dsh-agent-hub

dsh allows exactly one AgentFactory across the entire process, so switching agent engines is always a global decision. Want to run Claude Code? You disable the built-in loop and every session in the profile moves with you. For teams that juggle multiple agent tools, that all-or-nothing trade-off feels restrictive.

dsh-agent-hub takes that single factory slot and turns it into a per-session router. You pick an engine in the composer the same way you pick a model: at session creation. From that point on, the session stays bound to its engine; switching simply means starting a new one, and anything still running on the current engine is left untouched. Every engine's native output is translated into dsh's unified Message, ContentBlock, and StreamChunk types, so cross-engine sessions are stored, streamed, resumed, and traced identically. Engine affinity is durable: a resumed session returns to the engine that produced its history, and forked sessions and subagents inherit the parent's engine automatically.

It is built for developers who want to mix the built-in in-process loop, Claude Code, Codex, and Pi within a single dsh profile. The Claude Code engine derives its provider environment directly from dsh's LLM configuration, so a desktop-launched dsh needs no manual variable exports. Codex runs through dsh's subprocess sandbox with AGENTS.md injected via the skill-injection seam. Pi executes in a read-only sandbox, surfacing its context files and skill catalogs through the same channel. And a profile with no SDK installed at all still runs the built-in loop as before.