Retry-aware LLM wrapper with backoff, telemetry hooks, and fallback.

Wraps any LangEx.LLM provider with automatic retries on transient failures (rate limits, transport errors, server errors), linear backoff, and configurable callbacks for observability.

Usage

LangEx.LLM.Resilient.chat(LangEx.LLM.Anthropic, messages,
  model: "claude-opus-4-6",
  tools: tools,
  max_retries: 3,
  retry_base_ms: 3_000
)

Options

All options not listed below are forwarded to the underlying provider.

• :max_retries — maximum retry attempts (default 3)
• :retry_base_ms — base delay between retries in ms (default 3_000); actual delay is base * (attempt + 1) (linear backoff)
• :retryable? — fn(error_reason) -> boolean() to classify retryable errors (default: 429, transport errors, 5xx)
• :on_success — fn(attempt, duration_ms, ai, usage) -> any()
• :on_retry — fn(attempt, duration_ms, wait_ms, reason) -> any()
• :on_error — fn(attempt, duration_ms, reason) -> any()
• :fallback — fn() -> Message.AI.t() called on final failure; when nil, the error propagates as {:error, reason}