dsh-see-world

One of the most trust-eroding failure modes of LLM agents: asked about a current policy, price, or software version, the model answers from stale training data without ever looking it up—while simultaneously over-searching on perfectly local tasks like editing a file or running a calculation. dsh-see-world turns this from a matter of model self-discipline into a hard mechanism. As a decision-layer plugin for DeepSeek Harness, it evaluates each incoming message and decides whether fresh external information is needed. When it is, search-before-answer is enforced with cited sources and an explicit I am not sure when nothing is found. When it is not, the message passes through in milliseconds with zero network calls and zero visual noise.\n\nThe plugin decides when to search, not how. Search execution is delegated to whatever capability-layer provider is already present in DSH (Tavily, SearXNG, the built-in web_search tool, and so on). If a provider is detected it simply works; if none is installed the plugin degrades silently, logging its verdict without blocking the conversation. Privacy is a first-class concern: decision logs are local JSONL files that never leave the machine, judgment input is stored as a truncated summary by default, and the judge model can be pointed at a local Ollama instance so the entire evaluation stays on your hardware. A visual settings card exposes trigger sensitivity, domain allow and block lists, per-session search budgets, and a one-click connectivity test, with changes taking effect on the very next turn.\n\nWho it fits best: DSH users juggling mixed workloads of local code and file tasks alongside the occasional need to verify external facts; teams tired of agents confabulating current-events answers and wanting a mechanical guardrail; and privacy-conscious operators who want to audit the reason, latency, and token cost behind every search decision without any data leaving their machine.