AI News
All stories
Reported

DeepSeek open-sources dsh-superagent, a research prototype for ongoing coding tasks

DeepSeek has released dsh-superagent, a research prototype that manages long-running coding tasks on a single GitHub repository, each with its own agent session.

Open Source · · 2 min read

DeepSeek has open-sourced dsh-superagent, a research prototype that manages ongoing coding tasks for a single GitHub repository. It runs on DeepSeek Harness, and each task gets its own agent session and working directory. DeepSeek labels the code as research rather than a production-ready product. The accompanying article says the published implementation differs from the company's internal version, and some of the ideas it discusses are still future work.

The prototype targets a common problem: a coding task rarely ends with the first patch. A reviewer may ask for changes, a test may fail, or the agent may need a person to clarify intent. According to DeepSeek, dsh-superagent turns each accepted request into a tracked task assigned to a worker agent. A request can start from a GitHub issue. The system checks whether the request is authorized before it creates the task. The worker inspects the repository and edits code inside a directory dedicated to that task. A configured preparation command can create a Git worktree, which gives the worker a separate checkout and branch.

Follow-up messages route back to the same worker instead of being treated as new requests. If the agent needs to know whether a failing test reflects a regression or an intentional behavior change, it can ask a question and wait. The reply returns to the existing task, and the worker keeps its earlier investigation. DeepSeek says the system links tasks to their GitHub conversations, including pull-request discussions, so later feedback reaches the worker responsible for the change. Each task keeps the original request and its authorized scope, the worker session and execution history, pending follow-up messages, and the latest progress report or question for a person.

DeepSeek separates incoming information from permission to act. Allowed senders can make explicit requests, while proactive work must satisfy configured policies, so an ordinary discussion or automated notification should not become a coding task. Reading a related issue supplies context but does not authorize the worker to implement everything discussed there. Authorization to start a task is separate from approval of individual commands. DeepSeek says its workers run commands as the operator's OS user without per-command approval.

Recovery works in a similar way. dsh-superagent preserves inputs that arrived but have not yet reached the worker, and it restarts interrupted assignments with a recovery note. A service restart can still terminate running processes, so recovery means resuming the investigation rather than preserving every command that was in flight. A restarted worker can inspect the checkout and execution history to establish what finished before it continues. A web panel lists tasks waiting for answers or review, along with blocked and failed work.

To try it, users clone the prototype repository from GitHub and open a coding agent in that directory. DeepSeek recommends asking the agent to read SETUP.md and set up dsh-superagent together. The setup guide covers the required environment and repository configuration.

What changed

DeepSeek open-sourced dsh-superagent, a research prototype that manages ongoing coding tasks for a single GitHub repository and runs on DeepSeek Harness.

Type: New capability · Applies to: dsh-superagent, DeepSeek Harness

Why it matters

Companies that connect coding agents to repositories gain a model for task tracking and recovery, but the absence of per-command approval means the access risk falls on the operator's environment.

What you can do

Review the documented execution model and run the prototype in an isolated environment before connecting it to any production repository.

Original source

We have no link to the original source for this story yet.

Where we saw it

1 source

AI Engineering

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works