dsh-tang-governance

Throwing a single coding request at DeepSeek Harness in single-agent mode tends to produce the same failure pattern: ambiguous requirements get rushed into code, review becomes a rubber stamp, execution boundaries rely on prompt convention, and post-incident tracing is nearly impossible. dsh-tang-governance, inspired by the Tang dynasty Three Departments and Six Ministries, structures every request into an auditable pipeline-intake and scale classification, clarification-before-draft, independent review with countersignature, explicit user approval, responsibility-based execution across six ministry domains, independent acceptance, and a user-controlled closure gate-critical boundaries are enforced by runtime code rather than left to prompt adherence.

Core capabilities include scale-aware intake that keeps simple tasks out of the full-process loop, a native clarification interface that pauses for user input before any draft is written, independent Menxia review that can remand concrete unresolved issues, uniform Shangshu scanning across all six ministry domains that dispatches only departments with material contribution, durable structured case dossiers under .tang/cases/ with a cross-case shared knowledge base, and a responsive dashboard showing the organization map, phase timeline, per-agent token usage, and live status. The user can stop work at any time without discarding completed artifacts. The Three Departments can be routed to independent providers and models, while the Six Ministries inherit Shangshu model configuration and adapt reasoning effort per assignment.

This plugin is built for teams and solo developers who have outgrown single-agent mode and want clear role separation, a traceable approval chain, and runtime-enforced governance in their AI coding workflow-especially those already on DeepSeek Harness who need an auditable delivery process, engineering leads who refuse to leave critical boundaries to prompt probability, and technical managers who want every AI-assisted delivery to leave a complete evidence trail for audit and reuse.