PAIR Bridge 0.8.0: richer MCP inference and output budgets

[GermanMik]

PAIR Bridge 0.8.0 is out: a community-built MCP plugin for NVIDIA PAIR routing and configured local-model devices.

Highlights:

• Fresh per-model capability discovery reports supported, unsupported, or unknown.
• Structured JSON responses are validated locally against the caller's schema.
• Process-local job and queue visibility, plus atomic multi-job batches with independent outcomes. Jobs across physical PCs can overlap while each device remains serialized.
• Bounded embeddings and inline-image vision requests for compatible model instances that are already loaded.
• Removed the fixed 8,192-token max_tokens ceiling. Requests honor fresh model/instance limits and planned cold-load context without silent clamping; when the output limit is unknown, the engine decides and accounts for prompt/reasoning tokens.

The public release checkout passes 116 tests. GitHub Actions passes on Windows, macOS, and Ubuntu.

Release: https://github.com/GermanMik/pair-bridge/releases/tag/v0.8.0
Project: https://github.com/GermanMik/pair-bridge