v0.3.3-alpha — Generic provider params, admin UI, install fixes

Fixed

• Startup crash when .env has inline comments — systemd's EnvironmentFile doesn't strip inline # comments.
• Misleading "install as service" hint when run via curl | bash.
• install.sh didn't stop the service before updating.

Added

• Generic per-provider params config — override, suppress (null), or add any API parameter per provider:

  {"name": "kimi", "api_type": "openai", "url": "...", "params": {"temperature": 1, "top_p": 0.95}}

• Auto-retry on parameter rejection — on 400 errors like "invalid temperature", the rejected parameter is stripped and the request retried automatically.
• Admin dashboard: provider cards now show editable params (JSON), "Add Provider" form includes a params field.
• New API endpoint: PUT /v1/rlm/providers/params to update provider params at runtime.

Removed

• Hardcoded kimi-k2 model name checks in OpenAIProvider.

v0.3.2-alpha — External review fixes

What's Changed

Fixes for 5 issues found during an independent review of the proxy API:

1. Content-Type tolerance — /api/* (Ollama-native) endpoints now accept requests without Content-Type: application/json, matching real Ollama behavior
2. Missing model → 400 — model is now a required field; returns 400 Field required instead of silently falling back to root model
3. No more URL leaks — Error responses no longer expose internal infrastructure URLs (e.g. http://localhost:11434/...)
4. Actual token counts — Passthrough responses now include real prompt_tokens / completion_tokens from Ollama instead of 0
5. Config endpoint fix — Removed stale root_provider_url reference that caused /v1/rlm/config to 500

Full Changelog: v0.3.1-alpha...v0.3.2-alpha

v0.3.1-alpha — Cross-platform install scripts

What's new in v0.3.1-alpha

• Cross-platform install/uninstall scripts for Linux, macOS, and Windows
• Service support: systemd (Linux), launchd (macOS), NSSM (Windows)
• Docker support on all platforms via --docker flag
• Updated INSTALL.md with platform-specific instructions

Install

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/ghbalf/rlm_proxy/main/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/ghbalf/rlm_proxy/main/install.ps1 | iex

See INSTALL.md for full details.

Full Changelog: v0.3.0-alpha...v0.3.1-alpha

v0.3.0-alpha — First public release

RLM Proxy v0.3.0-alpha

Multi-provider LLM proxy implementing the Recursive Language Model inference paradigm (Zhang, Kraska, Khattab — MIT CSAIL, 2025).

Highlights

• RLM scaffold — sandboxed Python REPL, parallel sub-calls, LRU caching, smart history compaction, adaptive early-stop
• Multi-provider dispatch — Ollama + OpenAI-compatible APIs with provider/model routing, auto-discovery, least-loaded balancing
• Dual API — OpenAI (/v1/*) and Ollama (/api/*) endpoints, both fully dispatched
• Admin dashboard — web UI at /admin for managing providers, settings, and live metrics
• Production hardening — API key auth, request queue, retry with exponential backoff, circuit breaker, token budgets

Quick Install

curl -fsSL https://raw.githubusercontent.com/ghbalf/rlm_proxy/main/install.sh | bash

Or manually:

git clone https://github.com/ghbalf/rlm_proxy.git
cd rlm_proxy
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
python main.py
# Open http://localhost:8881/admin

What's alpha about it

• Tested with local Ollama; remote OpenAI-compatible providers implemented but need broader testing
• Anthropic provider planned but not yet implemented
• No WebSocket streaming support yet
• Admin UI is functional but basic

See README.md and INSTALL.md for full documentation.