Changelog

// Tracking the evolution of recursive reasoning.

April 25, 2026

OpenAI-Compatible Server and Lock-Free Registry

Today we are shipping a major architectural upgrade that makes Altum a drop-in replacement for standard LLM providers. By exposing an OpenAI-compatible API, you can now point your favorite agentic tools (like OpenCode or Codex) directly at an Altum instance.

  • OpenAI-Compatible Server: New serve-openai command starts an HTTP server supporting /v1/chat/completions and /v1/responses.
  • Lock-Free Tool Registry: Migrated from standard Mutex locks to an ArcSwap pattern for tool lookups. This allows high-concurrency environments to read the toolset without any contention.
  • Extensible Tool Injection: Use the configure_tools MCP tool or API to register custom functions that the RLM sandbox can invoke during reasoning.
  • RlmService Refactoring: Unified the core reasoning engine into a standalone service, ensuring consistent behavior across CLI, MCP, and HTTP interfaces.
  • Model Overrides: Specify the model and sub-model per-query rather than being locked to global defaults.
Performance Benchmark Results

Testing registry read latency under load (100 concurrent users, 1,000 reads each):

113.8ms
Standard Mutex Locking
18.3ms
Lock-Free ArcSwap

Finding: The move to lock-free reads resulted in a 6x speedup for high-concurrency tool lookups.

March 1, 2026

Initial Release

Altum RLM engine is officially open source. Featuring recursive depth management, sandboxed Python REPLs, and a persistent context store.