ReqLLM v1.29.0 released!

ReqLLM v1.29.0 has been released.

Security update

ReqLLM 1.29.0 fixes GHSA-6vww-5m28-f26v. Versions 1.9.0 through 1.28.0 did not verify TLS certificates for affected WebSocket sessions. Users of OpenAI Responses WebSocket, Realtime, or Codex WebSocket sessions should upgrade. HTTP streaming is not affected.

Features:

• streaming: expose builtin tool start chunks to stream consumers (#1133) by Vasilis Spilka

• retry: retry transient HTTP statuses with exponential backoff (#1124) by redac

• support gateway auth and model provider defaults (#1116) by mikehostetler

• bedrock: grammar-constrained structured output for Claude (#1113) by Max

• fireworks: support context_length_exceeded_behavior request parameter (#1110) by green-david

• openai: keep tool search items and namespaces on the Responses API (#1109) by Mark Ericksen

• reflection: add Reflection AI provider (#1108) by mikehostetler

• openrouter: support prompt_cache_key option (#1094) by deepfates

• azure: support Microsoft-Decision-1 evaluation (#1092) by neilberkman

Bug Fixes:

• streaming: verify TLS for HTTPS WebSocket URLs (#1140) by mikehostetler

• google: omit empty text from tool-only continuations (#1128) by Gabriel Rinaldi

• google: use header authentication by default (#1139) by neilberkman

• bedrock: map response roles explicitly (#1138) by neilberkman

• config: skip dotenv loading by default in releases (#1137) by neilberkman

• http: disable redirects by default (#1136) by neilberkman

• application: restrict Tidewave REPL to development (#1135) by neilberkman

• github_copilot: make gh token lookup opt-in (#1134) by neilberkman

• anthropic: make thinking off reachable and report dropped sampling params (#1131) by neilberkman

• anthropic: derive pricing context from confirmed request and usage facts (#1122) by mikehostetler

• usage: document cached input counts and retain fallback flag (#1118) by mikehostetler

• streaming: preserve JSON text and thinking content (#1120) by mikehostetler

• bedrock: validate prepared object stream options (#1114) by mikehostetler

• anthropic: reduce grammar-constrained schemas in their declared order, atom keys included (#1112) by Max

• deps: update Bandit and remove temporary CVE exception by mikehostetler

• ci: temporarily exempt the Bandit timeout advisory by mikehostetler

• anthropic: keep stop_reason in provider_meta so a paused turn is identifiable (#1107) by Mark Ericksen

• google: validate assistant prefill before sending requests (#1106) by mikehostetler

• billing: explain missing cached-token pricing (#1100) by mikehostetler

• anthropic: preserve adaptive reasoning on cloud providers (#1099) by mikehostetler

• google: enforce newer Gemini parameter constraints (#1098) by mikehostetler

Refactoring:

• schema: use MapSet transform constructor (#1123) by mikehostetler

ReqLLM Info

Classification: LLM Client Library
Description: An Elixir library built on Req and Finch with a unified interface to LLM APIs.
Forum/tag: req-llm

Click here to view ReqLLM’s main announcement thread.