Preface

DeepSeek Harness (DSH) native Goal mode is suited for “continuous progression” tasks: the model executes turn by turn until the goal is achieved or the budget is exhausted. However, when tasks span multiple sessions, require dependency ordering, subtask delegation, evidence retention, and explicit verification at completion, Goal mode alone often lacks persistent planning, recovery mechanisms, and completion gates.

Oh My DSH (also known as DSH Autopilot, npm package name dsh-autopilot) is a community-maintained workflow plugin that adds a control plane layer on top of DSH: persistent task graphs, bounded workers, fixed completion checks, failure recovery, and final reporting. This article is based on the repository README and package.json (version 0.1.0-alpha.3) and outlines installation and usage; the SkillHub directory page currently does not show standalone installation instructions, so the following commands refer to the GitHub documentation.

What is this

Oh My DSH is maintained by LiuMengxuan04, with the repository at liumengxuan04/oh-my-dsh. It is categorized as a “workflow” plugin in SkillHub. The project uses the MIT license, is a community implementation independent of DeepSeek, and is inspired by oh-my-codex (OMX). It composes and orchestrates DSH’s Goal, Agent, session, storage, subagent, Skill, and Cordis public interfaces.

In one sentence: it provides a durable, bounded autonomous orchestration layer for DSH long-running development tasks—plans survive restarts, execution has leases and limits, completion requires freeze checks and read-only reviews, and when state cannot be safely reconciled, it enters needs-attention rather than guessing and continuing.

Currently in developer preview 0.1.0-alpha.3; public APIs, plugin composition, configuration, and storage state may undergo incompatible changes between pre-release versions. DSH itself is still under active development.

Core Features

The following capabilities are summarized from the README, all available within a single Autopilot run.

Leases and Budgets

Each run holds a persistent lease, limiting rounds, active duration, validation counts, dynamic packages, sub-agent launches, and the scale of single-run delegation or review fan-out. The default profile is 1,024 Goal rounds and a seven-day active period; the deployment limit is 4,096 rounds and 30 active days. Paused periods do not count toward the active lease.

Requirements Interview and Three-Role Plan Review

After startup, a requirements interview is conducted and answers recorded. Then a dependency graph (DAG) is generated, which undergoes a fixed three-role plan review before execution begins.

Persistent DAG and Bounded Workers

The task graph persists dependencies, attempt counts, blocking items, and evidence. Execution supports managed one-shot delegation, resumable Team workers, bounded Ralph loops, fixed Workflow and Mission queues. Each completed task must include evidence.

Freeze Checks and Read-Only Completion Review

Project check findings are finite and manifest-based; selected check items, workspace, and related root manifest hashes are frozen during the run. If the root manifest changes, validation reports drift instead of silently switching check sets. After the completion strategy is frozen, selected project checks are run, followed by a fresh read-only reviewer inspecting the results; if it fails, a fix loop is entered; if it passes, the exact Goal is completed and a final report is delivered.

Recovery, Control, and Auditing

Supports recovery after compaction, Agent errors, plugin reloads, and process restarts. Provides pause, resume, stop, audit, and completion records. A text dashboard displays Goal, DAG, worker, validation, budget, and cleanup status.

Security Boundaries (Summary)

  • Does not replace DSH permissions or native Goal ownership; client Cordis code still goes through DSH’s native approval flow.
  • selfModification defaults to off.
  • host-only Cordis runs within the DSH process; VMs are not secure sandboxes and should only be enabled in trusted, disposable, credential-free environments or after OS isolation.
  • Cleanup debt blocks subsequent model work until managed resources are successfully removed.
  • Delivery prepares bounded local worktrees and command plans; it does not silently commit, push, open PRs, or merge.

Before enabling Host self-modification, external notifications, MCP, browser access, or delivery integration, read the Autonomy and security in the repository.

Environment and Dependencies

Install in an isolated profile and pin the following supported versions:

  • DeepSeek Harness 0.1.0-rc.6
  • Cordis 4.0.1
  • Node.js ^22.19.0 || >=24.0.0

Installation and Enabling

Install the pre-release package in a DSH profile named web, export configuration, and perform a health check:

dsh plugin --profile web add dsh-autopilot@next
dsh --profile web --dump-config
dsh plugin --profile web exec dsh-autopilot doctor --profile web

Open a top-level DSH session in the project directory to start a run. Uninstallation example:

dsh plugin --profile web remove dsh-autopilot
dsh --profile web --dump-config

Removing the package does not delete existing DSH transcripts or Autopilot storage records.

Plugins run with the current dsh process permissions; before installation, read the source code and MIT license, try it in an isolated profile, and back up important work.

Typical Usage

Starting an Autopilot Run

Explicitly invoke start in a DSH session; bare /autopilot is not an alias for status or start:

/autopilot start --rounds 256 --duration 2d Implement the requested change, add focused tests and documentation, run the repository checks, and report file-and-command evidence.

Viewing Status and Control

/autopilot status
/autopilot dashboard
/autopilot audit --limit 20
/autopilot pause
/autopilot resume [--duration 1d]
/autopilot stop

pause preserves the run but detaches its Goal and runtime bindings; resume requires manual triggering. stop revokes the run and clears matching native Goals, but does not delete DSH transcripts or Autopilot audit history.

Use status for a complete machine-readable run view; audit --json may contain targets, evidence, findings, and trusted Host source code, so use it with caution.

Viewing the Task Graph

The dashboard is a text snapshot, not a custom web graph:

/autopilot dashboard

The snapshot includes task dependencies and statuses, as well as Goals, workers, validation results, budget, and cleanup items; it is generated on request and does not auto-populate when the run ends.

Normal Flow Overview

  1. Autopilot interviews requirements and records answers.
  2. Writes the dependency graph; the plan undergoes fixed three-role review.
  3. Ready tasks are executed locally or via managed workers; completed items include evidence.
  4. Freezes the completion strategy and runs selected project checks.
  5. A read-only reviewer inspects results; on failure, enters a fix loop; on pass, completes the Goal and delivers the final report.

Use Cases and Notes

Suitable for: Repository changes spanning multiple turns and sessions; teams or individuals needing resumable plans, dependency ordering, evidence chains, and explicit completion gates; users who have verified the plugin workflow on DSH 0.1.0-rc.6.

Not ideal for: Single short tasks (native Goal is lighter); environments unwilling to accept pre-release API changes or unable to provide an isolated profile.

Must Note:

  • This is a community project, not an official DeepSeek product; SkillHub is an independent community directory, unrelated to DeepSeek / High-Flyer.
  • Pre-release versions may break compatibility; use an isolated profile, back up important work, and pin versions as listed in the documentation.
  • When state cannot be safely reconciled, the run stops in needs-attention, requiring manual intervention.
  • Building from source can reference the README’s pnpm install, pnpm run check, and pnpm pack processes; for daily installation, use the npm package dsh-autopilot@next.

Conclusion

Oh My DSH, within DSH’s “everything is a plugin” ecosystem, consolidates the persistent graph, bounded execution, freeze verification, and recovery capabilities needed for long-running autonomous development into an orchestration layer. If you are handling cross-session development tasks on DSH that require evidence and completion gates, start by installing dsh-autopilot@next in an isolated profile using the commands above and run a trial with /autopilot start.