(msg: dict) -> set[str]:
+    """
+    Extract tool_call IDs from an assistant message.
+
+    Supports both formats:
+    - OpenAI: {"role": "assistant", "tool_calls": [{"id": "..."}]}
+    - Anthropic: {"role": "assistant", "content": [{"type": "tool_use", "id": "..."}]}
+
+    Returns:
+        Set of tool_call IDs found in the message.
+    """
+    ids: set[str] = set()
+    if msg.get("role") != "assistant":
+        return ids
+
+    # OpenAI format: tool_calls array
+    if msg.get("tool_calls"):
+        for tc in msg["tool_calls"]:
+            tc_id = tc.get("id")
+            if tc_id:
+                ids.add(tc_id)
+
+    # Anthropic format: content list with tool_use blocks
+    content = msg.get("content")
+    if isinstance(content, list):
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "tool_use":
+                tc_id = block.get("id")
+                if tc_id:
+                    ids.add(tc_id)
+
+    return ids
+
+
+def _extract_tool_response_ids_from_message(msg: dict) -> set[str]:
+    """
+    Extract tool_call IDs that this message is responding to.
+
+    Supports both formats:
+    - OpenAI: {"role": "tool", "tool_call_id": "..."}
+    - Anthropic: {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "..."}]}
+
+    Returns:
+        Set of tool_call IDs this message responds to.
+    """
+    ids: set[str] = set()
+
+    # OpenAI format: role=tool with tool_call_id
+    if msg.get("role") == "tool":
+        tc_id = msg.get("tool_call_id")
+        if tc_id:
+            ids.add(tc_id)
+
+    # Anthropic format: content list with tool_result blocks
+    content = msg.get("content")
+    if isinstance(content, list):
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "tool_result":
+                tc_id = block.get("tool_use_id")
+                if tc_id:
+                    ids.add(tc_id)
+
+    return ids
+
+
+def _is_tool_response_message(msg: dict) -> bool:
+    """Check if message is a tool response (OpenAI or Anthropic format)."""
+    # OpenAI format
+    if msg.get("role") == "tool":
+        return True
+    # Anthropic format
+    content = msg.get("content")
+    if isinstance(content, list):
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "tool_result":
+                return True
+    return False
+
+
+def _remove_orphan_tool_responses(
+    messages: list[dict], orphan_ids: set[str]
+) -> list[dict]:
+    """
+    Remove tool response messages/blocks that reference orphan tool_call IDs.
+
+    Supports both OpenAI and Anthropic formats.
+    For Anthropic messages with mixed valid/orphan tool_result blocks,
+    filters out only the orphan blocks instead of dropping the entire message.
+    """
+    result = []
+    for msg in messages:
+        # OpenAI format: role=tool - drop entire message if orphan
+        if msg.get("role") == "tool":
+            tc_id = msg.get("tool_call_id")
+            if tc_id and tc_id in orphan_ids:
+                continue
+            result.append(msg)
+            continue
+
+        # Anthropic format: content list may have mixed tool_result blocks
+        content = msg.get("content")
+        if isinstance(content, list):
+            has_tool_results = any(
+                isinstance(b, dict) and b.get("type") == "tool_result" for b in content
+            )
+            if has_tool_results:
+                # Filter out orphan tool_result blocks, keep valid ones
+                filtered_content = [
+                    block
+                    for block in content
+                    if not (
+                        isinstance(block, dict)
+                        and block.get("type") == "tool_result"
+                        and block.get("tool_use_id") in orphan_ids
+                    )
+                ]
+                # Only keep message if it has remaining content
+                if filtered_content:
+                    msg = msg.copy()
+                    msg["content"] = filtered_content
+                    result.append(msg)
+                continue
+
+        result.append(msg)
+    return result
+
+
+def _ensure_tool_pairs_intact(
+    recent_messages: list[dict],
+    all_messages: list[dict],
+    start_index: int,
+) -> list[dict]:
+    """
+    Ensure tool_call/tool_response pairs stay together after slicing.
+
+    When slicing messages for context compaction, a naive slice can separate
+    an assistant message containing tool_calls from its corresponding tool
+    response messages. This causes API validation errors (e.g., Anthropic's
+    "unexpected tool_use_id found in tool_result blocks").
+
+    This function checks for orphan tool responses in the slice and extends
+    backwards to include their corresponding assistant messages.
+
+    Supports both formats:
+    - OpenAI: tool_calls array + role="tool" responses
+    - Anthropic: tool_use blocks + tool_result blocks
+
+    Args:
+        recent_messages: The sliced messages to validate
+        all_messages: The complete message list (for looking up missing assistants)
+        start_index: The index in all_messages where recent_messages begins
+
+    Returns:
+        A potentially extended list of messages with tool pairs intact
+    """
+    if not recent_messages:
+        return recent_messages
+
+    # Collect all tool_call_ids from assistant messages in the slice
+    available_tool_call_ids: set[str] = set()
+    for msg in recent_messages:
+        available_tool_call_ids |= _extract_tool_call_ids_from_message(msg)
+
+    # Find orphan tool responses (responses whose tool_call_id is missing)
+    orphan_tool_call_ids: set[str] = set()
+    for msg in recent_messages:
+        response_ids = _extract_tool_response_ids_from_message(msg)
+        for tc_id in response_ids:
+            if tc_id not in available_tool_call_ids:
+                orphan_tool_call_ids.add(tc_id)
+
+    if not orphan_tool_call_ids:
+        # No orphans, slice is valid
+        return recent_messages
+
+    # Find the assistant messages that contain the orphan tool_call_ids
+    # Search backwards from start_index in all_messages
+    messages_to_prepend: list[dict] = []
+    for i in range(start_index - 1, -1, -1):
+        msg = all_messages[i]
+        msg_tool_ids = _extract_tool_call_ids_from_message(msg)
+        if msg_tool_ids & orphan_tool_call_ids:
+            # This assistant message has tool_calls we need
+            # Also collect its contiguous tool responses that follow it
+            assistant_and_responses: list[dict] = [msg]
+
+            # Scan forward from this assistant to collect tool responses
+            for j in range(i + 1, start_index):
+                following_msg = all_messages[j]
+                following_response_ids = _extract_tool_response_ids_from_message(
+                    following_msg
+                )
+                if following_response_ids and following_response_ids & msg_tool_ids:
+                    assistant_and_responses.append(following_msg)
+                elif not _is_tool_response_message(following_msg):
+                    # Stop at first non-tool-response message
+                    break
+
+            # Prepend the assistant and its tool responses (maintain order)
+            messages_to_prepend = assistant_and_responses + messages_to_prepend
+            # Mark these as found
+            orphan_tool_call_ids -= msg_tool_ids
+            # Also add this assistant's tool_call_ids to available set
+            available_tool_call_ids |= msg_tool_ids
+
+        if not orphan_tool_call_ids:
+            # Found all missing assistants
+            break
+
+    if orphan_tool_call_ids:
+        # Some tool_call_ids couldn't be resolved - remove those tool responses
+        # This shouldn't happen in normal operation but handles edge cases
+        logger.warning(
+            f"Could not find assistant messages for tool_call_ids: {orphan_tool_call_ids}. "
+            "Removing orphan tool responses."
+        )
+        recent_messages = _remove_orphan_tool_responses(
+            recent_messages, orphan_tool_call_ids
+        )
+
+    if messages_to_prepend:
+        logger.info(
+            f"Extended recent messages by {len(messages_to_prepend)} to preserve "
+            f"tool_call/tool_response pairs"
+        )
+        return messages_to_prepend + recent_messages
+
+    return recent_messages
+
+
+async def _summarize_messages_llm(
+    messages: list[dict],
+    client: AsyncOpenAI,
+    model: str,
+    timeout: float = 30.0,
+) -> str:
+    """Summarize messages using an LLM."""
+    conversation = []
+    for msg in messages:
+        role = msg.get("role", "")
+        content = msg.get("content", "")
+        if content and role in ("user", "assistant", "tool"):
+            conversation.append(f"{role.upper()}: {content}")
+
+    conversation_text = "\n\n".join(conversation)
+
+    if not conversation_text:
+        return "No conversation history available."
+
+    # Limit to ~100k chars for safety
+    MAX_CHARS = 100_000
+    if len(conversation_text) > MAX_CHARS:
+        conversation_text = conversation_text[:MAX_CHARS] + "\n\n[truncated]"
+
+    response = await client.with_options(timeout=timeout).chat.completions.create(
+        model=model,
+        messages=[
+            {
+                "role": "system",
+                "content": (
+                    "Create a detailed summary of the conversation so far. "
+                    "This summary will be used as context when continuing the conversation.\n\n"
+                    "Before writing the summary, analyze each message chronologically to identify:\n"
+                    "- User requests and their explicit goals\n"
+                    "- Your approach and key decisions made\n"
+                    "- Technical specifics (file names, tool outputs, function signatures)\n"
+                    "- Errors encountered and resolutions applied\n\n"
+                    "You MUST include ALL of the following sections:\n\n"
+                    "## 1. Primary Request and Intent\n"
+                    "The user's explicit goals and what they are trying to accomplish.\n\n"
+                    "## 2. Key Technical Concepts\n"
+                    "Technologies, frameworks, tools, and patterns being used or discussed.\n\n"
+                    "## 3. Files and Resources Involved\n"
+                    "Specific files examined or modified, with relevant snippets and identifiers.\n\n"
+                    "## 4. Errors and Fixes\n"
+                    "Problems encountered, error messages, and their resolutions. "
+                    "Include any user feedback on fixes.\n\n"
+                    "## 5. Problem Solving\n"
+                    "Issues that have been resolved and how they were addressed.\n\n"
+                    "## 6. All User Messages\n"
+                    "A complete list of all user inputs (excluding tool outputs) to preserve their exact requests.\n\n"
+                    "## 7. Pending Tasks\n"
+                    "Work items the user explicitly requested that have not yet been completed.\n\n"
+                    "## 8. Current Work\n"
+                    "Precise description of what was being worked on most recently, including relevant context.\n\n"
+                    "## 9. Next Steps\n"
+                    "What should happen next, aligned with the user's most recent requests. "
+                    "Include verbatim quotes of recent instructions if relevant."
+                ),
+            },
+            {"role": "user", "content": f"Summarize:\n\n{conversation_text}"},
+        ],
+        max_tokens=1500,
+        temperature=0.3,
+    )
+
+    return response.choices[0].message.content or "No summary available."
+
+
+async def compress_context(
+    messages: list[dict],
+    target_tokens: int = DEFAULT_TOKEN_THRESHOLD,
+    *,
+    model: str = "gpt-4o",
+    client: AsyncOpenAI | None = None,
+    keep_recent: int = DEFAULT_KEEP_RECENT,
+    reserve: int = 2_048,
+    start_cap: int = 8_192,
+    floor_cap: int = 128,
+) -> CompressResult:
+    """
+    Unified context compression that combines summarization and truncation strategies.
+
+    Strategy (in order):
+    1. **LLM summarization** – If client provided, summarize old messages into a
+       single context message while keeping recent messages intact. This is the
+       primary strategy for chat service.
+    2. **Content truncation** – Progressively halve a per-message cap and truncate
+       bloated message content (tool outputs, large pastes). Preserves all messages
+       but shortens their content. Primary strategy when client=None (LLM blocks).
+    3. **Middle-out deletion** – Delete whole messages one at a time from the center
+       outward, skipping tool messages and objective messages.
+    4. **First/last trim** – Truncate first and last message content as last resort.
+
+    Parameters
+    ----------
+    messages        Complete chat history (will be deep-copied).
+    target_tokens   Hard ceiling for prompt size.
+    model           Model name for tokenization and summarization.
+    client          AsyncOpenAI client. If provided, enables LLM summarization
+                    as the first strategy. If None, skips to truncation strategies.
+    keep_recent     Number of recent messages to preserve during summarization.
+    reserve         Tokens to reserve for model response.
+    start_cap       Initial per-message truncation ceiling (tokens).
+    floor_cap       Lowest cap before moving to deletions.
+
+    Returns
+    -------
+    CompressResult with compressed messages and metadata.
+    """
+    # Guard clause for empty messages
+    if not messages:
+        return CompressResult(
+            messages=[],
+            token_count=0,
+            was_compacted=False,
+            original_token_count=0,
+        )
+
+    token_model = _normalize_model_for_tokenizer(model)
+    enc = encoding_for_model(token_model)
+    msgs = deepcopy(messages)
+
+    def total_tokens() -> int:
+        return sum(_msg_tokens(m, enc) for m in msgs)
+
+    original_count = total_tokens()
+
+    # Already under limit
+    if original_count + reserve <= target_tokens:
+        return CompressResult(
+            messages=msgs,
+            token_count=original_count,
+            was_compacted=False,
+            original_token_count=original_count,
+        )
+
+    messages_summarized = 0
+    messages_dropped = 0
+
+    # ---- STEP 1: LLM summarization (if client provided) -------------------
+    # This is the primary compression strategy for chat service.
+    # Summarize old messages while keeping recent ones intact.
+    if client is not None:
+        has_system = len(msgs) > 0 and msgs[0].get("role") == "system"
+        system_msg = msgs[0] if has_system else None
+
+        # Calculate old vs recent messages
+        if has_system:
+            if len(msgs) > keep_recent + 1:
+                old_msgs = msgs[1:-keep_recent]
+                recent_msgs = msgs[-keep_recent:]
+            else:
+                old_msgs = []
+                recent_msgs = msgs[1:] if len(msgs) > 1 else []
+        else:
+            if len(msgs) > keep_recent:
+                old_msgs = msgs[:-keep_recent]
+                recent_msgs = msgs[-keep_recent:]
+            else:
+                old_msgs = []
+                recent_msgs = msgs
+
+        # Ensure tool pairs stay intact
+        slice_start = max(0, len(msgs) - keep_recent)
+        recent_msgs = _ensure_tool_pairs_intact(recent_msgs, msgs, slice_start)
+
+        if old_msgs:
+            try:
+                summary_text = await _summarize_messages_llm(old_msgs, client, model)
+                summary_msg = {
+                    "role": "assistant",
+                    "content": f"[Previous conversation summary — for context only]: {summary_text}",
+                }
+                messages_summarized = len(old_msgs)
+
+                if has_system:
+                    msgs = [system_msg, summary_msg] + recent_msgs
+                else:
+                    msgs = [summary_msg] + recent_msgs
+
+                logger.info(
+                    f"Context summarized: {original_count} -> {total_tokens()} tokens, "
+                    f"summarized {messages_summarized} messages"
+                )
+            except Exception as e:
+                logger.warning(f"Summarization failed, continuing with truncation: {e}")
+                # Fall through to content truncation
+
+    # ---- STEP 2: Normalize content ----------------------------------------
+    # Convert non-string payloads to strings so token counting is coherent.
+    # Always run this before truncation to ensure consistent token counting.
+    for i, m in enumerate(msgs):
+        if not isinstance(m.get("content"), str) and m.get("content") is not None:
+            if _is_tool_message(m):
+                continue
+            if i == 0 or i == len(msgs) - 1:
+                continue
+            content_str = json.dumps(m["content"], separators=(",", ":"))
+            if len(content_str) > 20_000:
+                content_str = _truncate_middle_tokens(content_str, enc, 20_000)
+            m["content"] = content_str
+
+    # ---- STEP 3: Token-aware content truncation ---------------------------
+    # Progressively halve per-message cap and truncate bloated content.
+    # This preserves all messages but shortens their content.
+    cap = start_cap
+    while total_tokens() + reserve > target_tokens and cap >= floor_cap:
+        for m in msgs[1:-1]:
+            if _is_tool_message(m):
+                _truncate_tool_message_content(m, enc, cap)
+                continue
+            if _is_objective_message(m):
+                continue
+            content = m.get("content") or ""
+            if _tok_len(content, enc) > cap:
+                m["content"] = _truncate_middle_tokens(content, enc, cap)
+        cap //= 2
+
+    # ---- STEP 4: Middle-out deletion --------------------------------------
+    # Delete messages one at a time from the center outward.
+    # This is more granular than dropping all old messages at once.
+    while total_tokens() + reserve > target_tokens and len(msgs) > 2:
+        deletable: list[int] = []
+        for i in range(1, len(msgs) - 1):
+            msg = msgs[i]
+            if (
+                msg is not None
+                and not _is_tool_message(msg)
+                and not _is_objective_message(msg)
+            ):
+                deletable.append(i)
+        if not deletable:
+            break
+        centre = len(msgs) // 2
+        to_delete = min(deletable, key=lambda i: abs(i - centre))
+        del msgs[to_delete]
+        messages_dropped += 1
+
+    # ---- STEP 5: Final trim on first/last ---------------------------------
+    cap = start_cap
+    while total_tokens() + reserve > target_tokens and cap >= floor_cap:
+        for idx in (0, -1):
+            msg = msgs[idx]
+            if msg is None:
+                continue
+            if _is_tool_message(msg):
+                _truncate_tool_message_content(msg, enc, cap)
+                continue
+            text = msg.get("content") or ""
+            if _tok_len(text, enc) > cap:
+                msg["content"] = _truncate_middle_tokens(text, enc, cap)
+        cap //= 2
+
+    # Filter out any None values that may have been introduced
+    final_msgs: list[dict] = [m for m in msgs if m is not None]
+    final_count = sum(_msg_tokens(m, enc) for m in final_msgs)
+    error = None
+    if final_count + reserve > target_tokens:
+        error = f"Could not compress below target ({final_count + reserve} > {target_tokens})"
+        logger.warning(error)
+
+    return CompressResult(
+        messages=final_msgs,
+        token_count=final_count,
+        was_compacted=True,
+        error=error,
+        original_token_count=original_count,
+        messages_summarized=messages_summarized,
+        messages_dropped=messages_dropped,
+    )

autogpt_platform/backend/backend/util/prompt_test.py

@@ -1,10 +1,21 @@
 """Tests for prompt utility functions, especially tool call token counting."""
 
+from unittest.mock import AsyncMock, MagicMock
+
 import pytest
 from tiktoken import encoding_for_model
 
 from backend.util import json
-from backend.util.prompt import _msg_tokens, estimate_token_count
+from backend.util.prompt import (
+    CompressResult,
+    _ensure_tool_pairs_intact,
+    _msg_tokens,
+    _normalize_model_for_tokenizer,
+    _truncate_middle_tokens,
+    _truncate_tool_message_content,
+    compress_context,
+    estimate_token_count,
+)
 
 
 class TestMsgTokens:
@@ -276,3 +287,690 @@ class TestEstimateTokenCount:
 
         assert total_tokens == expected_total
         assert total_tokens > 20  # Should be substantial
+
+
+class TestNormalizeModelForTokenizer:
+    """Test model name normalization for tiktoken."""
+
+    def test_openai_models_unchanged(self):
+        """Test that OpenAI models are returned as-is."""
+        assert _normalize_model_for_tokenizer("gpt-4o") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("gpt-4") == "gpt-4"
+        assert _normalize_model_for_tokenizer("gpt-3.5-turbo") == "gpt-3.5-turbo"
+
+    def test_claude_models_normalized(self):
+        """Test that Claude models are normalized to gpt-4o."""
+        assert _normalize_model_for_tokenizer("claude-3-opus") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("claude-3-sonnet") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("anthropic/claude-3-haiku") == "gpt-4o"
+
+    def test_openrouter_paths_extracted(self):
+        """Test that OpenRouter model paths are handled."""
+        assert _normalize_model_for_tokenizer("openai/gpt-4o") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("anthropic/claude-3-opus") == "gpt-4o"
+
+    def test_unknown_models_default_to_gpt4o(self):
+        """Test that unknown models default to gpt-4o."""
+        assert _normalize_model_for_tokenizer("some-random-model") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("llama-3-70b") == "gpt-4o"
+
+
+class TestTruncateToolMessageContent:
+    """Test tool message content truncation."""
+
+    @pytest.fixture
+    def enc(self):
+        return encoding_for_model("gpt-4o")
+
+    def test_truncate_openai_tool_message(self, enc):
+        """Test truncation of OpenAI-style tool message with string content."""
+        long_content = "x" * 10000
+        msg = {"role": "tool", "tool_call_id": "call_123", "content": long_content}
+
+        _truncate_tool_message_content(msg, enc, max_tokens=100)
+
+        # Content should be truncated
+        assert len(msg["content"]) < len(long_content)
+        assert "…" in msg["content"]  # Has ellipsis marker
+
+    def test_truncate_anthropic_tool_result(self, enc):
+        """Test truncation of Anthropic-style tool_result."""
+        long_content = "y" * 10000
+        msg = {
+            "role": "user",
+            "content": [
+                {
+                    "type": "tool_result",
+                    "tool_use_id": "toolu_123",
+                    "content": long_content,
+                }
+            ],
+        }
+
+        _truncate_tool_message_content(msg, enc, max_tokens=100)
+
+        # Content should be truncated
+        result_content = msg["content"][0]["content"]
+        assert len(result_content) < len(long_content)
+        assert "…" in result_content
+
+    def test_preserve_tool_use_blocks(self, enc):
+        """Test that tool_use blocks are not truncated."""
+        msg = {
+            "role": "assistant",
+            "content": [
+                {
+                    "type": "tool_use",
+                    "id": "toolu_123",
+                    "name": "some_function",
+                    "input": {"key": "value" * 1000},  # Large input
+                }
+            ],
+        }
+
+        original = json.dumps(msg["content"][0]["input"])
+        _truncate_tool_message_content(msg, enc, max_tokens=10)
+
+        # tool_use should be unchanged
+        assert json.dumps(msg["content"][0]["input"]) == original
+
+    def test_no_truncation_when_under_limit(self, enc):
+        """Test that short content is not modified."""
+        msg = {"role": "tool", "tool_call_id": "call_123", "content": "Short content"}
+
+        original = msg["content"]
+        _truncate_tool_message_content(msg, enc, max_tokens=1000)
+
+        assert msg["content"] == original
+
+
+class TestTruncateMiddleTokens:
+    """Test middle truncation of text."""
+
+    @pytest.fixture
+    def enc(self):
+        return encoding_for_model("gpt-4o")
+
+    def test_truncates_long_text(self, enc):
+        """Test that long text is truncated with ellipsis in middle."""
+        long_text = "word " * 1000
+        result = _truncate_middle_tokens(long_text, enc, max_tok=50)
+
+        assert len(enc.encode(result)) <= 52  # Allow some slack for ellipsis
+        assert "…" in result
+        assert result.startswith("word")  # Head preserved
+        assert result.endswith("word ")  # Tail preserved
+
+    def test_preserves_short_text(self, enc):
+        """Test that short text is not modified."""
+        short_text = "Hello world"
+        result = _truncate_middle_tokens(short_text, enc, max_tok=100)
+
+        assert result == short_text
+
+
+class TestEnsureToolPairsIntact:
+    """Test tool call/response pair preservation for both OpenAI and Anthropic formats."""
+
+    # ---- OpenAI Format Tests ----
+
+    def test_openai_adds_missing_tool_call(self):
+        """Test that orphaned OpenAI tool_response gets its tool_call prepended."""
+        all_msgs = [
+            {"role": "system", "content": "You are helpful."},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {"id": "call_1", "type": "function", "function": {"name": "f1"}}
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "result"},
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (the tool response)
+        recent = [all_msgs[2], all_msgs[3]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the tool_call message
+        assert len(result) == 3
+        assert result[0]["role"] == "assistant"
+        assert "tool_calls" in result[0]
+
+    def test_openai_keeps_complete_pairs(self):
+        """Test that complete OpenAI pairs are unchanged."""
+        all_msgs = [
+            {"role": "system", "content": "System"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {"id": "call_1", "type": "function", "function": {"name": "f1"}}
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "result"},
+        ]
+        recent = all_msgs[1:]  # Include both tool_call and response
+        start_index = 1
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        assert len(result) == 2  # No messages added
+
+    def test_openai_multiple_tool_calls(self):
+        """Test multiple OpenAI tool calls in one assistant message."""
+        all_msgs = [
+            {"role": "system", "content": "System"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {"id": "call_1", "type": "function", "function": {"name": "f1"}},
+                    {"id": "call_2", "type": "function", "function": {"name": "f2"}},
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "result1"},
+            {"role": "tool", "tool_call_id": "call_2", "content": "result2"},
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (first tool response)
+        recent = [all_msgs[2], all_msgs[3], all_msgs[4]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the assistant message with both tool_calls
+        assert len(result) == 4
+        assert result[0]["role"] == "assistant"
+        assert len(result[0]["tool_calls"]) == 2
+
+    # ---- Anthropic Format Tests ----
+
+    def test_anthropic_adds_missing_tool_use(self):
+        """Test that orphaned Anthropic tool_result gets its tool_use prepended."""
+        all_msgs = [
+            {"role": "system", "content": "You are helpful."},
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_123",
+                        "name": "get_weather",
+                        "input": {"location": "SF"},
+                    }
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_123",
+                        "content": "22°C and sunny",
+                    }
+                ],
+            },
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (the tool_result)
+        recent = [all_msgs[2], all_msgs[3]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the tool_use message
+        assert len(result) == 3
+        assert result[0]["role"] == "assistant"
+        assert result[0]["content"][0]["type"] == "tool_use"
+
+    def test_anthropic_keeps_complete_pairs(self):
+        """Test that complete Anthropic pairs are unchanged."""
+        all_msgs = [
+            {"role": "system", "content": "System"},
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_456",
+                        "name": "calculator",
+                        "input": {"expr": "2+2"},
+                    }
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_456",
+                        "content": "4",
+                    }
+                ],
+            },
+        ]
+        recent = all_msgs[1:]  # Include both tool_use and result
+        start_index = 1
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        assert len(result) == 2  # No messages added
+
+    def test_anthropic_multiple_tool_uses(self):
+        """Test multiple Anthropic tool_use blocks in one message."""
+        all_msgs = [
+            {"role": "system", "content": "System"},
+            {
+                "role": "assistant",
+                "content": [
+                    {"type": "text", "text": "Let me check both..."},
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_1",
+                        "name": "get_weather",
+                        "input": {"city": "NYC"},
+                    },
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_2",
+                        "name": "get_weather",
+                        "input": {"city": "LA"},
+                    },
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_1",
+                        "content": "Cold",
+                    },
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_2",
+                        "content": "Warm",
+                    },
+                ],
+            },
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (tool_result)
+        recent = [all_msgs[2], all_msgs[3]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the assistant message with both tool_uses
+        assert len(result) == 3
+        assert result[0]["role"] == "assistant"
+        tool_use_count = sum(
+            1 for b in result[0]["content"] if b.get("type") == "tool_use"
+        )
+        assert tool_use_count == 2
+
+    # ---- Mixed/Edge Case Tests ----
+
+    def test_anthropic_with_type_message_field(self):
+        """Test Anthropic format with 'type': 'message' field (smart_decision_maker style)."""
+        all_msgs = [
+            {"role": "system", "content": "You are helpful."},
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_abc",
+                        "name": "search",
+                        "input": {"q": "test"},
+                    }
+                ],
+            },
+            {
+                "role": "user",
+                "type": "message",  # Extra field from smart_decision_maker
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_abc",
+                        "content": "Found results",
+                    }
+                ],
+            },
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (the tool_result with 'type': 'message')
+        recent = [all_msgs[2], all_msgs[3]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the tool_use message
+        assert len(result) == 3
+        assert result[0]["role"] == "assistant"
+        assert result[0]["content"][0]["type"] == "tool_use"
+
+    def test_handles_no_tool_messages(self):
+        """Test messages without tool calls."""
+        all_msgs = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there!"},
+        ]
+        recent = all_msgs
+        start_index = 0
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        assert result == all_msgs
+
+    def test_handles_empty_messages(self):
+        """Test empty message list."""
+        result = _ensure_tool_pairs_intact([], [], 0)
+        assert result == []
+
+    def test_mixed_text_and_tool_content(self):
+        """Test Anthropic message with mixed text and tool_use content."""
+        all_msgs = [
+            {
+                "role": "assistant",
+                "content": [
+                    {"type": "text", "text": "I'll help you with that."},
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_mixed",
+                        "name": "search",
+                        "input": {"q": "test"},
+                    },
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_mixed",
+                        "content": "Found results",
+                    }
+                ],
+            },
+            {"role": "assistant", "content": "Here are the results..."},
+        ]
+        # Start from tool_result
+        recent = [all_msgs[1], all_msgs[2]]
+        start_index = 1
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the assistant message with tool_use
+        assert len(result) == 3
+        assert result[0]["content"][0]["type"] == "text"
+        assert result[0]["content"][1]["type"] == "tool_use"
+
+
+class TestCompressContext:
+    """Test the async compress_context function."""
+
+    @pytest.mark.asyncio
+    async def test_no_compression_needed(self):
+        """Test messages under limit return without compression."""
+        messages = [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "user", "content": "Hello!"},
+        ]
+
+        result = await compress_context(messages, target_tokens=100000)
+
+        assert isinstance(result, CompressResult)
+        assert result.was_compacted is False
+        assert len(result.messages) == 2
+        assert result.error is None
+
+    @pytest.mark.asyncio
+    async def test_truncation_without_client(self):
+        """Test that truncation works without LLM client."""
+        long_content = "x" * 50000
+        messages = [
+            {"role": "system", "content": "System"},
+            {"role": "user", "content": long_content},
+            {"role": "assistant", "content": "Response"},
+        ]
+
+        result = await compress_context(
+            messages, target_tokens=1000, client=None, reserve=100
+        )
+
+        assert result.was_compacted is True
+        # Should have truncated without summarization
+        assert result.messages_summarized == 0
+
+    @pytest.mark.asyncio
+    async def test_with_mocked_llm_client(self):
+        """Test summarization with mocked LLM client."""
+        # Create many messages to trigger summarization
+        messages = [{"role": "system", "content": "System prompt"}]
+        for i in range(30):
+            messages.append({"role": "user", "content": f"User message {i} " * 100})
+            messages.append(
+                {"role": "assistant", "content": f"Assistant response {i} " * 100}
+            )
+
+        # Mock the AsyncOpenAI client
+        mock_client = AsyncMock()
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        mock_response.choices[0].message.content = "Summary of conversation"
+        mock_client.with_options.return_value.chat.completions.create = AsyncMock(
+            return_value=mock_response
+        )
+
+        result = await compress_context(
+            messages,
+            target_tokens=5000,
+            client=mock_client,
+            keep_recent=5,
+            reserve=500,
+        )
+
+        assert result.was_compacted is True
+        # Should have attempted summarization
+        assert mock_client.with_options.called or result.messages_summarized > 0
+
+    @pytest.mark.asyncio
+    async def test_preserves_tool_pairs(self):
+        """Test that tool call/response pairs stay together."""
+        messages = [
+            {"role": "system", "content": "System"},
+            {"role": "user", "content": "Do something"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {"id": "call_1", "type": "function", "function": {"name": "func"}}
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "Result " * 1000},
+            {"role": "assistant", "content": "Done!"},
+        ]
+
+        result = await compress_context(
+            messages, target_tokens=500, client=None, reserve=50
+        )
+
+        # Check that if tool response exists, its call exists too
+        tool_call_ids = set()
+        tool_response_ids = set()
+        for msg in result.messages:
+            if "tool_calls" in msg:
+                for tc in msg["tool_calls"]:
+                    tool_call_ids.add(tc["id"])
+            if msg.get("role") == "tool":
+                tool_response_ids.add(msg.get("tool_call_id"))
+
+        # All tool responses should have their calls
+        assert tool_response_ids <= tool_call_ids
+
+    @pytest.mark.asyncio
+    async def test_returns_error_when_cannot_compress(self):
+        """Test that error is returned when compression fails."""
+        # Single huge message that can't be compressed enough
+        messages = [
+            {"role": "user", "content": "x" * 100000},
+        ]
+
+        result = await compress_context(
+            messages, target_tokens=100, client=None, reserve=50
+        )
+
+        # Should have an error since we can't get below 100 tokens
+        assert result.error is not None
+        assert result.was_compacted is True
+
+    @pytest.mark.asyncio
+    async def test_empty_messages(self):
+        """Test that empty messages list returns early without error."""
+        result = await compress_context([], target_tokens=1000)
+
+        assert result.messages == []
+        assert result.token_count == 0
+        assert result.was_compacted is False
+        assert result.error is None
+
+
+class TestRemoveOrphanToolResponses:
+    """Test _remove_orphan_tool_responses helper function."""
+
+    def test_removes_openai_orphan(self):
+        """Test removal of orphan OpenAI tool response."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {"role": "tool", "tool_call_id": "call_orphan", "content": "result"},
+            {"role": "user", "content": "Hello"},
+        ]
+        orphan_ids = {"call_orphan"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        assert len(result) == 1
+        assert result[0]["role"] == "user"
+
+    def test_keeps_valid_openai_tool(self):
+        """Test that valid OpenAI tool responses are kept."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {"role": "tool", "tool_call_id": "call_valid", "content": "result"},
+        ]
+        orphan_ids = {"call_other"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        assert len(result) == 1
+        assert result[0]["tool_call_id"] == "call_valid"
+
+    def test_filters_anthropic_mixed_blocks(self):
+        """Test filtering individual orphan blocks from Anthropic message with mixed valid/orphan."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_valid",
+                        "content": "valid result",
+                    },
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_orphan",
+                        "content": "orphan result",
+                    },
+                ],
+            },
+        ]
+        orphan_ids = {"toolu_orphan"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        assert len(result) == 1
+        # Should only have the valid tool_result, orphan filtered out
+        assert len(result[0]["content"]) == 1
+        assert result[0]["content"][0]["tool_use_id"] == "toolu_valid"
+
+    def test_removes_anthropic_all_orphan(self):
+        """Test removal of Anthropic message when all tool_results are orphans."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_orphan1",
+                        "content": "result1",
+                    },
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_orphan2",
+                        "content": "result2",
+                    },
+                ],
+            },
+        ]
+        orphan_ids = {"toolu_orphan1", "toolu_orphan2"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        # Message should be completely removed since no content left
+        assert len(result) == 0
+
+    def test_preserves_non_tool_messages(self):
+        """Test that non-tool messages are preserved."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there!"},
+        ]
+        orphan_ids = {"some_id"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        assert result == messages
+
+
+class TestCompressResultDataclass:
+    """Test CompressResult dataclass."""
+
+    def test_default_values(self):
+        """Test default values are set correctly."""
+        result = CompressResult(
+            messages=[{"role": "user", "content": "test"}],
+            token_count=10,
+            was_compacted=False,
+        )
+
+        assert result.error is None
+        assert result.original_token_count == 0  # Defaults to 0, not None
+        assert result.messages_summarized == 0
+        assert result.messages_dropped == 0
+
+    def test_all_fields(self):
+        """Test all fields can be set."""
+        result = CompressResult(
+            messages=[{"role": "user", "content": "test"}],
+            token_count=100,
+            was_compacted=True,
+            error="Some error",
+            original_token_count=500,
+            messages_summarized=10,
+            messages_dropped=5,
+        )
+
+        assert result.token_count == 100
+        assert result.was_compacted is True
+        assert result.error == "Some error"
+        assert result.original_token_count == 500
+        assert result.messages_summarized == 10
+        assert result.messages_dropped == 5

autogpt_platform/frontend/src/app/(no-navbar)/onboarding/page.tsx

@@ -1,10 +1,9 @@
 "use client";
+import { getV1OnboardingState } from "@/app/api/__generated__/endpoints/onboarding/onboarding";
+import { getOnboardingStatus, resolveResponse } from "@/app/api/helpers";
 import { LoadingSpinner } from "@/components/atoms/LoadingSpinner/LoadingSpinner";
 import { useRouter } from "next/navigation";
 import { useEffect } from "react";
-import { resolveResponse, getOnboardingStatus } from "@/app/api/helpers";
-import { getV1OnboardingState } from "@/app/api/__generated__/endpoints/onboarding/onboarding";
-import { getHomepageRoute } from "@/lib/constants";
 
 export default function OnboardingPage() {
   const router = useRouter();
@@ -13,12 +12,10 @@ export default function OnboardingPage() {
     async function redirectToStep() {
       try {
         // Check if onboarding is enabled (also gets chat flag for redirect)
-        const { shouldShowOnboarding, isChatEnabled } =
-          await getOnboardingStatus();
-        const homepageRoute = getHomepageRoute(isChatEnabled);
+        const { shouldShowOnboarding } = await getOnboardingStatus();
 
         if (!shouldShowOnboarding) {
-          router.replace(homepageRoute);
+          router.replace("/");
           return;
         }
 
@@ -26,7 +23,7 @@ export default function OnboardingPage() {
 
         // Handle completed onboarding
         if (onboarding.completedSteps.includes("GET_RESULTS")) {
-          router.replace(homepageRoute);
+          router.replace("/");
           return;
         }
 

autogpt_platform/frontend/src/app/(platform)/auth/callback/route.ts

@@ -1,9 +1,8 @@
-import { getServerSupabase } from "@/lib/supabase/server/getServerSupabase";
-import { getHomepageRoute } from "@/lib/constants";
-import BackendAPI from "@/lib/autogpt-server-api";
-import { NextResponse } from "next/server";
-import { revalidatePath } from "next/cache";
 import { getOnboardingStatus } from "@/app/api/helpers";
+import BackendAPI from "@/lib/autogpt-server-api";
+import { getServerSupabase } from "@/lib/supabase/server/getServerSupabase";
+import { revalidatePath } from "next/cache";
+import { NextResponse } from "next/server";
 
 // Handle the callback to complete the user session login
 export async function GET(request: Request) {
@@ -27,13 +26,12 @@ export async function GET(request: Request) {
         await api.createUser();
 
         // Get onboarding status from backend (includes chat flag evaluated for this user)
-        const { shouldShowOnboarding, isChatEnabled } =
-          await getOnboardingStatus();
+        const { shouldShowOnboarding } = await getOnboardingStatus();
         if (shouldShowOnboarding) {
           next = "/onboarding";
           revalidatePath("/onboarding", "layout");
         } else {
-          next = getHomepageRoute(isChatEnabled);
+          next = "/";
           revalidatePath(next, "layout");
         }
       } catch (createUserError) {

autogpt_platform/frontend/src/app/(platform)/copilot/layout.tsx

@@ -1,6 +1,13 @@
-import type { ReactNode } from "react";
+"use client";
+import { FeatureFlagPage } from "@/services/feature-flags/FeatureFlagPage";
+import { Flag } from "@/services/feature-flags/use-get-flag";
+import { type ReactNode } from "react";
 import { CopilotShell } from "./components/CopilotShell/CopilotShell";
 
 export default function CopilotLayout({ children }: { children: ReactNode }) {
-  return <CopilotShell>{children}</CopilotShell>;
+  return (
+    <FeatureFlagPage flag={Flag.CHAT} whenDisabled="/library">
+      <CopilotShell>{children}</CopilotShell>
+    </FeatureFlagPage>
+  );
 }

autogpt_platform/frontend/src/app/(platform)/copilot/page.tsx

@@ -14,14 +14,8 @@ export default function CopilotPage() {
   const isInterruptModalOpen = useCopilotStore((s) => s.isInterruptModalOpen);
   const confirmInterrupt = useCopilotStore((s) => s.confirmInterrupt);
   const cancelInterrupt = useCopilotStore((s) => s.cancelInterrupt);
-  const {
-    greetingName,
-    quickActions,
-    isLoading,
-    hasSession,
-    initialPrompt,
-    isReady,
-  } = state;
+  const { greetingName, quickActions, isLoading, hasSession, initialPrompt } =
+    state;
   const {
     handleQuickAction,
     startChatWithPrompt,
@@ -29,8 +23,6 @@ export default function CopilotPage() {
     handleStreamingChange,
   } = handlers;
 
-  if (!isReady) return null;
-
   if (hasSession) {
     return (
       <div className="flex h-full flex-col">

autogpt_platform/frontend/src/app/(platform)/copilot/useCopilotPage.ts

@@ -3,18 +3,11 @@ import {
   postV2CreateSession,
 } from "@/app/api/__generated__/endpoints/chat/chat";
 import { useToast } from "@/components/molecules/Toast/use-toast";
-import { getHomepageRoute } from "@/lib/constants";
 import { useSupabase } from "@/lib/supabase/hooks/useSupabase";
 import { useOnboarding } from "@/providers/onboarding/onboarding-provider";
-import {
-  Flag,
-  type FlagValues,
-  useGetFlag,
-} from "@/services/feature-flags/use-get-flag";
 import { SessionKey, sessionStorage } from "@/services/storage/session-storage";
 import * as Sentry from "@sentry/nextjs";
 import { useQueryClient } from "@tanstack/react-query";
-import { useFlags } from "launchdarkly-react-client-sdk";
 import { useRouter } from "next/navigation";
 import { useEffect } from "react";
 import { useCopilotStore } from "./copilot-page-store";
@@ -33,22 +26,6 @@ export function useCopilotPage() {
   const isCreating = useCopilotStore((s) => s.isCreatingSession);
   const setIsCreating = useCopilotStore((s) => s.setIsCreatingSession);
 
-  // Complete VISIT_COPILOT onboarding step to grant $5 welcome bonus
-  useEffect(() => {
-    if (isLoggedIn) {
-      completeStep("VISIT_COPILOT");
-    }
-  }, [completeStep, isLoggedIn]);
-
-  const isChatEnabled = useGetFlag(Flag.CHAT);
-  const flags = useFlags<FlagValues>();
-  const homepageRoute = getHomepageRoute(isChatEnabled);
-  const envEnabled = process.env.NEXT_PUBLIC_LAUNCHDARKLY_ENABLED === "true";
-  const clientId = process.env.NEXT_PUBLIC_LAUNCHDARKLY_CLIENT_ID;
-  const isLaunchDarklyConfigured = envEnabled && Boolean(clientId);
-  const isFlagReady =
-    !isLaunchDarklyConfigured || flags[Flag.CHAT] !== undefined;
-
   const greetingName = getGreetingName(user);
   const quickActions = getQuickActions();
 
@@ -58,11 +35,8 @@ export function useCopilotPage() {
     : undefined;
 
   useEffect(() => {
-    if (!isFlagReady) return;
-    if (isChatEnabled === false) {
-      router.replace(homepageRoute);
-    }
-  }, [homepageRoute, isChatEnabled, isFlagReady, router]);
+    if (isLoggedIn) completeStep("VISIT_COPILOT");
+  }, [completeStep, isLoggedIn]);
 
   async function startChatWithPrompt(prompt: string) {
     if (!prompt?.trim()) return;
@@ -116,7 +90,6 @@ export function useCopilotPage() {
       isLoading: isUserLoading,
       hasSession,
       initialPrompt,
-      isReady: isFlagReady && isChatEnabled !== false && isLoggedIn,
     },
     handlers: {
       handleQuickAction,

autogpt_platform/frontend/src/app/(platform)/error/page.tsx

@@ -1,8 +1,6 @@
 "use client";
 
 import { ErrorCard } from "@/components/molecules/ErrorCard/ErrorCard";
-import { getHomepageRoute } from "@/lib/constants";
-import { Flag, useGetFlag } from "@/services/feature-flags/use-get-flag";
 import { useSearchParams } from "next/navigation";
 import { Suspense } from "react";
 import { getErrorDetails } from "./helpers";
@@ -11,8 +9,6 @@ function ErrorPageContent() {
   const searchParams = useSearchParams();
   const errorMessage = searchParams.get("message");
   const errorDetails = getErrorDetails(errorMessage);
-  const isChatEnabled = useGetFlag(Flag.CHAT);
-  const homepageRoute = getHomepageRoute(isChatEnabled);
 
   function handleRetry() {
     // Auth-related errors should redirect to login
@@ -30,7 +26,7 @@ function ErrorPageContent() {
       }, 2000);
     } else {
       // For server/network errors, go to home
-      window.location.href = homepageRoute;
+      window.location.href = "/";
     }
   }
 

autogpt_platform/frontend/src/app/(platform)/login/actions.ts

@@ -1,6 +1,5 @@
 "use server";
 
-import { getHomepageRoute } from "@/lib/constants";
 import BackendAPI from "@/lib/autogpt-server-api";
 import { getServerSupabase } from "@/lib/supabase/server/getServerSupabase";
 import { loginFormSchema } from "@/types/auth";
@@ -38,10 +37,8 @@ export async function login(email: string, password: string) {
     await api.createUser();
 
     // Get onboarding status from backend (includes chat flag evaluated for this user)
-    const { shouldShowOnboarding, isChatEnabled } = await getOnboardingStatus();
-    const next = shouldShowOnboarding
-      ? "/onboarding"
-      : getHomepageRoute(isChatEnabled);
+    const { shouldShowOnboarding } = await getOnboardingStatus();
+    const next = shouldShowOnboarding ? "/onboarding" : "/";
 
     return {
       success: true,

autogpt_platform/frontend/src/app/(platform)/login/useLoginPage.ts

@@ -1,8 +1,6 @@
 import { useToast } from "@/components/molecules/Toast/use-toast";
-import { getHomepageRoute } from "@/lib/constants";
 import { useSupabase } from "@/lib/supabase/hooks/useSupabase";
 import { environment } from "@/services/environment";
-import { Flag, useGetFlag } from "@/services/feature-flags/use-get-flag";
 import { loginFormSchema, LoginProvider } from "@/types/auth";
 import { zodResolver } from "@hookform/resolvers/zod";
 import { useRouter, useSearchParams } from "next/navigation";
@@ -22,17 +20,15 @@ export function useLoginPage() {
   const [isGoogleLoading, setIsGoogleLoading] = useState(false);
   const [showNotAllowedModal, setShowNotAllowedModal] = useState(false);
   const isCloudEnv = environment.isCloud();
-  const isChatEnabled = useGetFlag(Flag.CHAT);
-  const homepageRoute = getHomepageRoute(isChatEnabled);
 
   // Get redirect destination from 'next' query parameter
   const nextUrl = searchParams.get("next");
 
   useEffect(() => {
     if (isLoggedIn && !isLoggingIn) {
-      router.push(nextUrl || homepageRoute);
+      router.push(nextUrl || "/");
     }
-  }, [homepageRoute, isLoggedIn, isLoggingIn, nextUrl, router]);
+  }, [isLoggedIn, isLoggingIn, nextUrl, router]);
 
   const form = useForm<z.infer<typeof loginFormSchema>>({
     resolver: zodResolver(loginFormSchema),
@@ -98,7 +94,7 @@ export function useLoginPage() {
       }
 
       // Prefer URL's next parameter, then use backend-determined route
-      router.replace(nextUrl || result.next || homepageRoute);
+      router.replace(nextUrl || result.next || "/");
     } catch (error) {
       toast({
         title:

autogpt_platform/frontend/src/app/(platform)/signup/actions.ts

@@ -1,6 +1,5 @@
 "use server";
 
-import { getHomepageRoute } from "@/lib/constants";
 import { getServerSupabase } from "@/lib/supabase/server/getServerSupabase";
 import { signupFormSchema } from "@/types/auth";
 import * as Sentry from "@sentry/nextjs";
@@ -59,10 +58,8 @@ export async function signup(
     }
 
     // Get onboarding status from backend (includes chat flag evaluated for this user)
-    const { shouldShowOnboarding, isChatEnabled } = await getOnboardingStatus();
-    const next = shouldShowOnboarding
-      ? "/onboarding"
-      : getHomepageRoute(isChatEnabled);
+    const { shouldShowOnboarding } = await getOnboardingStatus();
+    const next = shouldShowOnboarding ? "/onboarding" : "/";
 
     return { success: true, next };
   } catch (err) {

autogpt_platform/frontend/src/app/(platform)/signup/useSignupPage.ts

@@ -1,8 +1,6 @@
 import { useToast } from "@/components/molecules/Toast/use-toast";
-import { getHomepageRoute } from "@/lib/constants";
 import { useSupabase } from "@/lib/supabase/hooks/useSupabase";
 import { environment } from "@/services/environment";
-import { Flag, useGetFlag } from "@/services/feature-flags/use-get-flag";
 import { LoginProvider, signupFormSchema } from "@/types/auth";
 import { zodResolver } from "@hookform/resolvers/zod";
 import { useRouter, useSearchParams } from "next/navigation";
@@ -22,17 +20,15 @@ export function useSignupPage() {
   const [isGoogleLoading, setIsGoogleLoading] = useState(false);
   const [showNotAllowedModal, setShowNotAllowedModal] = useState(false);
   const isCloudEnv = environment.isCloud();
-  const isChatEnabled = useGetFlag(Flag.CHAT);
-  const homepageRoute = getHomepageRoute(isChatEnabled);
 
   // Get redirect destination from 'next' query parameter
   const nextUrl = searchParams.get("next");
 
   useEffect(() => {
     if (isLoggedIn && !isSigningUp) {
-      router.push(nextUrl || homepageRoute);
+      router.push(nextUrl || "/");
     }
-  }, [homepageRoute, isLoggedIn, isSigningUp, nextUrl, router]);
+  }, [isLoggedIn, isSigningUp, nextUrl, router]);
 
   const form = useForm<z.infer<typeof signupFormSchema>>({
     resolver: zodResolver(signupFormSchema),
@@ -133,7 +129,7 @@ export function useSignupPage() {
       }
 
       // Prefer the URL's next parameter, then result.next (for onboarding), then default
-      const redirectTo = nextUrl || result.next || homepageRoute;
+      const redirectTo = nextUrl || result.next || "/";
       router.replace(redirectTo);
     } catch (error) {
       setIsLoading(false);

autogpt_platform/frontend/src/app/api/helpers.ts

@@ -181,6 +181,5 @@ export async function getOnboardingStatus() {
   const isCompleted = onboarding.completedSteps.includes("CONGRATS");
   return {
     shouldShowOnboarding: status.is_onboarding_enabled && !isCompleted,
-    isChatEnabled: status.is_chat_enabled,
   };
 }

autogpt_platform/frontend/src/app/page.tsx

@@ -1,27 +1,15 @@
 "use client";
 
-import { getHomepageRoute } from "@/lib/constants";
-import { Flag, useGetFlag } from "@/services/feature-flags/use-get-flag";
+import { LoadingSpinner } from "@/components/atoms/LoadingSpinner/LoadingSpinner";
 import { useRouter } from "next/navigation";
 import { useEffect } from "react";
 
 export default function Page() {
-  const isChatEnabled = useGetFlag(Flag.CHAT);
   const router = useRouter();
-  const homepageRoute = getHomepageRoute(isChatEnabled);
-  const envEnabled = process.env.NEXT_PUBLIC_LAUNCHDARKLY_ENABLED === "true";
-  const clientId = process.env.NEXT_PUBLIC_LAUNCHDARKLY_CLIENT_ID;
-  const isLaunchDarklyConfigured = envEnabled && Boolean(clientId);
-  const isFlagReady =
-    !isLaunchDarklyConfigured || typeof isChatEnabled === "boolean";
 
-  useEffect(
-    function redirectToHomepage() {
-      if (!isFlagReady) return;
-      router.replace(homepageRoute);
-    },
-    [homepageRoute, isFlagReady, router],
-  );
+  useEffect(() => {
+    router.replace("/copilot");
+  }, [router]);
 
-  return null;
+  return <LoadingSpinner size="large" cover />;
 }

autogpt_platform/frontend/src/components/layout/Navbar/components/NavbarLink.tsx

@@ -1,7 +1,6 @@
 "use client";
 
 import { IconLaptop } from "@/components/__legacy__/ui/icons";
-import { getHomepageRoute } from "@/lib/constants";
 import { cn } from "@/lib/utils";
 import { Flag, useGetFlag } from "@/services/feature-flags/use-get-flag";
 import { ListChecksIcon } from "@phosphor-icons/react/dist/ssr";
@@ -24,11 +23,11 @@ interface Props {
 export function NavbarLink({ name, href }: Props) {
   const pathname = usePathname();
   const isChatEnabled = useGetFlag(Flag.CHAT);
-  const homepageRoute = getHomepageRoute(isChatEnabled);
+  const expectedHomeRoute = isChatEnabled ? "/copilot" : "/library";
 
   const isActive =
-    href === homepageRoute
-      ? pathname === "/" || pathname.startsWith(homepageRoute)
+    href === expectedHomeRoute
+      ? pathname === "/" || pathname.startsWith(expectedHomeRoute)
       : pathname.includes(href);
 
   return (

autogpt_platform/frontend/src/hooks/useAgentGraph.tsx

@@ -66,7 +66,7 @@ export default function useAgentGraph(
   >(null);
   const [xyNodes, setXYNodes] = useState<CustomNode[]>([]);
   const [xyEdges, setXYEdges] = useState<CustomEdge[]>([]);
-  const betaBlocks = useGetFlag(Flag.BETA_BLOCKS);
+  const betaBlocks = useGetFlag(Flag.BETA_BLOCKS) as string[];
 
   // Filter blocks based on beta flags
   const availableBlocks = useMemo(() => {

autogpt_platform/frontend/src/lib/constants.ts

@@ -11,10 +11,3 @@ export const API_KEY_HEADER_NAME = "X-API-Key";
 
 // Layout
 export const NAVBAR_HEIGHT_PX = 60;
-
-// Routes
-export function getHomepageRoute(isChatEnabled?: boolean | null): string {
-  if (isChatEnabled === true) return "/copilot";
-  if (isChatEnabled === false) return "/library";
-  return "/";
-}

autogpt_platform/frontend/src/lib/supabase/helpers.ts

@@ -1,4 +1,3 @@
-import { getHomepageRoute } from "@/lib/constants";
 import { environment } from "@/services/environment";
 import { Key, storage } from "@/services/storage/local-storage";
 import { type CookieOptions } from "@supabase/ssr";
@@ -71,7 +70,7 @@ export function getRedirectPath(
   }
 
   if (isAdminPage(path) && userRole !== "admin") {
-    return getHomepageRoute();
+    return "/";
   }
 
   return null;

autogpt_platform/frontend/src/lib/supabase/middleware.ts

@@ -1,4 +1,3 @@
-import { getHomepageRoute } from "@/lib/constants";
 import { environment } from "@/services/environment";
 import { createServerClient } from "@supabase/ssr";
 import { NextResponse, type NextRequest } from "next/server";
@@ -67,7 +66,7 @@ export async function updateSession(request: NextRequest) {
 
     // 2. Check if user is authenticated but lacks admin role when accessing admin pages
     if (user && userRole !== "admin" && isAdminPage(pathname)) {
-      url.pathname = getHomepageRoute();
+      url.pathname = "/";
       return NextResponse.redirect(url);
     }
 

autogpt_platform/frontend/src/providers/onboarding/onboarding-provider.tsx

@@ -23,9 +23,7 @@ import {
   WebSocketNotification,
 } from "@/lib/autogpt-server-api";
 import { useBackendAPI } from "@/lib/autogpt-server-api/context";
-import { getHomepageRoute } from "@/lib/constants";
 import { useSupabase } from "@/lib/supabase/hooks/useSupabase";
-import { Flag, useGetFlag } from "@/services/feature-flags/use-get-flag";
 import Link from "next/link";
 import { usePathname, useRouter } from "next/navigation";
 import {
@@ -104,8 +102,6 @@ export default function OnboardingProvider({
   const pathname = usePathname();
   const router = useRouter();
   const { isLoggedIn } = useSupabase();
-  const isChatEnabled = useGetFlag(Flag.CHAT);
-  const homepageRoute = getHomepageRoute(isChatEnabled);
 
   useOnboardingTimezoneDetection();
 
@@ -150,7 +146,7 @@ export default function OnboardingProvider({
         if (isOnOnboardingRoute) {
           const enabled = await resolveResponse(getV1IsOnboardingEnabled());
           if (!enabled) {
-            router.push(homepageRoute);
+            router.push("/");
             return;
           }
         }
@@ -162,7 +158,7 @@ export default function OnboardingProvider({
           isOnOnboardingRoute &&
           shouldRedirectFromOnboarding(onboarding.completedSteps, pathname)
         ) {
-          router.push(homepageRoute);
+          router.push("/");
         }
       } catch (error) {
         console.error("Failed to initialize onboarding:", error);
@@ -177,7 +173,7 @@ export default function OnboardingProvider({
     }
 
     initializeOnboarding();
-  }, [api, homepageRoute, isOnOnboardingRoute, router, isLoggedIn, pathname]);
+  }, [api, isOnOnboardingRoute, router, isLoggedIn, pathname]);
 
   const handleOnboardingNotification = useCallback(
     (notification: WebSocketNotification) => {

autogpt_platform/frontend/src/services/environment/index.ts

@@ -83,6 +83,10 @@ function getPostHogCredentials() {
   };
 }
 
+function getLaunchDarklyClientId() {
+  return process.env.NEXT_PUBLIC_LAUNCHDARKLY_CLIENT_ID;
+}
+
 function isProductionBuild() {
   return process.env.NODE_ENV === "production";
 }
@@ -120,7 +124,10 @@ function isVercelPreview() {
 }
 
 function areFeatureFlagsEnabled() {
-  return process.env.NEXT_PUBLIC_LAUNCHDARKLY_ENABLED === "enabled";
+  return (
+    process.env.NEXT_PUBLIC_LAUNCHDARKLY_ENABLED === "true" &&
+    Boolean(process.env.NEXT_PUBLIC_LAUNCHDARKLY_CLIENT_ID)
+  );
 }
 
 function isPostHogEnabled() {
@@ -143,6 +150,7 @@ export const environment = {
   getSupabaseAnonKey,
   getPreviewStealingDev,
   getPostHogCredentials,
+  getLaunchDarklyClientId,
   // Assertions
   isServerSide,
   isClientSide,

autogpt_platform/frontend/src/services/feature-flags/FeatureFlagPage.tsx

@@ -0,0 +1,59 @@
+"use client";
+
+import { LoadingSpinner } from "@/components/atoms/LoadingSpinner/LoadingSpinner";
+import { useLDClient } from "launchdarkly-react-client-sdk";
+import { useRouter } from "next/navigation";
+import { ReactNode, useEffect, useState } from "react";
+import { environment } from "../environment";
+import { Flag, useGetFlag } from "./use-get-flag";
+
+interface FeatureFlagRedirectProps {
+  flag: Flag;
+  whenDisabled: string;
+  children: ReactNode;
+}
+
+export function FeatureFlagPage({
+  flag,
+  whenDisabled,
+  children,
+}: FeatureFlagRedirectProps) {
+  const [isLoading, setIsLoading] = useState(true);
+  const router = useRouter();
+  const flagValue = useGetFlag(flag);
+  const ldClient = useLDClient();
+  const ldEnabled = environment.areFeatureFlagsEnabled();
+  const ldReady = Boolean(ldClient);
+  const flagEnabled = Boolean(flagValue);
+
+  useEffect(() => {
+    const initialize = async () => {
+      if (!ldEnabled) {
+        router.replace(whenDisabled);
+        setIsLoading(false);
+        return;
+      }
+
+      // Wait for LaunchDarkly to initialize when enabled to prevent race conditions
+      if (ldEnabled && !ldReady) return;
+
+      try {
+        await ldClient?.waitForInitialization();
+        if (!flagEnabled) router.replace(whenDisabled);
+      } catch (error) {
+        console.error(error);
+        router.replace(whenDisabled);
+      } finally {
+        setIsLoading(false);
+      }
+    };
+
+    initialize();
+  }, [ldReady, flagEnabled]);
+
+  return isLoading || !flagEnabled ? (
+    <LoadingSpinner size="large" cover />
+  ) : (
+    <>{children}</>
+  );
+}

autogpt_platform/frontend/src/services/feature-flags/FeatureFlagRedirect.tsx

@@ -0,0 +1,51 @@
+"use client";
+
+import { LoadingSpinner } from "@/components/atoms/LoadingSpinner/LoadingSpinner";
+import { useLDClient } from "launchdarkly-react-client-sdk";
+import { useRouter } from "next/navigation";
+import { useEffect } from "react";
+import { environment } from "../environment";
+import { Flag, useGetFlag } from "./use-get-flag";
+
+interface FeatureFlagRedirectProps {
+  flag: Flag;
+  whenEnabled: string;
+  whenDisabled: string;
+}
+
+export function FeatureFlagRedirect({
+  flag,
+  whenEnabled,
+  whenDisabled,
+}: FeatureFlagRedirectProps) {
+  const router = useRouter();
+  const flagValue = useGetFlag(flag);
+  const ldEnabled = environment.areFeatureFlagsEnabled();
+  const ldClient = useLDClient();
+  const ldReady = Boolean(ldClient);
+  const flagEnabled = Boolean(flagValue);
+
+  useEffect(() => {
+    const initialize = async () => {
+      if (!ldEnabled) {
+        router.replace(whenDisabled);
+        return;
+      }
+
+      // Wait for LaunchDarkly to initialize when enabled to prevent race conditions
+      if (ldEnabled && !ldReady) return;
+
+      try {
+        await ldClient?.waitForInitialization();
+        router.replace(flagEnabled ? whenEnabled : whenDisabled);
+      } catch (error) {
+        console.error(error);
+        router.replace(whenDisabled);
+      }
+    };
+
+    initialize();
+  }, [ldReady, flagEnabled]);
+
+  return <LoadingSpinner size="large" cover />;
+}

autogpt_platform/frontend/src/services/feature-flags/feature-flag-provider.tsx

@@ -1,5 +1,6 @@
 "use client";
 
+import { LoadingSpinner } from "@/components/atoms/LoadingSpinner/LoadingSpinner";
 import { useSupabase } from "@/lib/supabase/hooks/useSupabase";
 import * as Sentry from "@sentry/nextjs";
 import { LDProvider } from "launchdarkly-react-client-sdk";
@@ -7,17 +8,17 @@ import type { ReactNode } from "react";
 import { useMemo } from "react";
 import { environment } from "../environment";
 
-const clientId = process.env.NEXT_PUBLIC_LAUNCHDARKLY_CLIENT_ID;
-const envEnabled = process.env.NEXT_PUBLIC_LAUNCHDARKLY_ENABLED === "true";
 const LAUNCHDARKLY_INIT_TIMEOUT_MS = 5000;
 
 export function LaunchDarklyProvider({ children }: { children: ReactNode }) {
   const { user, isUserLoading } = useSupabase();
-  const isCloud = environment.isCloud();
-  const isLaunchDarklyConfigured = isCloud && envEnabled && clientId;
+  const envEnabled = environment.areFeatureFlagsEnabled();
+  const clientId = environment.getLaunchDarklyClientId();
 
   const context = useMemo(() => {
-    if (isUserLoading || !user) {
+    if (isUserLoading) return;
+
+    if (!user) {
       return {
         kind: "user" as const,
         key: "anonymous",
@@ -36,15 +37,17 @@ export function LaunchDarklyProvider({ children }: { children: ReactNode }) {
     };
   }, [user, isUserLoading]);
 
-  if (!isLaunchDarklyConfigured) {
+  if (!envEnabled) {
     return <>{children}</>;
   }
 
+  if (isUserLoading) {
+    return <LoadingSpinner size="large" cover />;
+  }
+
   return (
     <LDProvider
-      // Add this key prop. It will be 'anonymous' when logged out,
-      key={context.key}
-      clientSideID={clientId}
+      clientSideID={clientId ?? ""}
       context={context}
       timeout={LAUNCHDARKLY_INIT_TIMEOUT_MS}
       reactOptions={{ useCamelCaseFlagKeys: false }}

autogpt_platform/frontend/src/services/feature-flags/use-get-flag.ts

@@ -1,6 +1,7 @@
 "use client";
 
 import { DEFAULT_SEARCH_TERMS } from "@/app/(platform)/marketplace/components/HeroSection/helpers";
+import { environment } from "@/services/environment";
 import { useFlags } from "launchdarkly-react-client-sdk";
 
 export enum Flag {
@@ -18,24 +19,9 @@ export enum Flag {
   CHAT = "chat",
 }
 
-export type FlagValues = {
-  [Flag.BETA_BLOCKS]: string[];
-  [Flag.NEW_BLOCK_MENU]: boolean;
-  [Flag.NEW_AGENT_RUNS]: boolean;
-  [Flag.GRAPH_SEARCH]: boolean;
-  [Flag.ENABLE_ENHANCED_OUTPUT_HANDLING]: boolean;
-  [Flag.NEW_FLOW_EDITOR]: boolean;
-  [Flag.BUILDER_VIEW_SWITCH]: boolean;
-  [Flag.SHARE_EXECUTION_RESULTS]: boolean;
-  [Flag.AGENT_FAVORITING]: boolean;
-  [Flag.MARKETPLACE_SEARCH_TERMS]: string[];
-  [Flag.ENABLE_PLATFORM_PAYMENT]: boolean;
-  [Flag.CHAT]: boolean;
-};
-
 const isPwMockEnabled = process.env.NEXT_PUBLIC_PW_TEST === "true";
 
-const mockFlags = {
+const defaultFlags = {
   [Flag.BETA_BLOCKS]: [],
   [Flag.NEW_BLOCK_MENU]: false,
   [Flag.NEW_AGENT_RUNS]: false,
@@ -50,17 +36,16 @@ const mockFlags = {
   [Flag.CHAT]: false,
 };
 
-export function useGetFlag<T extends Flag>(flag: T): FlagValues[T] | null {
+type FlagValues = typeof defaultFlags;
+
+export function useGetFlag<T extends Flag>(flag: T): FlagValues[T] {
   const currentFlags = useFlags<FlagValues>();
   const flagValue = currentFlags[flag];
+  const areFlagsEnabled = environment.areFeatureFlagsEnabled();
 
-  const envEnabled = process.env.NEXT_PUBLIC_LAUNCHDARKLY_ENABLED === "true";
-  const clientId = process.env.NEXT_PUBLIC_LAUNCHDARKLY_CLIENT_ID;
-  const isLaunchDarklyConfigured = envEnabled && Boolean(clientId);
-
-  if (!isLaunchDarklyConfigured || isPwMockEnabled) {
-    return mockFlags[flag];
+  if (!areFlagsEnabled || isPwMockEnabled) {
+    return defaultFlags[flag];
   }
 
-  return flagValue ?? mockFlags[flag];
+  return flagValue ?? defaultFlags[flag];
 }

classic/frontend/.gitignore

@@ -8,6 +8,7 @@
 .buildlog/
 .history
 .svn/
+.next/
 migrate_working_dir/
 
 # IntelliJ related

classic/frontend/build/web/flutter_service_worker.js

@@ -3,45 +3,45 @@ const MANIFEST = 'flutter-app-manifest';
 const TEMP = 'flutter-temp-cache';
 const CACHE_NAME = 'flutter-app-cache';
 
-const RESOURCES = {"canvaskit/skwasm.worker.js": "51253d3321b11ddb8d73fa8aa87d3b15",
-"canvaskit/skwasm.js": "95f16c6690f955a45b2317496983dbe9",
-"canvaskit/canvaskit.wasm": "d9f69e0f428f695dc3d66b3a83a4aa8e",
-"canvaskit/skwasm.wasm": "d1fde2560be92c0b07ad9cf9acb10d05",
-"canvaskit/canvaskit.js": "5caccb235fad20e9b72ea6da5a0094e6",
-"canvaskit/chromium/canvaskit.wasm": "393ec8fb05d94036734f8104fa550a67",
-"canvaskit/chromium/canvaskit.js": "ffb2bb6484d5689d91f393b60664d530",
-"icons/Icon-maskable-192.png": "c457ef57daa1d16f64b27b786ec2ea3c",
-"icons/Icon-maskable-512.png": "301a7604d45b3e739efc881eb04896ea",
+const RESOURCES = {"flutter.js": "6fef97aeca90b426343ba6c5c9dc5d4a",
 "icons/Icon-512.png": "96e752610906ba2a93c65f8abe1645f1",
+"icons/Icon-maskable-512.png": "301a7604d45b3e739efc881eb04896ea",
 "icons/Icon-192.png": "ac9a721a12bbc803b44f645561ecb1e1",
+"icons/Icon-maskable-192.png": "c457ef57daa1d16f64b27b786ec2ea3c",
 "manifest.json": "0fa552613b8ec0fda5cda565914e3b16",
-"favicon.png": "5dcef449791fa27946b3d35ad8803796",
-"version.json": "46a52461e018faa623d9196334aa3f50",
-"index.html": "e6981504a32bf86f892909c1875df208",
-"/": "e6981504a32bf86f892909c1875df208",
-"main.dart.js": "6fcbf8bbcb0a76fae9029f72ac7fbdc3",
-"assets/AssetManifest.json": "1b1e4a4276722b65eb1ef765e2991840",
-"assets/packages/cupertino_icons/assets/CupertinoIcons.ttf": "055d9e87e4a40dbf72b2af1a20865d57",
-"assets/packages/fluttertoast/assets/toastify.js": "56e2c9cedd97f10e7e5f1cebd85d53e3",
-"assets/packages/fluttertoast/assets/toastify.css": "a85675050054f179444bc5ad70ffc635",
+"index.html": "3442c510a9ea217672c82e799ae070f7",
+"/": "3442c510a9ea217672c82e799ae070f7",
 "assets/shaders/ink_sparkle.frag": "f8b80e740d33eb157090be4e995febdf",
-"assets/fonts/MaterialIcons-Regular.otf": "245e0462249d95ad589a087f1c9f58e1",
-"assets/assets/images/twitter_logo.png": "af6c11b96a5e732b8dfda86a2351ecab",
-"assets/assets/images/discord_logo.png": "0e4a4162c5de8665a7d63ae9665405ae",
-"assets/assets/images/google_logo.svg.png": "0e29f8e1acfb8996437dbb2b0f591f19",
-"assets/assets/images/autogpt_logo.png": "6a5362a7d1f2f840e43ee259e733476c",
-"assets/assets/images/github_logo.svg.png": "ba087b073efdc4996b035d3a12bad0e4",
-"assets/assets/scrape_synthesize_tree_structure.json": "a9665c1b465bb0cb939c7210f2bf0b13",
+"assets/assets/tree_structure.json": "cda9b1a239f956c547411efad9f7c794",
 "assets/assets/coding_tree_structure.json": "017a857cf3e274346a0a7eab4ce02eed",
 "assets/assets/general_tree_structure.json": "41dfbcdc2349dcdda2b082e597c6d5ee",
-"assets/assets/google_logo.svg.png": "0e29f8e1acfb8996437dbb2b0f591f19",
-"assets/assets/tree_structure.json": "cda9b1a239f956c547411efad9f7c794",
-"assets/assets/data_tree_structure.json": "5f9627548304155821968182f3883ca7",
 "assets/assets/github_logo.svg.png": "ba087b073efdc4996b035d3a12bad0e4",
+"assets/assets/images/discord_logo.png": "0e4a4162c5de8665a7d63ae9665405ae",
+"assets/assets/images/github_logo.svg.png": "ba087b073efdc4996b035d3a12bad0e4",
+"assets/assets/images/twitter_logo.png": "af6c11b96a5e732b8dfda86a2351ecab",
+"assets/assets/images/google_logo.svg.png": "0e29f8e1acfb8996437dbb2b0f591f19",
+"assets/assets/images/autogpt_logo.png": "6a5362a7d1f2f840e43ee259e733476c",
+"assets/assets/google_logo.svg.png": "0e29f8e1acfb8996437dbb2b0f591f19",
+"assets/assets/scrape_synthesize_tree_structure.json": "a9665c1b465bb0cb939c7210f2bf0b13",
+"assets/assets/data_tree_structure.json": "5f9627548304155821968182f3883ca7",
+"assets/fonts/MaterialIcons-Regular.otf": "245e0462249d95ad589a087f1c9f58e1",
 "assets/NOTICES": "28ba0c63fc6e4d1ef829af7441e27f78",
-"assets/AssetManifest.bin": "791447d17744ac2ade3999c1672fdbe8",
+"assets/packages/fluttertoast/assets/toastify.css": "a85675050054f179444bc5ad70ffc635",
+"assets/packages/fluttertoast/assets/toastify.js": "56e2c9cedd97f10e7e5f1cebd85d53e3",
+"assets/packages/cupertino_icons/assets/CupertinoIcons.ttf": "055d9e87e4a40dbf72b2af1a20865d57",
 "assets/FontManifest.json": "dc3d03800ccca4601324923c0b1d6d57",
-"flutter.js": "6fef97aeca90b426343ba6c5c9dc5d4a"};
+"assets/AssetManifest.bin": "791447d17744ac2ade3999c1672fdbe8",
+"assets/AssetManifest.json": "1b1e4a4276722b65eb1ef765e2991840",
+"canvaskit/chromium/canvaskit.wasm": "393ec8fb05d94036734f8104fa550a67",
+"canvaskit/chromium/canvaskit.js": "ffb2bb6484d5689d91f393b60664d530",
+"canvaskit/skwasm.worker.js": "51253d3321b11ddb8d73fa8aa87d3b15",
+"canvaskit/skwasm.js": "95f16c6690f955a45b2317496983dbe9",
+"canvaskit/canvaskit.wasm": "d9f69e0f428f695dc3d66b3a83a4aa8e",
+"canvaskit/canvaskit.js": "5caccb235fad20e9b72ea6da5a0094e6",
+"canvaskit/skwasm.wasm": "d1fde2560be92c0b07ad9cf9acb10d05",
+"favicon.png": "5dcef449791fa27946b3d35ad8803796",
+"version.json": "46a52461e018faa623d9196334aa3f50",
+"main.dart.js": "6fcbf8bbcb0a76fae9029f72ac7fbdc3"};
 // The application shell files that are downloaded before a service worker can
 // start.
 const CORE = ["main.dart.js",

classic/frontend/build/web/index.html

@@ -35,7 +35,7 @@
 
   <script>
     // The value below is injected by flutter build, do not touch.
-    const serviceWorkerVersion = "726743092";
+    const serviceWorkerVersion = "1550046101";
   </script>
   <!-- This script adds the flutter initialization JS code -->
   <script src="flutter.js" defer></script>