dsh-crew

DeepSeek Harness (DSH) is an agent harness that drives DeepSeek models through a Web UI, but many developers spend most of their time in Claude Code, Codex, Antigravity, or Grok. dsh-crew turns DSH agents into dispatchable native subagents inside those hosts: the orchestrator keeps its own model and task UI, while the actual work runs as a DSH worker with its own session, tools, sandbox, and preset — not as a bare model call. This brings capabilities that previously lived only in the DSH Web UI into the coding environment you already use.

The plugin provides flash / pro capability tiers plus off / high / max reasoning effort. tier_policy can clamp every dispatch to one tier at the tool layer, and escalate_on_failure retries a failed flash run once on pro based on actual results. Dispatch guardrails fire before anything spawns: worker nesting is capped at origin-chain depth 3, cycles are refused, and a second worker on an already-held workspace is rejected with the holder's information instead of being silently queued. A jobs board keeps running and finished jobs visible with tier, effort, live progress, and token usage, while the host side shows a HUD segment such as ⚙dsh 1▶pro 2m14s 21.7k/606 ✓3.

In hub mode every worker is a first-class DSH session, visible in the Web UI and grouped by working directory; without a DSH instance, dispatch falls back to a standalone runtime so CI and headless environments keep working. Although DSH's models are text-only, image description prefers DeepSeek's own vision model when available, then falls back to the CLIs you already have — Claude, Codex, Grok, Antigravity — or any OpenAI-compatible API you configure; image generation borrows the same CLIs' brush, and pasted images stay visible in the conversation while reaching the model as text.

This plugin is for teams already using Claude Code, Codex, Antigravity, or Grok who want to bring DSH agents, tools, sandboxes, and session history into their existing workflow. It is also a good fit for running DSH workers in CI or headless environments, or for anyone who wants tiered model dispatch, parallel fan-out, and observable job state without replacing their host UI.