dsh-cortex

Most agent pipelines route every single step to the strongest model, so cost scales linearly with task complexity while quality bottlenecks still sit at planning and acceptance. Cortex reframes the host primary model as a high-leverage supervisor that only plans, decomposes, routes, verifies, and accepts; the dense generation and stat-heavy subtasks are dynamically assigned to cheaper executor models, cutting the average cost per successful task while holding quality steady.

Under the hood, Cortex provides a recursive task tree, four-stage quality gates (Gate0 through Gate3), a four-pool budget engine with automatic degradation, fingerprint-based policy reuse, a failure-recovery decision tree, and a seven-dimension model evaluation system with Pareto-frontier tracking and drift monitoring. For purely statistical subtasks it can bypass the model entirely via the L0 engine for zero-call exact computation; for medium-sensitivity repetitive work, multi-round voting on weaker models approaches strong-model reliability at roughly one-hundredth the cost. All state is persisted and survives restarts, and a bundled Web console plus the Watchdog scheduler round out the orchestration loop.

If your project already runs on DeepSeek-Harness and your tasks involve multi-step planning, cross-model collaboration, or per-token cost sensitivity, Cortex is a drop-in orchestration layer: it installs via a single plugin command, auto-mounts on the next boot with no manual profile edits, and the supervisor drives the entire pipeline autonomously through 13 built-in tools.