feat(plugins): Ollama support

[Dialvive]

Summary

Adds a new plugin that lets Code Puppy connect to local inference servers — Ollama, LM Studio, vLLM, llama.cpp, and any other OpenAI Chat Completions-compatible endpoint — without sending requests to remote model providers.

The plugin is auto-discovered by the existing plugin loader. Zero existing files were modified.

User Story

As a user, I want to run Code Puppy with local models through Ollama (or any OpenAI Chat Completions-compatible endpoint), so that I don't connect with remote model providers, use the hardware on my machine to process requests, and reduce costs.

Acceptance Criteria

• A new plugin code_puppy/plugins/ollama/ is created following the plugin architecture
• The plugin registers an "ollama" model type via the register_model_type callback hook
• Users can configure Ollama models in ~/.code_puppy/extra_models.json using the existing custom_endpoint config format
• When no custom_endpoint is provided, the plugin defaults to http://localhost:11434/v1 with api_key "ollama"
• The OLLAMA_HOST environment variable overrides the default base URL
• The handler creates OpenAIChatModel (Chat Completions API), NOT OpenAIResponsesModel (Responses API)
• All existing Code Puppy features (tools, agents, sub-agents, plugins, streaming) work with Ollama models that support tool/function calling
• The plugin fails gracefully — never crashes the app
• Unit tests cover all paths with 95%+ coverage
• No existing files are modified — plugin loader auto-discovers the new directory
• Code passes ruff format and ruff check

Test Plan

#	Test Case	Expected Result
1	Handler with custom_endpoint config	Uses provided URL, api_key, headers via get_custom_config()
2	Handler without custom_endpoint	Defaults to http://localhost:11434/v1, api_key "ollama"
3	OLLAMA_HOST env var set	Overrides default URL, appends /v1 if missing
4	Handler returns OpenAIChatModel	NOT OpenAIResponsesModel — verified via isinstance check
5	Handler failure (e.g., bad config)	Returns None, does not raise
6	get_ollama_model_types() return structure	Returns [{"type": "ollama", "handler": callable}]
7	ModelFactory.get_model integration	Type "ollama" routes to the handler via callback
8	Plugin auto-discovery	register_callbacks.py is found by plugin loader

How to Use

1 — Install Ollama

brew install ollama

2 — Pull a model with tool calling support

ollama pull qwen3:8b         # ~6 GB VRAM  — good for testing
ollama pull qwen3:14b        # ~10 GB VRAM — better quality
ollama pull qwen3:30b        # ~20 GB VRAM — recommended

3 — Start the Ollama server

brew services start ollama

curl http://localhost:11434/api/tags

4 — Configure the model in Code Puppy

Create or edit ~/.code_puppy/extra_models.json:

{
  "ollama-qwen3": {
    "type": "ollama",
    "name": "qwen3:8b",
    "context_length": 32768
  }
}

Multiple models can be registered at once:

{
  "ollama-qwen3-8b": {
    "type": "ollama",
    "name": "qwen3:8b",
    "context_length": 32768
  },
  "ollama-qwen3-30b": {
    "type": "ollama",
    "name": "qwen3:30b",
    "context_length": 131072
  }
}

Custom host (different port, remote machine, LM Studio):

{
  "lmstudio-codellama": {
    "type": "ollama",
    "name": "codellama:34b",
    "context_length": 16384,
    "custom_endpoint": {
      "url": "http://192.168.1.50:1234/v1",
      "api_key": "lm-studio"
    }
  }
}

Alternatively, set OLLAMA_HOST to override the default endpoint without editing the config:

export OLLAMA_HOST=http://myserver:11434

5 — Run Code Puppy and switch to the local model

./code-puppy-dev --interactive

Inside the session:

/model ollama-qwen3

Or start directly on the model:

./code-puppy-dev --interactive --model ollama-qwen3

Testing

Unit tests (no Ollama required — fully mocked)

pytest tests/test_ollama_plugin.py -v

Coverage: 100% on plugin files across 15 test cases covering:

• custom_endpoint path (uses get_custom_config())
• Default localhost path
• OLLAMA_HOST env var — append /v1, already ends with /v1, trailing slash, empty string
• Returns OpenAIChatModel not OpenAIResponsesModel
• model_config["name"] resolution with fallback to config key
• Graceful None return on exception
• Handler registration structure

Verified End-to-End

• ollama serve + qwen3:8b on localhost
• Model appears in /model picker after configuring extra_models.json
• File read/write tool calls work
• Multi-step agentic tasks complete successfully
• No remote API calls made during local model session