Default tests are mock-only. Live provider calls are demonstrated by the
repository-level examples/*.exs scripts. These scripts run directly and make
real provider calls when the underlying provider dependency and credentials are
available.
Run examples from the repository root:
elixir examples/live_gemini_ex.exs
GeminiEx
export GEMINI_API_KEY=...
export INFERENCE_GEMINI_MODEL=gemini-3.1-flash-lite-preview
elixir examples/live_gemini_ex.exs
Agent Session Manager
The completion-only examples use Claude or Codex. Amp, Antigravity, and Cursor are recognized ASM providers, but Inference intentionally returns a typed unsupported-capability error for them until their SDK feature manifests prove completion-only execution.
elixir examples/asm_adapter/text_only.exs \
--provider codex \
--model gpt-6.1-sol \
--prompt "Reply with exactly: INFERENCE_ASM_OK"
elixir examples/asm_adapter/tools_unsupported.exs \
--provider codex \
--model gpt-6.1-sol
The 0.5.2 release was verified on 2026-10-27 with a live query through
Inference.Adapters.ASM, ASM 0.17.4, and Codex SDK 0.21.4 on the :sdk lane.
GPT-6.1 Sol at low reasoning effort returned exactly INFERENCE_052_CODEX_OK.
This verifies the completion path; other provider paths were not live-tested
for this dependency refresh.
ReqLlmNext
export GEMINI_API_KEY=...
export INFERENCE_REQLLM_NEXT_PROVIDER=google
export INFERENCE_REQLLM_NEXT_MODEL=gemini-3.1-flash-lite-preview
elixir examples/live_reqllm_next.exs
ReqLLM Compatibility
Installs the latest compatible Hex package, currently req_llm ~> 1.29.0.
export GEMINI_API_KEY=... # or GOOGLE_API_KEY=...
export INFERENCE_REQ_LLM_PROVIDER=google
export INFERENCE_REQ_LLM_MODEL=gemini-3.1-flash-lite-preview
elixir examples/live_req_llm.exs