endpoint: https://llm-gateway.example.com/v1
api_key: ""            # optional; env LLMTB_API_KEY wins
models:
  - uni/deepseek-v4-flash-0731
  - uni/qwen3.6-27b
  - uni/btl-4
concurrency: 8         # default when omitted; matches internal/config's default
request_timeout: 900s  # default when omitted is 300s
max_tokens_default: 40000  # default when omitted; a reasoning model can spend
                            # thousands of tokens on reasoning_content before
                            # writing an answer, so this must stay generous
max_retries: 2         # retries per request after the first try (default 2; 0 disables).
                       # Raise for flaky gateways: backoff is exponential and
                       # honors Retry-After.
# seed: 1234           # optional; sent with every request for backends that honor it
# Slow local models: give them their own timeout and queue depth.
# retry_timeouts: false     # default true; false stops a timed-out call re-running a full slow generation
# model_timeouts:           # overrides request_timeout per model
#   local/big-model: 45m
# model_concurrency:        # caps in-flight calls per model (global concurrency still applies)
#   local/big-model: 1
# stream: false             # true = SSE; detects a hung server after idle_timeout instead of the full request timeout
# idle_timeout: 5m          # max silence in a stream (must exceed prompt-processing time); default 5m