DSH‑Agent网关整体方案|v1.1 定稿
This is an independent personal architectural preprint for DSH‑AI‑Gateway (DeepSeek‑Harness agent‑internal gateway), no production‑ready source‑code implementation is provided. Analysis is built upon Kuangchao System Architecture Meta‑Theory v1.0‑public, baseline marker: KUANGCHAO‑BASE‑V1P0‑ROOT‑R4. This gateway acts as the single external interface inside DSH agent runtime. It follows the core principle of zero business control: the gateway only performs message receiving, format classification, tagging, buffering management and routing, without making any business semantic judgement. All business decision‑making belongs exclusively to the upstream Agent‑Loop LLM. The design defines bidirectional streams: forward stream (external signals entering DSH) and reverse stream (LLM instructions dispatched outward). Two‑level buffer mechanism with priority‑based eviction strategy is implemented. Three output channels are distinguished: confirmed channel, uncertain channel and asynchronous channel. An isolated feedback channel is responsible for buffer warning, overflow and internal‑exception reporting, separated from normal business traffic. This document fixes three mechanism gaps and eliminates ambiguous definitions from v1.0. Attached artifacts include hazard‑test‑parameter mapping table, on‑site troubleshooting worksheet template, three‑way defect classification table, plus a complete set of architecture‑level fault‑injection test suites covering steady‑state, offset‑state, conflict‑state, split‑state and crash‑state. Fifteen Mermaid flow‑chart & sequence‑diagram source codes are embedded inside markdown as architectural drawings. Disclaimer: Purely paper‑based architectural deduction, without real‑machine engineering verification. For technical exchange only, cannot be directly put into production. Not the official product design of any organization.
Authors
- Kuangchao (ORCID: https://orcid.org/0009-0003-5601-9448)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24151865
- Primary Topic
- Software System Performance and Reliability
- Type
- preprint