dsh-consult

Agent teams excel at splitting work and parallel execution—more hands on the same task. But the bottleneck that actually stalls projects is often a directional judgment: which architecture to commit to, where the security boundary sits, whether a release is truly ready. dsh-consult reserves model diversity for exactly those moments. At an explicit checkpoint, two to five independently configured models receive the same evidence brief and answer the same question, anonymously, without dividing labor or debating with one another.

The main agent gathers evidence, implements, and validates before invoking the panel. What it receives is a set of de-branded, independent opinions. It compares where they converge, notes where they diverge, catches what one model missed that another did not—and makes the call. Consultants have no tools, no workspace access, and token spend occurs only at the moment the operator deliberately triggers the checkpoint.

It is built for engineers on DeepSeek Harness facing high-leverage decisions: architecture selection, security and data-retention review, migration or rollback planning, release-readiness gates, or any scenario where a conclusion might be one model's preference rather than a robust judgment. Routine edits and already-settled plans do not need a panel—the design is a deliberate, high-value checkpoint, not a tax on every turn.