Support multiple AI providers (OpenAI-compatible API)

[aliciapaz]

Context

The gem was hardcoded to OpenAI's proprietary Responses API (client.responses.create()), making it impossible to use with other AI providers like Minimax, Ollama, Groq, or any OpenAI-compatible service. Users who don't use OpenAI were locked out entirely.

Closes #3

What Changed

Switched the AI layer from OpenAI's Responses API to the standard Chat Completions API, which is the universal interface that all OpenAI-compatible providers implement.

Key files modified:

• templates/config.js — Provider-agnostic config: reads AI_API_KEY (falls back to OPENAI_API_KEY), supports AI_BASE_URL for custom endpoints, default model changed to gpt-4o
• templates/ai.js — Replaced client.responses.create() → client.chat.completions.create(), updated response parsing (resp.choices[0].message.content instead of resp.output_text), added null-content guard for truncated/refused responses
• install_generator.rb — Updated post-install instructions to reference AI_API_KEY and AI_BASE_URL
• README.md — Updated requirements, setup steps, config table, and "How it works" section
• agent_e2e.gemspec — Updated summary and description

Notable decisions:

• Kept the openai npm package as the SDK — it natively supports custom baseURL, so no new dependencies needed
• OPENAI_API_KEY is still supported as a fallback so existing users don't break
• Added explicit error when AI_API_KEY is missing (previously the OpenAI SDK would throw an opaque error)

How to Test

1. OpenAI (default): Set AI_API_KEY=sk-... in .env, run bin/e2e — should work identically to before
2. Backward compat: Set OPENAI_API_KEY=sk-... (old var name) — should still work via fallback
3. Custom provider: Set AI_API_KEY, AI_BASE_URL, and AI_MODEL for a compatible provider (e.g., Ollama at http://localhost:11434/v1)
4. Missing key: Remove both env vars — should get a clear error message and exit code 1
5. Null response: If the model returns content: null (e.g., token limit hit), should get "AI returned empty response" error instead of a crash

Deployment Tasks

None

🤖 Generated with Claude Code