ToolCallAdvisor causes infinite loop when a tool returns long descriptive text on invalid input (empty string parameter)

When a tool is called with an empty string for a required parameter (e.g., inputData = ""), and the tool implementation returns a long descriptive “task prompt” instead of a concise error message, the ToolCallAdvisor enters an infinite loop. The model keeps calling the same tool with the same empty argument, consuming excessive tokens and never terminating.

Expected behavior
The ToolCallAdvisor should detect repeated identical tool calls (same name + same arguments) and break the loop after a configurable number of attempts, or at least allow the model to receive a clear error that stops the cycle.

Actual behavior
The advisor continues indefinitely, because the tool’s output (a long text that looks like a valid response) is interpreted as a normal tool result, and the model decides to call the same tool again with the same empty input.

Steps to Reproduce

1. Define a tool that requires a non‑empty inputData parameter:

   @Tool(description = "Analyzes data")
   public String analyze(@ToolParam(required = true) String inputData) {
       if (inputData == null || inputData.trim().isEmpty()) {
           // ❌ Returning a long descriptive text
           return "## Role\nYou are an expert...\n## Task\nPlease provide data...";
       }
       return "analysis result";
   }
   

2. Configure ChatClient with ToolCallAdvisor:

   var advisor = ToolCallAdvisor.builder()
       .toolCallingManager(toolCallingManager)
       .build();
   var chatClient = ChatClient.builder(chatModel)
       .defaultAdvisors(advisor)
       .defaultTools(analyzeTool)
       .build();
   

3. Ask a question that leads the model to call the tool with an empty inputData (e.g., the model hallucinates the required data).
4. Observe infinite tool‑call loop in logs.

Environment

• Spring AI version: 2.0.0-M2
• Java version: 17
• LLM provider: Doubao / OpenAI (any)
• Spring Boot version: 3.2.x

Logs (excerpt)

2026-04-03 17:34:52.476 ... "tool_calls":[{"function":{"arguments":" {\"inputData\": \"\", ...", "name":"audioChronicleInsight"}}]
2026-04-03 17:34:52.600 ... Successful execution of tool: audioChronicleInsight
... (repeats many times with same empty inputData)

Possible workarounds

• Validate input in the tool and throw an exception or return a short, clear error message like "ERROR: inputData cannot be empty. Please call searchAudioRecordings first."
• Manually implement a loop counter in a custom CallAdvisor.

Suggested improvements

• Provide a built‑in mechanism in ToolCallAdvisor to limit the number of identical consecutive tool calls (same name + same arguments).
• Add a configuration option like .maxIdenticalToolCalls(int) (default 2 or 3).
• Allow users to specify a ToolCallValidator that can reject a tool call before execution.

Labels

type: bug
area: tool-calling
priority: high
status: triage-needed

Start by reading ToolCallAdvisor and its builder, then trace how repeated tool calls are processed through the toolCallingManager. Define completion as detecting identical tool name and arguments, stopping after a configurable limit, and returning a clear error; verify the reproduced empty-input loop terminates.