test-gateway

Test the local gateway end to end

Verified procedure for exercising the gateway happy path from a clean state. Run everything from the repo root with uv run so the project environment is used.

1. Interactive first-run setup (TUI)

exp run on a real TTY with an uninitialized root opens the first-run wizard (exp/cli/gateway/setup.py). Pick a fresh root, then:

uv run exp run --root /path/to/fresh/.exp --port 8822

Known-good prompt answers (OpenAI provider, real key in OPENAI_API_KEY):

Prompt	Answer
Provider connection name	openai
Provider adapter	openai
Credential environment variable name	OPENAI_API_KEY
Base URL	empty (provider default)
Exact provider model ID	gpt-5.4-mini
Exact logical model identity	gpt-5.4-mini
Public model alias	mini
Default identity	accept default
Create this gateway configuration?	y

Identifiers must match ^[a-z][a-z0-9]*(?:[._-][a-z0-9]+)*$ (lowercase, dot/underscore/hyphen separators).

On success it prints "Gateway ready at http://127.0.0.1:8822/v1", the usage URL, and a one-time virtual key (export OPENAI_API_KEY=exp_vk_...), then serves uvicorn. The raw key is shown exactly once: capture it immediately.

Non-TTY contexts (pipes, --json, --non-interactive) never prompt; they exit 2 with gateway_not_initialized plus the exact non-interactive setup commands.

2. Verify the served routes

With K set to the issued virtual key:

curl http://127.0.0.1:8822/v1/models -H "Authorization: Bearer $K"                  # expect alias "mini"
curl http://127.0.0.1:8822/v1/chat/completions -H "Authorization: Bearer $K" \
  -H "Content-Type: application/json" \
  -d '{"model":"mini","messages":[{"role":"user","content":"Reply with exactly: gateway-ok"}]}'
# stream:true returns SSE chunks ending in "data: [DONE]"
curl http://127.0.0.1:8822/v1/responses -H "Authorization: Bearer $K" \
  -H "Content-Type: application/json" -d '{"model":"mini","input":"Say hi"}'
curl -o /dev/null -w '%{http_code}' http://127.0.0.1:8822/v1/models                  # keyless: expect 401