78 ways we've caught an agent wasting tokens (a few of our favorites)

Security rules get most of the attention, but roughly half of PinkyAI’s catalog isn’t about risk at all — it’s about waste. An AI agent working agentically racks up token spend in specific, repeatable ways, and most of them are cheap to catch before they happen. A few real examples from the catalog:

The duplicate tool-call suppressor. Warns before re-running a Read or Bash call that’s already been made, identically, earlier in the same session. Small on its own; adds up fast across a long session.

The read-batching advisor. After several single-file reads land in the same directory in a short window, this suggests one Glob/Grep instead — usually cheaper than the file-by-file approach it’s replacing.

The stale test/build rerun advisor. If nothing has changed since a test or build command last ran this session, re-running it is very likely to produce the same result. This one flags that before the rerun happens, not after.

The inline-heredoc advisor. A large heredoc embedded directly in a shell command bloats the command’s own token cost. Writing the same content to a file first is usually cheaper — this advisory is what nudges toward that.

The repo-map advisor. After enough single-file reads pile up, a cheap directory-plus-symbol overview is often what the situation actually called for from the start.

Every one of these is advisory, not a security block — you can always proceed as-is. They exist because the pattern is common enough, and the fix is cheap enough, that it’s worth a nudge before the token spend happens rather than a retrospective report after.

That’s 5 of 78. The full catalog — what’s fire-tracked, what config key controls it, and where it surfaces — ships with the engine and is what pinkyai status counts against.