Skip to content

v0.4.1

Latest
Latest

Choose a tag to compare

View all tags
@codelion codelion released this 02 Nov 05:00
v0.4.1
f74e678
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's Changed

Action required if you use these features

  • executecode is now disabled by default (#337, #302). The plugin runs Python from requests and LLM responses with no sandboxing. To keep using it, start optillm with --enable-unsafe-plugins=executecode or set OPTILLM_ENABLE_UNSAFE_PLUGINS=executecode. Only enable it in single-user, air-gapped environments.
  • MCP secrets could be written to logs. The MCP plugin wrote configured env and headers values, which can include API keys, to ~/.optillm/logs/mcp_plugin.log. It now logs only the names. If you use the plugin, check that file for old secrets and rotate any you find.

New

  • Streaming hints plugin (#352, #206): streaming_hints watches the model's reasoning as it is generated and injects relevant hints from a memory bank at moments of uncertainty.
  • Legacy completions endpoint: the inbuilt inference server now exposes POST /v1/completions.
  • LLM gateway guide: docs/llm-gateway.md covers using the proxy plugin for routing and inference optimization together.

Security

  • readurls: blocks SSRF destinations, including CGNAT/shared and multicast ranges.
  • z3: statically gates solver code to the declared z3/sympy subset and closes string-based evasions of the AST attribute denylist.

Fixes

  • max_completion_tokens (#350): pass-through and proxy requests no longer add max_tokens, which o-series and gpt-5 models reject with a 400. The local inference server now honors max_completion_tokens.
  • mcts: follows tree selection through to a leaf.
  • mars: populates diverse aggregation groups, preserves full agent IDs when applying verification, and matches solution authors across string and integer IDs.
  • bon: excludes truncated and blank candidates.
  • rstar: backpropagates rewards from evaluated rollout nodes and preserves signed and decimal answers.
  • deepconf: votes and calibrates as in the DeepConf reference implementation.
  • entropy_decoding: keeps the nucleus head and applies min_p in sampling.

Project

  • CI now runs every test file, and MCP tests run on pull requests.
  • The README has been streamlined.

Thanks to Parman Mohammadalizadeh, akushonkamen, ZY Liu, flyingdoubleG, and Daniel Peng for their contributions.

Full Changelog: v0.4.0...v0.4.1

Assets 2
Loading