fullstack-expert

Coding agents write fast, but speed often means skipping verification, self-attesting results, and drawing no hard boundaries around sensitive operations. fullstack-expert does not bolt on a second agent framework. Instead, it lays a disciplined engineering loop over the Harness composition you already run: understand, inspect, design, plan, implement, test, review, verify — each phase backed by a deterministic state machine and explicit evidence records, so "the agent finished" becomes auditable fact rather than a claim.

Two deep model-facing operations carry the weight. fullstack_plan performs a bounded repository inspection and produces a vertical-slice plan; fullstack_check runs read-only structure, command, and optional Supabase detections, recording non-verification review facts as explicit evidence. A permission classifier and an approval-aware pre-execute gate delegate sensitive operations to native Harness approval and fail closed when approval is unavailable. A complexity governor weighs requirement scope, reuse, native platform, dependencies, and safety controls before letting work proceed. Critically, verification evidence must come from a native Harness check — V0.1 rejects self-attested "pass" records outright.

If you ship features, fix bugs, refactor, or build applications through a DeepSeek Harness workflow and want the agent\u2019s behavior to be traceable, bounded, and evidence-backed, this plugin is a lightweight methodology layer, not another heavy framework. It reuses your existing dsh-fs, dsh-shell, dsh-tools, and dsh-skill rows, stays provider-neutral with no hard-coded LLM adapter, and plugs in without reshaping your stack.