One terminal shell, many coding agents, and an independent referee

What it does

DSCODE is a terminal-based coding agent for macOS built on the DeepSeek Harness. It wraps a persistent shell, file editor, and test runner into a single shared runtime that the TUI, CLI, and external scripts all attach to simultaneously. Rather than firing off one-off tasks, it keeps state alive so agents can iterate, delegate, and review within the same long-lived session.

The interesting bit

The project assumes most coding agents work alone, and deliberately does the opposite: every session on your machine is visible to the others. Agents can hand off tasks via a persisted mailbox, delegate to child agents working in isolated Git worktrees, and submit their changes to an independent reviewer model that approves or rejects based on your actual instructions—not a hard-coded rule table.

Key highlights

• Persistent shared runtime: the TUI, CLI commands, and scripts all share one live shell session with full cwd and environment state.
• Agent-to-agent delegation: parent agents spawn children in isolated Git worktrees, verify the results, and merge them back; a colour-coded kanban dashboard tracks progress.
• Independent diff review: a separate read-only model inspects Git diffs or workspace snapshots against your instructions before a turn ends.
• Cross-session memory and discovery: background extraction of reusable experience, plus session cards that advertise recent topics so other agents know whom to ask.
• Multi-provider routing: native DeepSeek Harness integration with support for OpenCode Go, OpenRouter, and curated models from Anthropic, OpenAI, Google, and xAI.

Caveats

• macOS 14+ only; Node.js 22.19+ and Google Chrome are strict, explicitly stated requirements.
• Permission naming and sandbox behaviour have shifted across recent releases—auto was renamed auto-review in 0.7.28, and a resumed session regains Computer Use tools only after the skill is manually reloaded.
• Some features remain incomplete: trigger failures currently notify nobody, per the 0.7.24 release notes.

Verdict

Worth exploring if you want a long-running, multi-agent coding environment inside your terminal with built-in oversight and sandboxing. Skip it if you are not on macOS or prefer lightweight, stateless LLM wrappers that do not manage their own runtime and inter-agent mailboxes.