dsh-shadow-mind

In DSH (DeepSeek Harness), the main agent typically handles all reasoning and execution in isolation, with no external perspective to cross-check conclusions or surface blind spots. dsh-shadow-mind introduces a set of "Shadow Mind" agents that are randomly activated after each main-agent turn by a probabilistic heartbeat scheduler. Each shadow carries its own responsibility description, tool allowlist, and per-run timeout, working in parallel to perform fact-checking, independent review, and multi-path cognitive exploration.

Core mechanics include: every shadow receives a sanitized main-session trajectory and an explicit tool allowlist (read-only by default); runs that exceed their timeout are interrupted and their slots released; tool-call arguments are redacted before forwarding so credentials never leak; and finished subagents are cleaned up automatically at epoch boundaries. Users can inspect state, probe a shadow manually, or pause/resume the system via /shadow commands, and can programmatically create, update, enable, or disable shadow definitions through model-callable management tools. Global configuration stays in sync across the DSH web settings panel, write_shadow_config tool calls, and manual edits to settings.yaml.

Ideal for DSH users tackling long-document analysis, multi-step code review, or technical design validation where multi-perspective cross-checking adds real value. It also suits workflow builders who want a composable, configurable cognitive layer attached to their main agent. The plugin is currently a functional prototype with core scheduling, restricted agents, timeout enforcement, and lifecycle management working end-to-end.