# Eval Agent Toolbox

Minimal tool gateway for `auxiliary_metrics.py`.

Public API:
- `eval_agent.tool_box.call_vision(text, image_paths, results_dir=...) -> str`
- `eval_agent.tool_box.call_tool(name, payload) -> str`

The agent only needs the public API. Provider details are hidden in:
- `eval_agent/tool_box/_internal/`

## Vision usage controls

Environment variables:
- `EVAL_TOOLBOX_VISION_MAX_CALLS` (default `2`)
- `EVAL_TOOLBOX_VISION_MAX_IMAGES` (default `2`)
- `EVAL_TOOLBOX_VISION_MODEL` (default `gemini-2.5-flash`)
- `EVAL_TOOLBOX_GCP_PROJECT` (default `research-01-268019`)
- `EVAL_TOOLBOX_GCP_LOCATION` (default `global`)
- `EVAL_TOOLBOX_VISION_SYSTEM` (optional system instruction)

Usage is logged to:
- `<experiment_root>/eval_agent_memory/tool_usage.json`

## Template

Reference example:
- `eval_agent/templates/auxiliary_metrics_template.py`