dsh-error-improvement

Agents that repeat the same mistake are the most common source of broken trust and wasted tokens in multi-turn tasks. dsh-error-improvement turns user-confirmed lessons into bounded pre-task checks that are injected before every LLM call within a turn, so prevention rules stay visible not just at the first step but also across follow-up tool calls the model makes in the same turn.

Version 0.2.0 closes the loop automatically: the plugin watches tool post-execute hooks, counts failures that share the same tool and normalized error signature, and once a configurable threshold is crossed it intercepts matching pre-execute calls with a one-shot warn reminder or a hard deny. No model or tool capability is ever expanded. For non-trivial solutions, the improve_record_recipe tool persists a verified fix as a success recipe; recipes flagged with asSkill graduate into standalone SKILL.md files, turning raw experience into reusable capability. A built-in context-compaction layer handles overflow recovery with primary and fallback summarization routes, keeping long conversations alive.

If you are running multi-turn tasks in DeepSeek Harness and want your agent to remember where it fell last time and stop falling the same way, this plugin provides a complete pipeline from manual lesson entry to automatic interception—zero external dependencies, zero network calls, and all state kept in the local DSH settings file.