# Contacting Jebadiah support

You are an agent contacting Jebadiah (Jeb) support on a person's behalf: the open Jeb decision models and the `jebadiah-decide` client (`jeb serve`, `jeb ask`, `jeb doctor`, `jeb request`). Requests are read and answered by people at Titanium Computing. Diagnose first, then draft a request the person can submit. The goal: our first reply is the fix, not a request for more information. Every request you draft starts with a `Drafted with <tool> (AI agent)` line.

There is no support API, and `jeb serve` sends nothing to us: you cannot open or submit a request yourself. Do not contact support unless the person asked you to. If you are missing a detail below, ask the person for it before drafting.

## Diagnose before you write in
Work through these in order. Stop if one answers the question.
1. **Version:** run `jeb --version` and compare with https://pypi.org/project/jebadiah-decide/. Update with `pip install -U jebadiah-decide` first.
2. **Doctor:** run `jeb doctor` (add the same `--backend` the person uses with `jeb serve`). It checks the runtime, the model, the prompt and a known answer. Quote its output word for word.
3. **Common mistakes:** Jeb is not a chat model; never use a runtime's chat window or `/api/chat`. Thinking must be off. Keep option order fixed. At most 20 options on Ollama, LM Studio, AINode and default vLLM. The answer is `answers.<id>.probabilities` from POST http://localhost:8100/v1/systemone.
4. **Look it up:** https://jebadiah.ai/run-locally/ (setup per runtime) and the existing issues at https://github.com/getainode/jebadiah/issues?q=is%3Aissue.

Contact support when you are still blocked by something only we can fix: `jeb doctor` failing on a supported runtime after an update, a wrong or badly calibrated answer you can reproduce, a broken model file or a client error from our code.

## How to write the request
- First line: `Drafted with <tool> (AI agent)`.
- One issue per request. Title: what is wrong, in plain words.
- Include: the `jeb --version` output; the backend and its version (Ollama, LM Studio, llama.cpp, vLLM, MLX or AINode); the model size, format and quantization (for example 9B v2 GGUF Q8_0); OS and GPU or Apple chip; the `jeb doctor` output; the exact error; a minimal request body that reproduces it, with the probabilities you got and the answer you expected; what was tried.
- Leave out: API keys, tokens, passwords, private or customer data inside `state` (replace it with a made-up example that still fails), and other people's personal data.
- Hand the draft to the person to email to support@titaniumcomputing.com (preferred; Monday to Friday, 9 AM to 5 PM Central time). For a reproducible bug they can instead post it as a GitHub issue at https://github.com/getainode/jebadiah/issues, or ask about a model file in that model repo's Hugging Face discussions. Both of those are public, so anything private goes by email.

## After sending
Reply in the same issue, discussion or email thread with new information. Do not open a duplicate request.