#Agent Systems

dsh-skin-chatlab 2.x: A Switchable Chat Workbench for DSH Web

Six independently published skins, aggregate bundle loading, avatar identity convergence, partial-install fallback, and a clean uninstall boundary for DSH Web.

From “It Runs” to Maintainable: Architecture and Stability Lessons from Building DSH Plugins

A practical account of developing and debugging ChatLab, Agent Suite, and Quota Router, covering DSH plugin mechanics, common preview-version stability problems, and the boundaries that are easiest to get wrong when building plugins yourself.

Case

12 min read

• #Technology
• #Agent Systems
• #Developer Productivity

Decompose First, Then Schedule: A Review of Multi-Model Task Decomposition, Capability Switching, and Subtask Routing

This paper examines task decomposition, subtask capability classification, model selection, and bounded execution-time fallback, while reducing the broader Agent scheduling problem to an engineering slice of dsh-quota-router.

13 min read

My DSH Workbench: Nine Plugins

DSH has been out only a few days, and I've already installed nine plugins on it. Notes on what each one fills in, how it works, and which one is the most useful.

Short

5 min read

• #Technology
• #Agent Systems
• #Developer Productivity

dsh-skin-chatlab: An Extensible Chat Skin for the DSH Web

Using a base-plus-skin architecture to give the DSH Web GUI a switchable, uninstallable Feishu-style chat skin that touches no existing plugin: workspaces become project groups, sessions become contacts, and the chat becomes bubbles.

Defining the Boundaries Before Bringing AI Capability into an Engineering Organization

A position paper based on the author's experience driving AI infrastructure and pilots in a real engineering organization: it proposes four kinds of boundaries—context, responsibility, authorization, and measurement—and offers a falsifiable pilot protocol plus a minimal harness as a demonstration of putting them into practice.

9 min read

From Prompt to Narrative Production: CanonLoom 0.2.0's Auditable Human-AI Collaboration Architecture for Long-Form Fiction

A system design paper: modeling long-form fiction collaboration as recoverable, auditable, author-approved state transitions, and proposing an evaluation scheme that does not mistake mechanical scores for literary value.