# Default server / model for tool-eval-bench.
# Copy to .env and adjust for your setup.

# Option A: full URL (preferred)
# TOOL_EVAL_BASE_URL=http://your-server:8080

# Option B: host + port separately (used when BASE_URL is empty)
TOOL_EVAL_HOST=your-server
TOOL_EVAL_PORT=8080

# Backend label for reports (optional: auto-detected, otherwise unknown)
# The request format follows the endpoint, independently of this label.
# TOOL_EVAL_BACKEND=vllm

# Model name/path — leave empty for auto-detection from /v1/models
TOOL_EVAL_MODEL=

# API key — leave empty if not required
TOOL_EVAL_API_KEY=

# ---------------------------------------------------------------------------
# Decision judge for optional answer audits (docs/decision-models.md).
# These only configure the connection: pass --decision-judge to audit a run.
# The model is auto-detected from the judge's /v1/models when left empty.
# The key is sent only to the judge, and the benchmark key never is.
# ---------------------------------------------------------------------------
# TOOL_EVAL_DECISION_JUDGE_BASE_URL=http://localhost:8084/v1
# TOOL_EVAL_DECISION_JUDGE_MODEL=
# TOOL_EVAL_DECISION_JUDGE_API_KEY=

# ---------------------------------------------------------------------------
# Named providers, for A/B runs from one .env.
#
# `tool-eval-bench run --provider gemini` (or TOOL_EVAL_PROVIDER=gemini) reads
# TOOL_EVAL_GEMINI_BASE_URL / _API_KEY / _MODEL instead of the generic values
# above. The name is free-form: VLLM_A and VLLM_B work as well as the vendor
# names. gemini, openai and anthropic also set the report's backend label and
# skip engine probing. --base-url, --api-key and --model still override.
# ---------------------------------------------------------------------------

# Gemini, native API (https://ai.google.dev/api/generate-content).
# Do not include /openai in the path: that is the OpenAI-compatible layer.
# TOOL_EVAL_GEMINI_BASE_URL=https://generativelanguage.googleapis.com
# TOOL_EVAL_GEMINI_API_KEY=
# TOOL_EVAL_GEMINI_MODEL=gemini-2.5-pro

# OpenAI (https://developers.openai.com/api/reference).
# TOOL_EVAL_OPENAI_BASE_URL=https://api.openai.com
# TOOL_EVAL_OPENAI_API_KEY=
# TOOL_EVAL_OPENAI_MODEL=gpt-5

# Anthropic, native Messages API (https://platform.claude.com/docs/en/api/messages).
# Detected from the host. Add /v1/chat/completions to the URL to benchmark the
# OpenAI-compatible layer instead.
# TOOL_EVAL_ANTHROPIC_BASE_URL=https://api.anthropic.com
# TOOL_EVAL_ANTHROPIC_API_KEY=
# TOOL_EVAL_ANTHROPIC_MODEL=claude-opus-5

# A gateway that serves the Messages API for other models (OpenCode Zen here).
# A URL ending in /messages selects the format; a gateway root that serves
# several formats needs --format anthropic.
# OpenCode Go routes by conversation: it refuses a request without a stable
# x-opencode-session and asks clients to name themselves. Headers are
# NAME=VALUE pairs separated by ';'. --header and --session-header override.
# TOOL_EVAL_ZEN_BASE_URL=https://opencode.ai/zen/go/v1/messages
# TOOL_EVAL_ZEN_API_KEY=
# TOOL_EVAL_ZEN_MODEL=
# TOOL_EVAL_ZEN_HEADERS=User-Agent=my-coding-agent/1.0
# TOOL_EVAL_ZEN_SESSION_HEADER=x-opencode-session

# A second self-hosted box, for engine-vs-engine comparisons.
# TOOL_EVAL_LOCAL_BASE_URL=http://gpu-box:8080