Skip to content

RELEASE_NOTES.md

Latest commit

 

History

History
67 lines (51 loc) · 2.29 KB

RELEASE_NOTES.md

File metadata and controls

67 lines (51 loc) · 2.29 KB
Raw

Release Notes

Codex Open NVIDIA NIM Preview

This release introduces the first provider-focused Codex Open build. The main goal is to let users run a local Codex CLI against NVIDIA NIM models while keeping the local build separate from any global Codex installation.

Added

  • Built-in nvidia-nim model provider.
  • NVIDIA NIM hosted endpoint default: https://integrate.api.nvidia.com/v1.
  • Optional CODEX_NVIDIA_NIM_BASE_URL override for private or self-hosted NIM deployments.
  • NVIDIA_API_KEY authentication.
  • Remote model discovery from NIM-compatible /v1/models.
  • Chat-completions inference through /v1/chat/completions.
  • Streaming assistant output for NIM chat models.
  • Thinking display when a model streams delta.reasoning_content.
  • NIM-specific performance guidance that encourages faster workspace search commands such as rg, rg --files, and git ls-files.

Fixed

  • NVIDIA NIM no longer routes to /v1/responses, which returns 404 Not Found for hosted NIM chat models.
  • Provider selection is explicit: use model_provider="nvidia-nim" and the full model slug in model, for example model="z-ai/glm-5.1".

Known Limits

  • Provider-specific model controls such as custom thinking budgets are not yet exposed as first-class Codex config.
  • Model availability depends on the NVIDIA account, region, and endpoint.
  • Tool behavior depends on each NIM model's OpenAI-compatible tool-call support.
  • Image generation, web search, and OpenAI-hosted provider features are disabled for nvidia-nim.

Upgrade Notes

Use a local CODEX_HOME for this build if you do not want to touch global Codex state:

$env:CODEX_HOME="D:\path\to\codex\.codex-local"

Build and run the local binary:

cd codex-rs
$env:CARGO_HOME="$(Resolve-Path ..)\.cargo-local"
$env:CARGO_TARGET_DIR="$(Resolve-Path .)\target-local"
cargo build -j 1 -p codex-cli

$env:NVIDIA_API_KEY="nvapi-your-key"
& ".\target-local\debug\codex.exe" `
  -c 'model_provider="nvidia-nim"' `
  -c 'model="z-ai/glm-5.1"'

Requesting Future Work

Additional providers and model-specific controls will be prioritized from real requests. Open a Provider or Model Request issue with official docs, model ids, auth details, streaming format, tool support, and reasoning/thinking fields.