DeepSeek Harness 0.1.1: images, subagents, Web UI upgrades

DeepSeek Harness shipped four releases between August 17 and 21, 2026 — 0.1.0-rc.7, 0.1.0-rc.8, 0.1.1-rc.1, and 0.1.1-rc.2 — with the latest being 0.1.1-rc.2. Each release has its own focus: multimodal image input, Codex and Claude Code subagents, and Web UI plus storage improvements. Before upgrading, note that the SQLite storage format changed in rc.8 — back up your ~/.dsh data first.

Overview: four releases in August and how to read this article

This article walks through each release in turn, listing every change under New / Fixes / Improvements, and closes with upgrade notes. DeepSeek Harness (DSH) is DeepSeek AI's open-source agent framework built on an "everything is a plugin" philosophy (What is DSH). Between August 17 and 21, the team published four releases on npm. The timeline and one-line positioning of each (source):

Below, each version is expanded in full.

0.1.0-rc.7 (2026-08-17): subagents in Job Panel, plugin settings cards, reasoning effort

rc.7 moved Codex and Claude Code subagent tasks into a unified Job Panel, opened settings cards to plugins, and added a low reasoning effort for DeepSeek models.

New

• Plugin settings cards: plugins can now register their own settings cards, so their configuration is managed right in the settings UI without jumping to external pages (source).
• Subagent tasks in Job Panel: execution status of Codex and Claude Code subagents shows up in the Job Panel, so all tasks are tracked in one place.
• Durable image attachments for MCP/ACP: image attachments passed through MCP/ACP tools can now persist.
• Nested image forwarding in PTC Mode: PTC mode (formerly Code mode) can forward nested images.
• DeepSeek low reasoning effort: a new low tier for the reasoning effort parameter, with high still the default; use low when you need faster responses.
• Built-in preset renamed: the English Code mode preset is renamed PTC mode — functionality is unchanged, but old docs and configs should use the new name.

Fixes

• Persistent Bash lag in the Minimal preset.
• Stack overflow when paginating large histories.
• Sessions that could not continue after max-tokens truncation.
• Cursor misalignment in Safari.
• Issues around node-pty 1.2 beta.

Improvements

• Question cards are collapsible and keep your draft, so collapsing never loses input.
• Better interactions in the dynamic Cordis plugin panel.

0.1.0-rc.8 (2026-08-19): multimodal image input, on-demand subagents, SQLite rebuild

rc.8 is the biggest update of the batch: native multimodal image input, on-demand Codex and Claude Code subagents, and a rebuilt SQLite backend — faster and smaller, but with an incompatible storage format.

New

• Multimodal image input (source):
  • DeepSeek model adapters can enable native image requests via configuration;
  • commands such as /goal and /plan accept image input directly;
  • the @ menu in the composer references files and sessions.
• On-demand subagents: Claude Code and Codex are both packaged as Profile Bundles and install on demand with no manual dependency setup.
• Codex non-interactive permission mode: Codex supports a non-interactive permission mode and multiple named instances.
• Windows PTY persistent PowerShell: a persistent PowerShell session on Windows, enabled by default in the Minimal preset.
• Python SDK covers the 4 built-in Agent presets.

Fixes

• Model requests failing due to oversized images or excessive accumulated image payloads.
• Leftover prefixes after cancelling streamed generation.
• Compatibility issues with the OpenAI-compatible gateway.

Improvements

• Layout: home directory abbreviated as ~, narrow-screen composer layout, feedback UI.
• Interactions: sidebar search focus response, workflow panels, model picker, retrying failed local file opens.
• Search: web_search supports concurrent queries.
• Subagents: reportDelivery reports back promptly and wakes the parent task.
• Startup: local dsh web opens the browser automatically; dependency download size is reduced.
• Performance: forking large-history sessions is faster; the SQLite backend reads, writes, and forks faster and uses less storage, but the data format is incompatible.
• Branding: the official brand usage guidelines were published; "DeepSeek Harness" is a registered trademark.

0.1.1-rc.1 (2026-08-21): vision model, multiline input, sandbox security fix

rc.1 completed the model side of the multimodal story — a new DeepSeek-V4-Flash-Vision-Exp vision model — and fixed a sandbox security vulnerability.

New

• Multimodal vision model: the DeepSeek adapter adds the vision model DeepSeek-V4-Flash-Vision-Exp (source).

Fixes

• @ reference layout issue.
• Bubblewrap sandbox vulnerability: confined processes inside the sandbox could escape via /proc/<pid>/root; this is now patched.

Improvements

• Markdown tables adapt responsively.
• Cache hit rate displays full precision at 99.x%.
• Subagent session titles can be switched.
• ask_user_question supports multiline answers: auto-wrapping, with Shift+Enter for manual newlines.

0.1.1-rc.2 (2026-08-21, latest): image upload and preprocessing improvements

rc.2 is the current latest release, polishing the last mile of the multimodal experience: image uploads now prefer the Files API with file reuse, and images are auto-resized and converted before sending. It shipped on the same day as rc.1 and is a small experience-focused release.

Improvements (experience)

• Files API uploads and reuse: the DeepSeek adapter prioritizes uploading images via the Files API and reuses already-uploaded files, so the same image is not uploaded twice (source).
• Image preprocessing: images are auto-resized and converted to the format the model requires before they are sent.

Upgrading to the latest release: four steps, back up first

Upgrading takes four steps: check versions → back up → upgrade by install method → verify and check plugins. The SQLite storage format changed incompatibly in 0.1.0-rc.8, so step two is not optional.

Step 1: check your current version and the latest version

# 1. Your current running version
dsh --version

# 2. The latest release on npm (0.1.1-rc.2 as of this writing)
npm view @deepseek-ai/dsh version

Step 2: back up your session data

cp -r ~/.dsh/sessions ~/.dsh-sessions-backup

Step 3: upgrade by install method

# npx: nothing is installed locally — re-run to get the latest
npx @deepseek-ai/dsh web

# npm global install: one command updates to the latest
npm update -g @deepseek-ai/dsh

# Built from source: pull and rebuild
git pull && pnpm install && pnpm run build

Step 4: verify the version and check plugins

# 1. Confirm the version is in place
dsh --version

# 2. List plugins in the current profile to spot any load failures
dsh plugin --profile web list

Run your main workflows after upgrading and confirm plugin compatibility. Two reminders:

1. rc is still a preview: an rc suffix means release candidate and interfaces can change; if a plugin becomes incompatible after the upgrade, check its compatible version and reinstall it from the DSH Plugin Hub.
2. Branding: the team clarified that "DeepSeek Harness" is a registered trademark, and third parties should follow the official brand usage guidelines (rc.8).

dsh-plugin-hub · Settings

See How to update to the latest DeepSeek Harness for version checks and more upgrade commands.

Sources: v0.1.0-rc.7 Release, v0.1.0-rc.8 Release, v0.1.1-rc.1 Release, v0.1.1-rc.2 Release