Summary

Adds improved feature-gated local model support for OpenAI-compatible endpoints:

• ollama/<model>
• vllm/<model>
• llamacpp/<model>
• local://<model>

This wires local routing through LiteLLM, CLI /model, backend validation/catalogs, and the web model selector. Local model UI/API support is gated by ENABLE_LOCAL_MODELS=true; local inference servers must already be running.

Review Follow-up

Addressed review feedback:

• Warn when reasoning_effort is passed to local models in non-strict mode.
• Keep strict-mode local reasoning_effort rejection explicit.
• Removed redundant whitespace validation.
• Removed fragile sys.path injection in local model validation tests.
• Added explicit feature-flag coverage for custom local IDs.
• Documented that local endpoint env vars must remain server-controlled.

Validation

• uv --cache-dir /tmp/uv-cache run pytest - 178 passed, 3 skipped
• npm run build from frontend/ - passed, with existing Vite chunk-size warning
• python -m py_compile agent/core/llm_params.py backend/model_catalog.py - passed