dsh-ponytail

AI coding agents have a habit: they hand-roll components the platform already ships, pull in a new dependency for a two-line helper, and split a three-field validation into six files and a class. The code has barely started running before the token bill spikes and review effort doubles. dsh-ponytail targets exactly this problem. Before every inference it injects a set of minimal-code rules that make the model ask itself first: can I use a native API? Can the stdlib handle it? Is there existing code I should reuse? Only after those questions are answered does it start writing, cutting redundancy at the source.

The rules follow a seven-step ladder: platform-native capability, standard library, codebase reuse, installed dependencies, one-liner, minimum working code. The model stops at the first rung that applies and never goes lower. Four intensity levels (Off, Lite, Full, Ultra) switch in real time right in the composer, taking effect on the very next message. Toggles and levels persist across restarts. One rule is non-negotiable at every level: trust-boundary validation, loss-prevention error handling, security, and accessibility are never stripped.

Built for developers and teams who work daily in DSH Desktop, want to keep token spend in check, and prefer a lean code style. The goal is not to debate your requirements back; it is to make the AI behave like a lazy-but-reliable senior engineer: reuse before rewrite, one line before fifty, but never drop the guardrails that matter.