| name | model-proxy |
| description | Use when configuring, running, or verifying the UTK model proxy — the OpenAI-compatible local context gateway for GitHub Copilot — including history compaction into recoverable [utk-block] summaries, deferred tool discovery (utk_find_tool), repeated/stale tool-output dedupe, route-specific compactors, observe-only cache-volatility detection, expand_context and proof recovery endpoints, prompt-compression policy, upstream provider routing (github-models, azure-ai-inference, azure-openai, openai), and .utk/model-proxy artifacts |
UTK Model Proxy
Use this skill to configure, run, and verify @utk/model-proxy — UTK's OpenAI-compatible local proxy that reduces repeated context before requests reach the upstream provider.
The proxy is context-gateway, not a public CLI or VS Code extension. It forwards Chat Completions, Responses, and Models requests while keeping originals local under .utk/model-proxy and returning compact, recoverable model-visible text. Default behavior is local-first: remote compressors and model downloads stay disabled unless explicitly enabled.
References
references/pipeline.md: request pipeline (normalize → policy → budget → prompt optimize → content route → retention → tool discovery → persist → forward → recover) and local-first compaction behavior.references/endpoints-and-recovery.md: HTTP endpoints plus expand_context, proof, and find_tool recovery contracts.references/configuration.md: [model_proxy] policy defaults, provider routing, and metrics.
Commands
npm run build -w @utk/model-proxy
npm test -w @utk/model-proxy
Keep detok MCP and preToolUse hook guidance in the detoks skill and the utk-detoks plugin; keep tool-hook mediation in the utk skill.
