dsh-humanize

Complex agent workflows often lack structured mechanisms for tracking, recovery, or result verification, leading to inefficiencies and errors. dsh-humanize addresses this by embedding Humanize discipline into a DeepSeek Harness agent preset, centering workflow construction around the agent itself. This means tasks are organized into explicit flows with judge commands at each stage, hash-locked identities for immutability, and review gates to ensure only approved steps proceed—moving away from memory-dependent execution to a more auditable and reliable system.

The plugin's core capabilities revolve around its flow kernel: it supports multi-stage goals where each phase includes a judge command that must exit successfully to count. Locks are generated via semantic normalization and SHA-256 hashing to prevent tampering, while review gates use HMAC signatures for binding decisions that are irreversible. Event logs (events.jsonl) serve as the single source of truth, enabling resumable runs that pick up from where they left off after interruptions. With a suite of 15 model tools (e.g., flow_draft, flow_lock, edit_anchored), it manages the entire flow lifecycle and allows customization of tools, skills, and personas to adapt to domains like quantitative research, academic writing, or competitive modeling.

dsh-humanize is ideal for developers, researchers, and domain experts who need structured, recoverable agent workflows. Whether you're building coding agents or specialized systems for multi-phase tasks, this preset offers a robust foundation for enforcing discipline, ensuring traceability, and enabling customization—helping you focus on task logic rather than reinventing workflow infrastructure.