Send real traffic through an installed gateway and grade the response.
Pods, Gateway, and EPP resources can all be ready while inference routing,
streaming, or MCP session handling is broken. This exercises the customer path
instead of control-plane readiness.
Checks, in order of what they need:
- auth enforced — always runs; an unauthenticated request must be rejected
with the data plane's own 401. Needs no credential.
- model routing — with --smoke-api-key: an unknown model must be rejected
with a 400, proving the key authenticated.
- completion — with --smoke-model: a real request returns a completion.
- streaming — with --smoke-model: a streamed response stays valid to
completion.
- MCP session — with --smoke-mcp-route: initialize then list tools.
A check whose inputs are absent is skipped, not failed, so environment and
credential limits stay distinct from product failures. Exit is non-zero only
when a check finds the gateway itself at fault.
This is the same set tare install and tare upgrade run after they apply; run it
standalone to re-check after enabling a model, adding a route, or an incident.
Examples:
# Credential-free: prove the gateway rejects anonymous traffic
tare smoke https://proxy.example.com
# Full path against an enabled model (spends a few tokens)
tare smoke https://proxy.example.com \
--smoke-api-key sk-... --smoke-model gpt-4o-mini
# Check an MCP route as well
tare smoke https://proxy.example.com \
--smoke-api-key sk-... --smoke-mcp-route /mcp/<profile-id>
Usage:
tare smoke <serve-url> [flags]
Flags:
--smoke-api-key string Inference API key for the post-apply gateway smoke tests. Prefer TARS_API_KEY — a key passed here lands in shell history and the process list. Mint one from the management-plane dashboard. Without it, only the credential-free auth check runs.
--smoke-mcp-route string MCP route path to establish a post-apply session against (e.g. /mcp/<profile-id>). Omit to skip.
--smoke-model string Model to send a real post-apply request to (e.g. gpt-4o-mini). Requires --smoke-api-key. Spends a few tokens on the named model to prove inference and streaming work end to end. Omit to skip.
--smoke-timeout duration Per-request timeout for the post-apply gateway smoke tests. (default 30s)
Global Flags:
-q, --quiet Suppress progress output; only errors and the final summary go to stderr
-v, --verbose Verbose output: plain-mode progress lines plus full per-blob chunk logs to stderr