Restores session-scoped fixtures and pytest_asyncio decorators that were
accidentally changed, causing "RuntimeError: Event loop is closed" in
test_authorize_creates_code_in_database. Also regenerates openapi.json.
### Changes 🏗️
- Added AI SDK integration for chat streaming with proper message
handling
- Implemented custom to_sse method in StreamToolOutputAvailable to
exclude non-spec fields
- Modified stream_chat_completion to reuse message IDs for tool call
continuations
- Created new Copilot 2.0 UI with AI SDK React components
- Added streamdown and related packages for markdown rendering
- Built reusable conversation and message components for the chat
interface
- Added support for tool output display in the chat UI
### Checklist 📋
#### For code changes:
- [x] I have clearly listed my changes in the PR description
- [x] I have made a test plan
- [x] I have tested my changes according to the test plan:
- [x] Start a new chat session and verify streaming works correctly
- [x] Test tool calls and verify they display properly in the UI
- [x] Verify message continuations don't create duplicate messages
- [x] Test markdown rendering with code blocks and other formatting
- [x] Verify the UI is responsive and scrolls correctly
#### For configuration changes:
- [x] `.env.default` is updated or already compatible with my changes
- [x] `docker-compose.yml` is updated or already compatible with my
changes
- [x] I have included a list of my configuration changes in the PR
description (under **Changes**)
---------
Co-authored-by: Lluis Agusti <hi@llu.lu>
Co-authored-by: Ubbe <hi@ubbe.dev>
When a credential is deleted but the node still references its ID,
CredentialsSelect now treats the stale ID as unselected and falls
back to the first available credential instead of showing the raw ID.
MCP credentials use per-server dynamic OAuth handlers, not a static
handler registered in HANDLERS_BY_NAME. The delete endpoint now
creates a dynamic handler from credential metadata for token
revocation instead of failing with "Provider 'mcp' does not support
OAuth".
- Validate token_endpoint in OAuth metadata before accessing it
- Check authorization_servers list is non-empty before indexing
- Use provider_matches() (renamed from private _provider_matches) in
creds_manager for Python 3.13 StrEnum compatibility
- Fill in MCP block documentation with technical explanation and use cases
- Set customized_name in metadata when MCP and Agent blocks are created
(both legacy and new builder) so titles persist through save/load
- Remove convoluted agent_name fallback from NodeHeader and getNodeTitle
- Add custom block-level validation in graph pre-run checks so MCP tool
arguments are validated before execution
- Fix server_name fallback to URL hostname in discover_tools endpoint
Backend: Add required-field validation in MCPToolBlock.run() before
calling the MCP server. The executor-level validation is bypassed for
MCP blocks because get_input_defaults() flattens tool_arguments,
stripping tool_input_schema from the validation context.
Frontend: NodeHeader now derives the MCP server label from the server
URL hostname when server_name is missing (pruned by pruneEmptyValues).
This fixes the title for existing blocks that don't have customized_name
in metadata.
When the MCP server returns a null server_name, fall back to the URL
hostname so customized_name is always set in metadata. This prevents
the title from degrading to "MCP:" after save and reload.
- CredentialsSelect: default to first available credential instead of
"None" when credentials exist, reorder options to show credentials
before the "None" option, and notify parent on auto-select
- Revert CredentialsGroupedView user auto-select effect (now handled
at the CredentialsSelect level)
- Block.tsx: persist MCP block title as customized_name in metadata
so it survives save/load
Add auto-selection for user credentials (like MCP OAuth) in the
CredentialsGroupedView run dialog. When exactly one credential matches
the provider, type, and discriminator values (e.g. MCP server URL),
it is pre-selected instead of defaulting to "None (skip this credential)".
Frontend: Include credentials field in MCP block's dynamic input schema
so users can select OAuth credentials from the node form. Separate
credentials from tool_arguments in FormCreator to store them at the
correct level in hardcodedValues.
Backend: Add _auto_lookup_credential fallback in MCPToolBlock.run() for
legacy nodes that don't have credentials explicitly set. This resolves
the credential by matching mcp_server_url in stored OAuth metadata.
verify_state_token and get_creds_by_provider compared provider strings
with ==, which failed when OAuth states were stored with the buggy
"ProviderName.MCP" format from Python 3.13's str(Enum) behavior.
Also fix double-append in store_state_token where the state was written
once via edit_user_integrations and again via a redundant manual block.
- Add missing `refresh_if_needed` mock to test_discover_tools_auto_uses_stored_credential
so it returns the stored credential instead of a MagicMock
- Fix credential auto-unselect clearing MCP credentials on initial render:
skip the "unselect if not available" check when the saved credentials
list is empty (empty list means not loaded yet, not invalid)
- Fix pyright errors in graph_test.py by properly typing frozenset[CredentialsType]
- Fix executor validation crash when credentials is empty {} by nullifying
the field before JSON schema validation
- Exclude MCP Tool block from e2e block discovery test (requires dialog)
- Normalize provider string in CredentialsMetaResponse to handle Python 3.13
str(Enum) bug for stored credentials
- Fix get_host() to match MCP provider regardless of enum string format
The discover_tools endpoint was reading raw access tokens from stored
credentials without checking if they had expired. This caused users
to be prompted to re-authenticate every time the token expired (~1h).
Now uses creds_manager.refresh_if_needed() to transparently refresh
expired tokens before using them.
The pytest_asyncio fixture changes with loop_scope="session" caused
"Event loop is closed" errors in all 31 oauth_test.py tests on CI.
MCP tests have their own conftest override and don't need these changes.
The tests used MagicMock for block.input_schema but didn't mock
get_required_fields(), causing the "required missing creds" test to
silently treat all credentials as optional.
Adds diagnostic logging when the `type vector does not exist` error
occurs in raw SQL queries.
## Problem
We're seeing intermittent "type vector does not exist" errors on
dev-behave ([Sentry
issue](https://significant-gravitas.sentry.io/issues/7205929979/)). The
pgvector extension should be in the search_path, but occasionally
queries fail to resolve the vector type.
## Solution
When a query fails with this specific error, we now log:
- `SHOW search_path` - what schemas are being searched
- `SELECT current_schema()` - the active schema
- `SELECT current_user, session_user, current_database()` - connection
context
This diagnostic info will help identify why the vector extension isn't
visible in certain cases.
## Changes
- Added `_log_vector_error_diagnostics()` helper function in
`backend/data/db.py`
- Wrapped SQL execution in try/except to catch and diagnose vector type
errors
- Original exception is re-raised after logging (no behavior change)
## Testing
This is observational/diagnostic code. It will be validated by waiting
for the error to occur naturally on dev and checking the logs.
## Rollout
Once we've captured diagnostic logs and identified the root cause, this
logging can be removed or reduced in verbosity.
- Auto-select credential when exactly one match exists (even for
optional fields). Only skip auto-select for optional fields with
multiple choices.
- In executor, catch ValueError from creds_manager.acquire() for
optional credential fields — fall back to running without credentials
instead of crashing when stale IDs reference deleted credentials.
The sortByUnsetFirst comparator in splitCredentialFieldsBySystem
caused credential inputs to jump positions every time a credential
was selected (set fields moved to bottom, unset moved to top).
Remove the sort to keep stable ordering.
The model_validator on CredentialsMetaInput normalizes legacy
"ProviderName.MCP" format for Pydantic validation, but validate_data()
uses raw JSON schema which bypasses Pydantic. Write normalized values
back to input_data after Pydantic processes them so both validation
paths see correct data.
- Add model_validator on CredentialsMetaInput to auto-normalize old
"ProviderName.MCP" format to "mcp" at the model level, eliminating
the need for string parsing hacks in every consumer.
- Fix aggregate_credentials_inputs to check block schema defaults when
determining if credentials are required, not just node metadata.
MCP blocks with default={} are always optional regardless of metadata.
Also search for credentials stored with "ProviderName.MCP" (from the
Python 3.13 str(Enum) bug) during both discover-tools auto-lookup and
OAuth callback cleanup. Remove the temporary debug endpoint.
Python 3.13 changed str(StrEnum) to return "ClassName.MEMBER" instead of
the plain value. This caused MCP credentials to be stored with provider
"ProviderName.MCP" instead of "mcp", leading to type/provider mismatch
errors during graph validation and execution.
Fix: Pass the enum directly to Pydantic (which extracts .value automatically),
matching the pattern used by all other OAuth handlers. Use .value explicitly
only in non-Pydantic contexts (string comparisons, API calls).
Python 3.13 changed str(StrEnum) to return "ClassName.MEMBER" instead of
the plain value. This caused MCP credentials to be stored with provider
"ProviderName.MCP" instead of "mcp", leading to type/provider mismatch
errors during graph validation and execution.
- _on_graph_activate: Clear stale credential references for optional
fields instead of blocking the save. Checks both node metadata
(credentials_optional) and block schema (field not in required_fields).
- _validate_node_input_credentials: Use block schema's required_fields
as fallback for credentials_optional check, so MCP blocks with
default={} credentials are properly treated as optional.
- Set credentials_optional metadata on new MCP nodes in the frontend.
Prevent MCP credential cross-contamination where a credential for one
server (e.g. Sentry) fills credential fields for other servers (e.g.
Linear). Adds matchesDiscriminatorValues() to match credentials by host
against discriminator_values from the schema.
- Extract shared OAuth popup utility (oauth-popup.ts) used by both
MCPToolDialog and useCredentialsInput, eliminating ~200 lines of
duplicated BroadcastChannel/postMessage/localStorage listener code
- Add mcpOAuthCallback to credentials provider so MCP credentials
are added to the in-memory cache after OAuth (fixes credentials not
appearing in the credential picker after OAuth via MCPToolDialog)
- Fix oauth_test.py async fixtures missing loop_scope="session"
- Add MCP token refresh handler in creds_manager for dynamic endpoints
- Fix enum string representation in CredentialsFieldInfo.combine()
- Replace manual credential_id field with CredentialsMetaInput pattern
- Fix credential deduplication so different MCP server URLs get separate
credential entries in the task credentials panel
- Add descriptive display names (e.g. "MCP: mcp.sentry.dev")
- Fix OAuth popup callback by adding mcp_callback route to middleware
exclusion list and adding localStorage polling fallback
- Fix SSRF test fixture to patch Requests constructor directly
- Add MCP server URL matching for credential auto-assignment
- Return CredentialsMetaResponse from MCP OAuth callback
- Support MCP-specific OAuth flow in frontend credential input
- Filter MCP credentials by server URL in frontend
- Add test coverage for credential deduplication logic
- Use isinstance(creds, APIKeyCredentials) instead of hasattr check
- Rewrite integration tests to use user_id param and mock _resolve_auth_token
- Fix f-string and line-length formatting issues in routes.py
The single-pass regex `/<[^>]+>/g` can be bypassed with nested tags
like `<scr<script>ipt>`. Loop until no more tags are found.
Note: React auto-escapes JSX so this is defense-in-depth.
User-provided MCP server URLs should not bypass SSRF IP-blocking
validation. Remove trusted_origins from all MCP code so that
private/internal IPs are properly blocked. Keep ThreadedResolver
in HostResolver fallback for DNS reliability in subprocess
environments.
- Support MCP servers that serve OAuth metadata directly without
protected-resource metadata (e.g. Linear) by falling back to
discover_auth_server_metadata on the server's own origin
- Omit resource_url when no protected-resource metadata exists to
avoid token audience mismatch errors (RFC 8707 resource is optional)
- Add Mcp-Session-Id header tracking per MCP Streamable HTTP spec
- Fall back to server_url credential lookup when credential_id is
empty (pruneEmptyValues strips it from saved graphs)
- Use ThreadedResolver instead of c-ares AsyncResolver to avoid DNS
failures in forked subprocess environments
- Simplify OAuth UX: single "Sign in & Connect" button on 401,
remove sticky localStorage URL prefill
- Clean up stale MCP credentials on re-authentication
[SECRT-1912: Investigate & eliminate chat session start
latency](https://linear.app/autogpt/issue/SECRT-1912)
### Changes 🏗️
- Add timing logs to `backend.api.features.chat` in `routes.py`,
`service.py`, and `stream_registry.py`
- Remove unneeded DB join in `create_chat_session`
### Checklist 📋
#### For code changes:
- [x] I have clearly listed my changes in the PR description
- [x] I have made a test plan
- [x] I have tested my changes according to the test plan:
- CI checks
BroadcastChannel can silently fail in some browser scenarios. Added:
- localStorage as third communication method in callback page
- storage event listener in dialog
- Popup close detection that checks localStorage directly
- Cleaned up auth-required box styling (gray instead of amber)
_resolve_auth_token now checks token expiry and refreshes using
MCPOAuthHandler with metadata (token_url, client_id, client_secret)
stored during the OAuth callback flow.
The dynamic get_input_defaults returns only tool_arguments, so the
execution engine loses block-level fields like server_url. Reconstruct
the full Input from node.input_default and set tool_arguments from the
resolved dynamic input, matching the AgentExecutorBlock pattern.
The block framework's CredentialsField requires credentials to always be
present, which doesn't work for public MCP servers. Replace it with a
plain credential_id field and manual resolution from the credential store,
allowing both authenticated and public MCP servers to work seamlessly.
- Remove credentials field from MCP dynamic schema since auth is handled
by the dialog's OAuth flow (the standard credentials widget doesn't
support MCP as a provider and fails with 404)
- Simplify FormCreator MCP handling — all form fields are tool arguments
- Disable auto-connect on dialog open; pre-fill last URL instead so user
can edit before connecting
Fixes
[#11800](https://github.com/Significant-Gravitas/AutoGPT/issues/11800)
## Problem
The FileInput component crashed with `TypeError: e.startsWith is not a
function` when the value was an object (from external API) instead of a
string.
## Example Input Object
When using the external API
(`/external-api/v1/graphs/{id}/execute/{version}`), file inputs can be
passed as objects:
```json
{
"node_input": {
"input_image": {
"name": "image.jpeg",
"type": "image/jpeg",
"size": 131147,
"data": "/9j/4QAW..."
}
}
}
```
## Changes
- Updated `getFileLabelFromValue()` to handle object format: `{ name,
type, size, data }`
- Added type guards for string vs object values
- Graceful fallback for edge cases (null, undefined, empty object)
## Test cases verified
- Object with name: returns filename
- Object with type only: extracts and formats MIME type
- String data URI: parses correctly
- String file path: extracts extension
- Edge cases: returns "File" fallback
### Changes 🏗️
Fixes
[**AUTOGPT-SERVER-1TN**](https://autoagpt.sentry.io/issues/?query=AUTOGPT-SERVER-1TN)
(~39K events since Feb 2025) and related connection issues
**6JC/6JD/6JE/6JF** (~6K combined).
#### Problem
When the RabbitMQ TCP connection drops (network blip, server restart,
etc.):
1. `connect_robust` (aio_pika) automatically reconnects the underlying
AMQP connection
2. But `AsyncRabbitMQ._channel` still references the **old dead
channel**
3. `is_ready` checks `not self._channel.is_closed` — but the channel
object doesn't know the transport is gone
4. `publish_message` tries to use the stale channel →
`ChannelInvalidStateError: No active transport in channel`
5. `@func_retry` retries 5 times, but each retry hits the same stale
channel (it passes `is_ready`)
This means every connection drop generates errors until the process is
restarted.
#### Fix
**New `_ensure_channel()` helper** that resets stale channels before
reconnecting, so `connect()` creates a fresh one instead of
short-circuiting on `is_connected`.
**Explicit `ChannelInvalidStateError` handling in `publish_message`:**
1. First attempt uses `_ensure_channel()` (handles normal staleness)
2. If publish throws `ChannelInvalidStateError`, does a full reconnect
(resets both `_channel` and `_connection`) and retries once
3. `@func_retry` provides additional retry resilience on top
**Simplified `get_channel()`** to use the same resilient helper.
**1 file changed, 62 insertions, 24 deletions.**
#### Impact
- Eliminates ~39K `ChannelInvalidStateError` Sentry events
- RabbitMQ operations self-heal after connection drops without process
restart
- Related transport EOF errors (6JC/6JD/6JE/6JF) should also reduce
- Add MCPToolDialog with tool discovery, OAuth sign-in, and card-based tool selection
- Add OAuth callback route using BroadcastChannel API for popup communication
- Add API client methods for MCP discovery, OAuth login, and callback
- Register MCP API routes on the backend REST API
- Render dynamic input schema for MCP blocks (credentials + tool params)
in both legacy and new builder CustomNode components
- Nest MCP tool argument values under tool_arguments in hardcodedValues
- Display tool name with server name prefix in block header
- Add backend route tests for discovery, OAuth login, and callback endpoints
Filters out blocks that are unsuitable for standalone execution from
CoPilot's block search and execution. These blocks serve graph-specific
purposes and will either fail, hang, or confuse users when run outside
of a graph context.
**Important:** This does NOT affect the Builder UI which uses
`load_all_blocks()` directly.
### Changes 🏗️
- **find_block.py**: Added `EXCLUDED_BLOCK_TYPES` and
`EXCLUDED_BLOCK_IDS` constants, skip excluded blocks in search results
- **run_block.py**: Added execution guard that returns clear error
message for excluded blocks
- **content_handlers.py**: Added filtering to
`BlockHandler.get_missing_items()` and `get_stats()` to prevent indexing
excluded blocks
**Excluded by BlockType:**
| BlockType | Reason |
|-----------|--------|
| `INPUT` | Graph interface definition - data enters via chat, not graph
inputs |
| `OUTPUT` | Graph interface definition - data exits via chat, not graph
outputs |
| `WEBHOOK` | Wait for external events - would hang forever in CoPilot |
| `WEBHOOK_MANUAL` | Same as WEBHOOK |
| `NOTE` | Visual annotation only - no runtime behavior |
| `HUMAN_IN_THE_LOOP` | Pauses for human approval - CoPilot IS
human-in-the-loop |
| `AGENT` | AgentExecutorBlock requires graph context - use `run_agent`
tool instead |
**Excluded by ID:**
| Block | Reason |
|-------|--------|
| `SmartDecisionMakerBlock` | Dynamically discovers downstream blocks
via graph topology |
### Checklist 📋
#### For code changes:
- [x] I have clearly listed my changes in the PR description
- [x] I have made a test plan
- [x] I have tested my changes according to the test plan:
- [ ] Search for "input" in CoPilot - should NOT return AgentInputBlock
variants
- [ ] Search for "output" in CoPilot - should NOT return
AgentOutputBlock
- [ ] Search for "webhook" in CoPilot - should NOT return trigger blocks
- [ ] Search for "human" in CoPilot - should NOT return
HumanInTheLoopBlock
- [ ] Search for "decision" in CoPilot - should NOT return
SmartDecisionMakerBlock
- [ ] Verify functional blocks still appear (e.g., "email", "http",
"text")
- [ ] Verify Builder UI still shows ALL blocks (no regression)
#### For configuration changes:
- [x] `.env.default` is updated or already compatible with my changes
- [x] `docker-compose.yml` is updated or already compatible with my
changes
- [x] I have included a list of my configuration changes in the PR
description (under **Changes**)
No configuration changes required.
---
Resolves: [SECRT-1831](https://linear.app/autogpt/issue/SECRT-1831)
🤖 Generated with [Claude Code](https://claude.ai/code)
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> **Low Risk**
> Behavior change is limited to CoPilot’s block discovery/execution
guards and is covered by new tests; main risk is inadvertently excluding
a block that should be runnable.
>
> **Overview**
> CoPilot now **filters out graph-only blocks** from `find_block`
results and prevents them from being executed via `run_block`, returning
a clear error when a user attempts to run an excluded block.
>
> `find_block` introduces explicit exclusion lists (by `BlockType` and a
specific block ID), over-fetches search results to maintain up to 10
usable matches after filtering, and adds debug logging when results are
reduced. New unit tests cover both the search filtering and the
`run_block` execution guard; a minor cleanup removes an unused `pytest`
import in `execution_queue_test.py`.
>
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
bc50755dcf. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: Nicholas Tindle <ntindle@users.noreply.github.com>
Co-authored-by: Otto <otto@agpt.co>
