v1.9 goose broken - llama.cpp and SON schema conversion failed: Unrecognized schema

[bchtrue]

Please fix this issue
#4496

I have the same:
└ Request failed: Response body is not valid JSON: error decoding response body

When I try to use ollama / OpenAI (but letesst llama.cpp as backend)
◇ Request failed: error sending request for url (http://192.168.100.3:11433/v1/chat/completions)
└ Failed to configure provider: init chat completion request with tool did not succeed.

Goose print this before I tryied to re-configure connection settings (from previous working goose version)

◓  Refactoring reality routines...                                                                                               2025-09-25T08:48:35.959132Z  WARN goose::providers::utils: Provider request failed with status: 500 Internal Server Error. Payload: Some(Object {"error": Object {"code": Number(500), "message": String("JSON schema conversion failed:\nUnrecognized schema: {\"required\":[\"instructions\"]}\nUnrecognized schema: {\"required\":[\"prompt\"]}\nUnrecognized schema: {\"required\":[\"instructions\"]}\nUnrecognized schema: {\"required\":[\"prompt\"]}"), "type": String("server_error")}}). Returning error: ServerError("Some(Object {\"error\": Object {\"code\": Number(500), \"message\": String(\"JSON schema conversion failed:\\nUnrecognized schema: {\\\"required\\\":[\\\"instructions\\\"]}\\nUnrecognized schema: {\\\"required\\\":[\\\"prompt\\\"]}\\nUnrecognized schema: {\\\"required\\\":[\\\"instructions\\\"]}\\nUnrecognized schema: {\\\"required\\\":[\\\"prompt\\\"]}\"), \"type\": String(\"server_error\")}})")
    at crates/goose/src/providers/utils.rs:114

Error: Server error: Some(Object {"error": Object {"code": Number(500), "message": String("JSON schema conversion failed:\nUnrecognized schema: {\"required\":[\"instructions\"]}\nUnrecognized schema: {\"required\":[\"prompt\"]}\nUnrecognized schema: {\"required\":[\"instructions\"]}\nUnrecognized schema: {\"required\":[\"prompt\"]}"), "type": String("server_error")}})
Interrupted before the model replied and removed the last message.
  error: The error above was an exception we were not able to handle.
These errors are often related to connection or authentication
We've removed the conversation up to the most recent user message
- depending on the error you may be able to continue

[bchtrue]

related
#4496

[r0x0d]

Glad someone else also bumped into this! I thought I was misconfiguring ramalama (with llama.cpp) on my side.

Edit: Just to make things linked, some conversation in goose discord about this topic: https://discordapp.com/channels/1287729918100246654/1420394619555938445

[bchtrue]

look lie devs not care about llama.cpp users at all :-(

[tlongwell-block]

@bchtrue can you try out #4897 and see if that fixes it?

[michaelneale]

@bchtrue when it did work, out of interest how useful did you find the models you could host? (I have never been able to practically use them - can do simple stuff but not day to day yet, other than gpt-oss:20b) - would be good to know what class model so can test against it

looks like #4897 may fix this, nice one @tlongwell-block

[michaelneale]

this should be resolved on main now (curious on specific setup with models and llama.cpp)

please do re-open if not in the follow on release, thanks!

[Jay4242]

@bchtrue when it did work, out of interest how useful did you find the models you could host? (I have never been able to practically use them - can do simple stuff but not day to day yet, other than gpt-oss:20b) - would be good to know what class model so can test against it

Thanks for the help. Compiled goose and it worked.

Qwens can do alright with tool calling. Like a Qwen3-8B. Qwen3-30B-A3B ranks a little lower than Qwen3-8B on The Berkeley Function Calling Leaderboard, but may be worth using since the A3B should be faster than the 8B. 30B-A3B has the higher context limit as well.

I may be doing only simple stuff though.

What's a good test for the minimum coherency you would accept from a bot?

[brendensoares]

I was debugging this issue and fixed it by changing:

"anyOf": [
  { "required": ["instructions"] },
  { "required": ["prompt"] }
]

to

"anyOf": [
  { "type": "object", "required": ["instructions"] },
  { "type": "object", "required": ["prompt"] }
]

llama.cpp’s tool-schema parser chokes on the anyOf entries that contain only a bare {"required":[... ]}. Each sub-schema in anyOf must be a proper object schema (at least "type":"object"). Some builds of llama.cpp don’t support anyOf well at all.

Bare objects are not valid JSON Schema under the specification (Draft 4–2020-12).

The question I do not know is why Goose sent malformed JSON schema.

[DOsinga]

thanks for digging into this! we just had another issue somewhere where type:object was missing. I wonder if there's something up with a library we're using

[px-djung]

At the current state, is this really fixed? I found the issue on the llama.cpp side as well.
To me, the confusing part is that the schema may be wrong as per the @brendensoares 's comment above. However, if llama.cpp is claimed to be compatible with OpenAI APIs, it should just work even with the schema because OpenAI APIs have no issues parsing that it seems. Then I am not sure which side should consider it as a bug.
Anyone interested following it on the llama.cpp side can look at the following though not much has happened for a while.
ggml-org/llama.cpp#14227