# llm-bench configuration.
# Copy to .env and edit. Env vars already in your shell take precedence
# over values here (so you can override per-run without editing the file).

# The OpenAI-compatible chat-completions URL of the model you're testing.
# Examples:
#   LM Studio (local):     http://127.0.0.1:1234/v1/chat/completions
#   Ollama (local):        http://127.0.0.1:11434/v1/chat/completions
#   vLLM / llama-server:   http://127.0.0.1:8000/v1/chat/completions
#   Ollama Cloud:          https://api.ollama-cloud.com/v1/chat/completions
#   OpenAI:                https://api.openai.com/v1/chat/completions
ENDPOINT=http://127.0.0.1:1234/v1/chat/completions

# The model id the endpoint expects in the "model" field of requests.
# For LM Studio this is the load identifier; for Ollama it's the tag;
# for vLLM/llama-server it's the served model name; for cloud it's the
# API model string (e.g. "gpt-4o", "claude-sonnet-5" via a compat layer).
MODEL=local

# Bearer token, sent as "Authorization: Bearer <key>" if non-empty.
# Leave blank for local endpoints (LM Studio, local Ollama, local vLLM).
# Set for cloud endpoints or any endpoint that requires auth.
API_KEY=

# Per-turn generation cap (max_tokens in the request body).
# Coding suite default is 8000; agentic benches default to 16000.
MAX_TOKENS=8000

# Sampling temperature. 0 = greedy (deterministic, recommended for
# coding/tool-use benchmarks). Higher = more creative/noise.
TEMPERATURE=0

# Per-request wall timeout in seconds. Bump for slow local models.
HTTP_TIMEOUT=900

# Agentic loop cap — max turns the agent can take before giving up.
# Only used by agentic-fix, agentic-build, agentic-multitask.
MAX_TURNS=40

# Suppress reasoning/thinking via an empty <think></think> assistant
# prefill. Works on Qwen3-family and other models that use the think-tag
# convention. Set to "1" to enable. Some backends also honor a
# chat_template_kwargs.enable_thinking field (the no-think instrument
# tests both paths).
NOTHINK=

# --- build/test timeouts (agentic benches) ---
BUILD_TIMEOUT=180
TEST_TIMEOUT=180
SHELL_TIMEOUT=60
SERVER_TIMEOUT=20