[BUG]: GPT-5 models cannot use tools #4385
Description
How are you running AnythingLLM?
AnythingLLM desktop app
What happened?
GPT-5 family models does not support message.role = "function". AnythingLLM uses it and the request to OpenAI crashes with 400 Unsupported value.
Here are logs:
[backend] info: [AgentHandler] [debug]: @agent is attempting to call `web-scraping` tool
[backend] info: [EncryptionManager] Loaded existing key & salt for encrypting arbitrary data.
[collector] info: -- Working URL https://docs.anythingllm.com/installation-desktop/debug => (text) --
[Electron:Collector] Page Content loaded - cleaning up timers
[Electron:Collector] Scraping page content as text
[collector] info: Cleaning up request handler for request ID.
[backend] info: [TokenManager] Initialized new TokenManager instance for model: gpt-5-mini
[backend] error: error: 400 Unsupported value: 'messages[2].role' does not support 'function' with this model. Error: 400 Unsupported value: 'messages[2].role' does not support 'function' with this model.
at r.complete (/Applications/AnythingLLM.app/Contents/Resources/backend/server.js:313:8244)
at process.processTicksAndRejections (node:internal/process/task_queues:95:5)
at async Dl.handleExecution (/Applications/AnythingLLM.app/Contents/Resources/backend/server.js:332:406)
at async Dl.handleExecution (/Applications/AnythingLLM.app/Contents/Resources/backend/server.js:332:1345)
at async Dl.reply (/Applications/AnythingLLM.app/Contents/Resources/backend/server.js:332:274)
at async Dl.chat (/Applications/AnythingLLM.app/Contents/Resources/backend/server.js:313:46286)
at async Dl.start (/Applications/AnythingLLM.app/Contents/Resources/backend/server.js:313:45720)
at async aK (/Applications/AnythingLLM.app/Contents/Resources/backend/server.js:369:2717)
at async /Applications/AnythingLLM.app/Contents/Resources/backend/server.js:369:2243
[backend] info: [AgentHandler] End 4292d7b4-6a9e-477e-af0b-6362b6fa756a::openai:gpt-5-mini
This discussion suggests that GPT-5 family models changed the API and are now using "tool" as message type for tool calls, not "function".
Are there known steps to reproduce?
- Set GPT-5 / GPT-5-mini (I guess even GPT-5-nano) as model
- Send this message:
@agent Collect the text of this page: https://docs.anythingllm.com/installation-desktop/debug(or any other tool call)
(Older models like GPT-4.1 work fine.)