Addition of safe MCP Tools breaking the chat responsiveness after one use.

Nobody has claimed this yet.

## MCP Tool Stuck Repeatedly Responding to First Message Only

### Which version of LM Studio?
LM Studio 0.3.24

### Which operating system?
Windows

### What is the bug?
When using MCP servers (e.g., PostgreSQL MCP), after the first successful tool call, the model becomes stuck in a loop where it only responds to the FIRST message/query, regardless of what new messages are sent. Every subsequent message triggers the MCP tool to re-execute the original first query instead of processing the new input.

### Expected Behavior
- Each new message should be processed independently
- MCP tools should respond to the current query, not repeat the first one
- The chat context should properly update after each interaction

### Actual Behavior
1. First message with MCP tool call works correctly (e.g., "List all tables")
2. Tool executes and returns correct results for first query
3. Send a completely different second message (e.g., "Count users in database")
4. MCP tool executes but returns the SAME results from the first query (table list)
5. All subsequent messages trigger the first query's response
6. The model appears unable to "move on" from the initial tool call
7. Chat is effectively stuck in a loop responding to a ghost of the first message

### Steps to Reproduce
1. Configure PostgreSQL MCP server in `mcp.json`
2. Load a model with tool calling support
3. Start a new chat
4. First message: "Show me all tables in my database"
- Result: Correctly lists tables
5. Second message: "How many users are in the users table?"
- Result: Shows table list again (first query response)
6. Third message: "What is 2+2?" (non-database query)
- Result: Either shows table list again OR model finally responds but ignores MCP tools
7. Any subsequent database query continues returning the first query's results

No source files, tests, or entry points are identified; begin by reproducing the flow with the PostgreSQL MCP server configured in mcp.json, using the listed three-message sequence. Done means each later message is processed independently and does not re-execute the first tool query.

Written by the indexing model from the issue text.