by qiz029
A DeepSeek coding agent harness: persistent shell, Ultra subagents, auto approval, Chrome MCP and session telemetry
# Add to your Claude Code skills
git clone https://github.com/qiz029/dscodeLast scanned: 9/21/2026
{
"issues": [
{
"file": "README.md",
"line": 77,
"type": "remote-install",
"message": "Install command (remote install script piped to a shell — review the source before running): \"curl -fsSL https://raw.githubusercontent.com/qiz029/dscode/main/install.sh \\| sh\"",
"severity": "low"
}
],
"status": "PASSED",
"scannedAt": "2026-09-21T09:37:45.185Z",
"npmAuditRan": true,
"pipAuditRan": true,
"promptInjectionRan": true
}See how dscode compares with popular alternatives.
dscode is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by qiz029. A DeepSeek coding agent harness: persistent shell, Ultra subagents, auto approval, Chrome MCP and session telemetry. It has 1,021 GitHub stars.
Yes. dscode passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/qiz029/dscode" and add it to your Claude Code skills directory (see the Installation section above).
dscode is primarily written in JavaScript. It is open-source under qiz029 on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh dscode against similar tools.
No comments yet. Be the first to share your thoughts!
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
Write code in your terminal. Plug scripts into a live session. Hand tasks between agents.

Core features · What's new · Quick start · Commands · Changelog · Docs
DSCODE is a terminal coding agent for macOS, built on DeepSeek Harness. A persistent shell reads and writes code and runs tests; the TUI, the CLI and your scripts all share one session runtime instead of each starting their own. It installs as a pinned, reproducible harness—DSH dependencies, TUI and plugins are versioned and verified together.
Most coding agents work alone. DSCODE is built on the opposite assumption: sessions on your machine are visible to each other, so one can hand a task over, another can review the diff, and an independent reviewer decides approvals from your instruction rather than from a rule table.
See it in 60 seconds. Three things most coding agents cannot do, and where to look:
| In one terminal | What it shows |
|---|---|
/btw why is the cache cold on the first turn? |
A side question runs in its own read-only child session and answers in a panel; the exchange never enters the main conversation. |
dscode send <session-id> --steer "review the change in parser.ts and reply" |
Work handed to another session on this machine; it can read your transcript, answer, and hand the result back. |
/permission auto-review · /review-usage |
Approvals decided by an independent reviewer from your instruction, with what it allowed and what it cost. |
The 90-second demo script has the shot list, the exact commands, and how to record it.
| Feature | What you get |
|---|---|
| Agentic coding loop | A persistent shell that keeps cwd, environment and background jobs, plus file edits, search, patch application and tests. Project instructions, skills, plan, goal and hooks are wired in. |
| Session bridge | Start a task in the TUI, then add requirements, read output or subscribe to progress from another terminal (dscode sessions, send, read, watch). Every source enters the same runtime and context, and readers never take the session write lock. |
| Agent-to-agent tasks | The agent finds, reads and messages other sessions with list_sessions, read_session, send_session and reply_session, choosing queue, steer or defer. A persisted mailbox, retry de-duplication and a finite budget bound message loss, double processing and wake-up loops. |
| Session cards | Each session advertises its project, workspace and the topics of its last five user requests—enough to pick the right collaborator without reading its transcript. Cards describe what the user asked for, not conclusions. |
| Cross-session memory | Reusable experience is extracted in the background and retrieved together with its workspace and source messages. Memory can be disabled per session or globally, and its model usage is tracked separately. |
| Effort and sub-agents | Ultra uses the model's max reasoning and decides how far to investigate, delegate and verify; below Ultra the agent can still delegate, but only when a task clearly warrants it. Parents pick a separate effort per child, and children that edit can work in isolated Git worktrees created from a clean HEAD. |
| Independent review | After a code change passes its relevant checks, the agent sends the Git diff—or, outside a repository, the changes since a snapshot taken when the task began—to a separate read-only model and fixes concrete findings before ending the turn. /review runs the same reviewer by hand, scoped to the staging area, a base branch, a commit or a path. |
| Non-interactive runs | dscode exec "prompt", or git diff | dscode exec "review this", runs a full turn in scripts and CI: the reply streams to stdout, tool activity and the session id go to stderr, and the exit code reflects the turn. --json and --resume are supported. |
| Terminal UX | Six interface languages, select-and-copy text, scrollback through history, a verbose view of thinking and tool calls, an input area pinned to the bottom, a sub-agent overview, and footer TPS / context / cost / cache figures. |
| Guardrails | A workspace-write sandbox by default and Auto permission review by an independent model, switchable to human approval. Ultra grants no extra permissions and Computer Use keeps human authorisation. |
| Agent email | A local [ToAgent] inbox shared by all sessions: Enter steers a message into the live session as user-selected context, and with Gmail configured the agent can send mail through the same approval flow. |
0.7.32 — OpenCode Go now signs in only with an OpenCode account: /opencode login shows a code, opens the console and signs in in the background. The login renews itself and sends requests to the inference endpoint the console names for your organisation. The console API key from 0.7.31 is no longer read, so run /opencode login once. The footer shows the subscription's usage of its five-hour, weekly and monthly caps, and when an exhausted one lifts. The consent page names the OpenCode CLI, whose login client DSCODE uses.
0.7.31 — OpenCode Go joins /provider: /provider opencode-go stores the key from the OpenCode console and serves Go's chat-completions models — DeepSeek, GLM, Kimi, MiMo, LongCat, Hy and Space Bunny — with the reasoning efforts each one accepted on the live gateway, the session id Go asks for, and reasoning passed back across tool calls. A Go session searches the web through Exa, as OpenCode does, with no extra key. Qwen, MiniMax and Go's Grok and GPT models use other protocols and are not offered yet.
0.7.30 — /delegate <task> turns the main agent into a coordinator: it plans the task on a board with priorities and dependencies, runs ready parts as child agents in isolated Git worktrees, verifies each result and merges the accepted ones in dependency order, staged and uncommitted. /delegate-dashboard shows the board as a colour-coded kanban — pending, running, verifying, complete. Parents now run up to five children at once below Ultra and twenty at Ultra, and the welcome snowflake is redrawn.
0.7.29 — /account signs the DeepSeek route in with a DeepSeek account through the system browser: the grant is stored on this machine, a signed-in machine needs no DEEPSEEK_API_KEY, and the command also reads the account identity, recharge and bonus balances, and signs out. The profile composes a loopback callback server for it, which carries a route only while a sign-in runs; DSCODE_ACCOUNT_LOGIN=0 removes it. Two composition fixes came with it: the host-plane workflow engine is disabled again after an upstream rename silently voided the patch that disabled it, and the redundant PTC runtime row is gone now that the base composition ships one.
0.7.28 — The DSH runtime moves from 0.1.5-rc.2 to 0.1.7-alpha.2, bringing sandbox escalation with justification on the retry shell, durable image offload, MCP resource tools, subagent permission inheritance and compaction retry recovery. The reviewed permission preset is renamed auto → auto-review, because DSH now reserves auto for its own integration: use /permission auto-review, dscode exec --permission auto-review and permission: auto-review. Computer Use keeps live activation and its native helper, but a resumed session regains its execution tools only once the computer-use skill is loaded again.
0.7.27 — Fixes fresh Hub installation failing after Harness composition generates its empty cordis.yml. The launcher carries a narrowly patched Hub CLI, and release verification now checks the native locked installer, launcher first start and Hub doctor.
0.7.26 — Triggers now support fresh or persistent sessions, durable cron and delay jobs, and supervised script sources. Manage them from /trigger, the CLI or agent tools with the same workspace and approval boundaries. The TUI stays in DSCODE mode and model switches reset unsupported effort values to the target default. TPS uses API-reported tokens with request-count exponential smoothing; runtime metrics and skills sit on the left, while cost, DeepSeek balance, peak/off-peak and cache align right or move together onto a third row. OpenRouter content-policy rejections now explain why the turn stopped.
0.7.25 — Running the installer again now keeps a tar installation current instead of refusing: an older one is updated through its own dscode update (state migrated, th