Skip to content

PAIR Bridge v0.5.0 — Smarter local models for Codex

Choose a tag to compare

View all tags
@GermanMik GermanMik released this 20 Sep 11:20
v0.5.0
c876f81

PAIR Bridge connects Codex to installed local models through NVIDIA PAIR or directly to configured LM Studio devices. This release expands /pair from basic model access into a managed workflow for discovery, task-based selection, comparison, diagnostics, and resource planning.

Highlights

  • Project and plugin rename: codex-pair-bridge is now pair-bridge. Existing ~/.codex-pair-bridge.json remains readable; ~/.pair-bridge.json takes precedence.
  • Smarter model selection: live inventory checks and an in-lock recheck prevent requests to missing models. Task profiles can select an already installed chat model; explicit model and device choices remain available.
  • Second opinions: compare two selected local models and retain their provenance. Codex must verify disputed claims against the source material.
  • Resource planning: per-device queues, sanitized request diagnostics, and pair_memory_plan use LM Studio's read-only CLI estimate at the planned context, including the actual context of loaded instances. A configured memory budget blocks estimated use above its limit with 10% headroom.
  • Explicit downloads: a model request never downloads weights. A separate one-use download plan can check an unambiguous Hugging Face GGUF file's size and revision and local disk space before a confirmed download.
  • Integrations and guides: Oh My Pi setup, optional Jev Choice/Score adapters, updated English and Russian documentation, and a refreshed MCP architecture illustration.

Upgrade

For the renamed plugin, follow the migration instructions. New installations can run:

codex plugin marketplace add GermanMik/pair-bridge
codex plugin add pair-bridge@pair-bridge

Validation and limits

The release commit passed 52 self-tests in a clean checkout, including MCP tool discovery. A read-only memory estimate was checked on Alfred. No release validation downloaded weights or performed a cold load. Actual free RAM/VRAM and remote LM Studio storage paths remain separate from the CLI estimate; live Jev, OMP, and two-model workflows still need device-specific verification. See TODO for the remaining work.

Assets 2
Loading