Leitarkkk/dsh-research-nudge

README excerpt

dsh-research-nudge English 简体中文 An advisory research-debt guard for DeepSeek Harness. It notices when an agent spends a long stretch reading, editing, executing, and repeating failures without consulting external evidence, then adds a short reminder to the next model step. It does not call an LLM, perform a search, block a tool, or force the agent to browse. The reminder explicitly allows self-contained work to continue normally. The problem Agents sometimes fall into a local trial-and-error loop: A search of the official docs, an exact error message, or an existing GitHub issue may resolve that uncertainty faster. dsh-research-nudge turns the growing cost of the local loop into a deterministic score called Research Debt . Research Debt: a calculated example This is a hypothetical sequence calculated from the documented default weights, not production telemetry. It shows exactly how the score would cross the threshold: Step Signal Added debt Total : : Read local code ordinary tool +1 1 Edit a file mutation +2 3 Run and fail execution + failure +1 +4 8 Edit again mutation +2 10 Run and hit the equivalent failure again execution + failure + repeated failure +1 +4 +6 21 The default th…

Category