Using Flockdeck

Running Ornith models

Ornith is a family of open-weight coding models from DeepReinforce, released under the MIT license. Version 1.5 comes as a 9B dense model, a 35B mixture of experts (35B-A3B, about 3B active) and a 397B mixture of experts. The model's source and serving notes are at https://github.com/ornith-ai/Ornith-1.

Flockdeck has no Ornith-specific support. Ornith is served behind an OpenAI-compatible API with tool calling, and that is what the OpenAI-compatible endpoint agent talks to (see Agents and models). This setup has been checked against a stand-in OpenAI-compatible endpoint, not a live Ornith server. There, a model id containing a slash was sent unchanged, tools were offered, the reasoning stream showed as thinking, and a tool call ran and its result came back. How a real Ornith model calls tools, and how its pane's status reads, are untested.

What the server has to do

The chat client sends tools with every request and reads the reasoning stream from reasoning_content or reasoning. If the server does not parse these, tool calls and thinking arrive as plain text in the answer and the agent cannot act on them. Ornith's README serves it with both parsers on. For vLLM (0.19.1 or later, per the README) that is:

vllm serve ornith-ai/Ornith-1.5-9B \
    --served-model-name Ornith-1.5-9B \
    --port 8000 \
    --enable-auto-tool-choice --tool-call-parser qwen3_xml \
    --reasoning-parser qwen3 \
    --trust-remote-code

The README's recipe includes --trust-remote-code; it does not say whether the flag is required. The 35B-A3B and 397B models need more GPUs, so add --tensor-parallel-size and a context length for your hardware, as the README does. SGLang (0.5.9 or later) uses --tool-call-parser qwen3_coder --reasoning-parser qwen3. Take the full flags from the README. Its serve commands name the repository deepreinforce-ai/..., but the checkpoints are published under ornith-ai/, so check the id on Hugging Face before serving.

For Ollama, the model card for the 1.5 9B GGUF (https://huggingface.co/ornith-ai/Ornith-1.5-9B-GGUF) gives ollama run ornith-1.5:9b, and llama-server -hf hf.co/ornith-ai/Ornith-1.5-9B-GGUF --port 8000 -c 262144 for llama.cpp. The main README still shows the 1.0 build under deepreinforce-ai. Ollama may report no tool support for the model. A model Ollama says cannot use tools is retried without them, and then it cannot edit files or run commands.

The model id you give Flockdeck is whatever the server calls the model: the --served-model-name above, or the id a hosted provider lists. A slash in the id is sent as it is. /models on the endpoint shows what it calls them.

Using the built-in endpoint agent

Give the OpenAI-compatible endpoint agent the server's address, then choose a model.

flockdeck keys endpoint openai-compatible http://127.0.0.1:8000/v1

You can do the same from the picker: choose OpenAI-compatible endpoint and type the address. An address on this machine needs no key. For a hosted provider, use its documented base URL and set the key with flockdeck keys set openai-compatible.

The built-in agent lists no models. In a pane, /model lists what the endpoint's /models reports, and /model <number> or /model <id> switches. A chat started with no model on a local server that offers exactly one picks that one.

Naming the model once

To have Ornith in the picker with a default model, add an agent of your own to agents.json (see Adding your own):

{
  "version": 1,
  "agents": [
    {
      "id": "ornith",
      "name": "Ornith 9B",
      "runner": "api",
      "api": {
        "wire": "openai",
        "baseURL": "http://127.0.0.1:8000/v1"
      },
      "models": [
        { "id": "Ornith-1.5-9B", "name": "Ornith 1.5 9B", "tier": "small" }
      ]
    }
  ]
}

One vLLM or SGLang process serves one model, and an agent has one address. To use a second model, such as the 35B-A3B, run a second server on another port and add a second entry with its own id, baseURL and model. For a hosted provider that serves several models at one address, one entry can list them all. Leave keyEnv out for a server on this machine; for a hosted one add "keyEnv": ["ORNITH_API_KEY"] or set the key with flockdeck keys set.

Because the address is not the vendor's own, these panes show tokens and no dollar cost.

Limits

Flockdeck does not set sampling parameters. Ornith's README recommends a temperature of 0.6 for precise coding work and 1.0 for general tasks, with top_k 20. Set those as defaults on the server, since the chat client sends only the model, messages and tools.

Only Ollama's "does not support tools" error triggers the retry without tools. vLLM started without --enable-auto-tool-choice refuses the request, and the error shows in the pane.