Add optional multi-provider LLM code generation to the OpenCAD agent

[Copilot]

OpenCAD’s agent could only follow a built-in deterministic planning path. This change adds an optional multi-provider LLM path for generating OpenCAD scripts in the same style as the repository examples, while keeping the existing execution flow unchanged.

• Agent API

  • Extends /chat request handling with:
    • llm_provider
    • llm_model
    • generate_code
  • Extends the response with optional generated_code
  • Keeps the default planner behavior intact when code generation is not requested

• Provider abstraction

  • Adds a small LiteLLM-backed integration layer for provider-agnostic code generation
  • Supports provider-qualified model routing (for example openai/gpt-4o-mini)
  • Validates provider/model configuration early instead of silently accepting partial LLM config

• Example-style code generation

  • Adds a dedicated code-generation prompt builder that includes curated scripts from examples/
  • Uses those examples to steer output toward the existing fluent OpenCAD style
  • Falls back to deterministic built-in script generation when no LLM model is configured

• Planner compatibility

  • Preserves the current deterministic operation-planning path for feature execution
  • Adds deterministic script generation patterns for common requests such as mounting brackets and PCB carriers

• Packaging, docs, and examples

  • Adds an optional llm extra for LiteLLM support
  • Documents the new request fields and the code-generation mode in the architecture/runtime docs
  • Adds a runnable usage example under examples/agents/
  • Documents how to run the agent example with deterministic fallback by default or with provider/model environment variables for LiteLLM-backed generation
  • Adds explicit Claude and Gemini example commands in examples/agents/README.md
  • Adds focused tests that execute the new example script and keep the provider-specific example documentation valid

Example request shape:

{
  "message": "Generate a mounting bracket script with corner fasteners and a center cutout",
  "tree_state": { "root_id": "root", "nodes": {} },
  "conversation_history": [],
  "reasoning": true,
  "generate_code": true,
  "llm_provider": "openai",
  "llm_model": "gpt-4o-mini"
}

Example generated code shape:

from opencad import Part, Sketch

bracket_profile = (
    Sketch(name="Generated Bracket Profile")
    .rect(80, 30)
    .circle(3, center=(8, 8), subtract=True)
    .circle(3, center=(72, 8), subtract=True)
    .circle(5, center=(40, 15), subtract=True)
)

Part(name="Generated Mounting Bracket").extrude(
    bracket_profile,
    depth=4,
    name="Bracket Body",
).fillet(edges="top", radius=0.75, name="Bracket Edge Relief")

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Allow for different llm providers</issue_title>
<issue_description>We need a library that can handle llm calls from different providers.
They should also be able to generate code similar to the the code in the examples directory.
</issue_description>

Comments on the Issue (you are @copilot in this section)

• Fixes Allow for different llm providers #12

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

[isayahc]

Provide an example of how it is to be used in example/agentsdirectory

[Copilot]

Provide an example of how it is to be used in example/agentsdirectory

Added a runnable example under examples/agents plus README/docs references, and added a test that executes the example script to keep it valid. Addressed in cde94e3.

[isayahc]

make sure you also include a claude and gemini example for llms

[Copilot]

make sure you also include a claude and gemini example for llms

Added explicit Claude and Gemini usage examples to examples/agents/README.md and a focused test to keep them in place. Addressed in 02f7e5b.

[isayahc]

does this library need litellm to run?

[Copilot]

does this library need litellm to run?

No — litellm is optional. The library still runs without it for the default deterministic planner and deterministic code-generation fallback. It’s only needed when you want provider-backed LLM code generation by setting generate_code with an LLM model/provider.