PAIR Bridge 0.7.2

Changelog

0.7.2

• Add configured Unsloth Studio devices for model inventory, memory estimates, load/unload, direct chat, background jobs, and the Hermes gateway.
• Require Unsloth models to be loaded through PAIR Bridge before direct chat so memory preflight and the configured unload allowlist apply.
• Keep Unsloth memory estimates independent of the LM Studio lms CLI.
• Expand English and Russian setup, safety, and tool documentation for Unsloth Studio.

v0.7.0

Full Changelog: v0.6.2...v0.7.0

PAIR Bridge 0.6.2

PAIR Bridge 0.6.2 fixes an empty-inventory failure seen on Alfred.

pair_devices and LM Studio's native API could see installed models while the PAIR/OpenAI-compatible /v1/models router catalog was empty because no model was loaded. pair_list now falls back to the installed inventories of configured online devices and keeps device, installed/loaded state, context, and offline-device provenance explicit. A non-empty router catalog remains authoritative.

No model is downloaded or loaded by listing. The returned notice directs agents to pair_smart_ask, or to an explicit device load and ask.

Validation: 55 self-tests passed; live MCP reproduction on Alfred confirmed six installed models behind an empty router catalog.

PAIR Bridge 0.6.1

PAIR Bridge 0.6.1 fixes capacity reporting when the Bridge runs locally on Windows, including on Alfred.

The Windows adapter now reads installed and available physical memory from the operating system. This allows pair_devices and cold-load preflight to use a fresh RAM sample instead of reporting unknown capacity. Remote Windows sampling over SSH remains supported.

Validation: 53 self-tests passed on macOS; plugin manifest validation passed. The installed 0.6.0 build on Alfred had reported partial capacity despite seeing its six installed models; this patch addresses that observed issue.

PAIR Bridge 0.6.0

PAIR Bridge 0.6.0 improves model selection and long-running requests across configured LM Studio devices.

• Samples available system memory, NVIDIA VRAM, and disk space on Mac and Windows devices. Cold-load preflight checks the intended context against fresh capacity data; unknown values remain explicit.
• Adds opt-in, local benchmark suites for already loaded models. pair_smart_ask can use recent pass rate and latency for its task profile. Benchmark journals contain metrics, never prompts or answers.
• Adds pair_job_start, pair_job_status, pair_job_cancel, and pair_job_recover for bounded background requests, progress, partial responses, cancellation, and prompt-free recovery metadata.
• Updates English and Russian documentation, the /pair skill, and ProofLoop evidence.

Validation: 52 self-tests passed on macOS; plugin manifest and skill validation passed. A live cancellation test completed on the Mac device. A representative live benchmark and full cold-load cycle remain to be verified. Cancelling a bridge request stops its HTTP stream but does not guarantee that LM Studio stops generation internally.

PAIR Bridge v0.5.0 — Smarter local models for Codex

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.

v0.4.0 — /pair and device model management

Adds the pair skill for the Codex slash menu and device-scoped LM Studio model management.

• Inspect installed models and loaded instances with pair_devices and pair_list(device).
• Load installed models with pair_load and unload exact instances with pair_unload.
• Target a specific device with pair_ask(device), or keep PAIR routing by omitting device.
• Remote devices use temporary loopback SSH tunnels with existing host authentication.
• Updated English/Russian setup and privacy documentation.

Management requires explicitly configured LM Studio native v1 endpoints. Automatic cluster management discovery, Ollama management, downloading/deleting weights, and engine installation are not included.

Validation: 17 tests passed on macOS, Windows and Linux. Live Mac and Windows device queries passed; Windows model load/unload was verified through SSH.

Open a new Codex task after updating to discover the pair skill and five MCP tools.

v0.3.1 — Consistent command documentation

All English and Russian usage examples and plugin starter prompts now use pair_list and pair_ask. Added explicit parameter examples and clarified that these are MCP tools used in Codex, not shell or slash commands. The diagram already uses the short names. No inference behavior changed.

v0.3.0 — Short tools and illustrated guides

Simpler tools, clearer setup

• Renamed MCP tools to pair_list and pair_ask.
• Added an illustrated explanation of Codex → MCP bridge → PAIR → local models.
• Reworked the English guide and added a complete Russian guide.
• Added setup steps, example prompts, configuration tables and troubleshooting.

Upgrade

codex plugin marketplace upgrade codex-pair-bridge
codex plugin add codex-pair-bridge@codex-pair-bridge

Open a new Codex task after upgrading. This release removes the old tool names; update saved prompts that refer to them. Existing router configuration remains compatible.

Codex PAIR Bridge v0.2.0

Initial public release of Codex PAIR Bridge, an independent community MCP plugin for consulting local models through NVIDIA PAIR.

• Installable GitHub marketplace with portable uv-based startup and locked dependencies.
• Model catalog and explicit model-query tools.
• Configurable endpoint, optional bearer token, and cross-process inference locking.
• MIT license and privacy documentation.
• Nine tests passed on macOS, Windows, and Linux. GitHub installation and a real PAIR inference also verified on macOS.

Install:

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

Requires uv on PATH and your own running PAIR router. Open a new Codex task after installation.