dsh-pain-point-check

You are running a multi-turn agent workflow and the model has already fired two shots at the same problem, both missing. Instead of pausing to ask where exactly it is stuck, it keeps engineering the next experiment around its current hypothesis — confirmation-biased design, sunk-cost lock-in, and the urge to close the narrative combine to strip the agent of meta-cognition. The official repeat-tool-reminder is advisory: a nudge the model can simply ignore. What you need is a guard that actually locks the door.

dsh-pain-point-check is that guard. After two non-converged experiments it does three things: injects three mandatory questions into the next agent step — what is the current critical blocker, what did the last negative result actually rule out, and which path is most cost-effective (not necessarily cheapest); denies every non-investigative tool call until the model writes all three answers in its reply text (an investigation allowlist of read, grep, glob, web_search, etc. remains available); and automatically lifts the gate once the answers are detected. Thresholds and the allowlist are configurable in your profile.

Built for developers running long-horizon agent tasks in DeepSeek Harness or sharing workflows across a team. If your agent keeps pressing forward even though the direction is clearly wrong, or you want to turn the negative-result-to-hypothesis-revision loop from a soft suggestion into a hard constraint, this plugin is designed for exactly that scenario.