Kyma API: send a test completion

Run one real, billed chat completion through Kyma API with the user's explicit approval and report back the output and what it cost. Use when the user asks to test a model, check that a model actually answers, compare two models on the same prompt, or says something like "reply with pong via <model>".

Send a test completion through Kyma API

send_message is the only billed tool in this plugin. It runs a real completion against a real model and charges the account.

When to use

• "Test ." / "Does actually work?"
• "Ask to reply with pong."
• "Run this prompt through and and compare."

The approval rule

• send_message requires allow=true. Never pass it without the user's go-ahead in this conversation.
• Ask once, in plain terms: which model, the exact prompt, and roughly what it will cost. Then wait.
• Without an Allow there is no charge and no call. Silence is not approval, and neither is an earlier "yes" to a different prompt or a different model.
• One approval covers one call. A second model, a retry, or a longer prompt needs a fresh ask.

Steps

1. Confirm the model id with get_model if the user gave a friendly name or an alias, so you bill the model they meant.
2. Check the budget first — get_credits and get_spend — and say the remaining amount out loud. See the check-spend-and-credits skill.
3. Show the user the exact model id and prompt you are about to send, with a cost estimate from list_pricing for that model.
4. On their explicit go-ahead, call send_message with allow=true. Keep max_tokens small for a smoke test; a "reply with pong" check does not need a long ceiling.
5. Report the model's reply verbatim, then the actual cost from the response, then the remaining budget.

Comparing models

Run the same prompt against each model as separate approved calls, with the same max_tokens. Put the replies side by side with each one's cost and latency. Do not judge a model on a single short completion — say what the sample does and does not show.

Do not

• Do not use send_message for work the user did not ask to be billed for, such as summarizing, checking a fact, or reformatting something.
• Do not retry a failed call automatically. Report the failure and ask.
• Do not attribute a model's reply to any inference backend. Report the model id that answered and nothing about where it ran.