verbosity

Detects: a call generating a large amount of output with no max_tokens cap set at all.

Output tokens are billed at 4–6× the rate of input tokens on most models, so unconstrained generation is disproportionately expensive relative to its token count.

How it works

A call is flagged when it has no max_tokens set and its realized output exceeds a configured ceiling (1,000 tokens by default). Both conditions matter: a capped call is never flagged, regardless of how much it generates — the developer already made a deliberate choice about the ceiling, and Wattage doesn't second-guess it without knowing whether the output was substantively useful.

Known limitation

The fuller version of this detector (per the original design) uses per-task-type expected bands — a short "classify" step should generate very little output; a "draft a document" step legitimately generates a lot. That needs a task-type classifier, which doesn't exist yet in this codebase. Using a single global ceiling risks flagging legitimate long-form generation as if it were verbosity. Wattage documents this rather than fabricating a task-type signal it doesn't actually have; this is a natural direction for a future contribution (see CONTRIBUTING.md).

wasted_tokens here means "tokens beyond your configured ceiling" — a policy estimate you set the threshold for, not a factual claim that those specific tokens were unnecessary. Judging that for certain would require evaluating the actual content, which needs the optional LLM judge (off by default).

Fix

Prompt for concise or structured/extractive output (JSON, a fixed schema), or use the provider's effort/verbosity control for the step. A max_tokens cap also works, but set it generously — a cap tight enough to truncate mid-answer forces retries that cost more than the excess ever did.

Quality risk: low

Capping output length is usually safe, but an overly aggressive cap can truncate a genuinely long, useful answer — hence "low" rather than "none".