DeepSeek Harness UX

English | 中文

A calmer Web experience for long-running work in DeepSeek Harness.

DeepSeek Harness is already an extensible coding-agent harness. This community edition concentrates on the part people watch and operate while the agent is working: session recovery, progress, multi-turn reading, actions, files, and workspace navigation.

It is not a different agent. It preserves the upstream plugin architecture, agent loop, tools, permissions, sandboxing, and main-model input. The changes are intentionally concentrated in the Web presentation layer.

This is an unofficial, community-maintained project and not a DeepSeek distribution. DeepSeek Harness and related names belong to their respective owners.

Why use it?

Long agent tasks produce a lot of useful evidence, but presenting every event at equal visual weight makes it hard to answer three basic questions: Is it still working? What is it doing now? Is it finished?

DeepSeek Harness UX turns that event stream into one stable process surface:

• While a task runs, the current stage and a short trail of completed stages remain visible instead of becoming a wall of logs.
• Technical reasoning, tool calls, questions, and approvals remain available under Run details; they are not deleted or disguised as the final answer.
• When the task finishes, the process folds automatically and the result becomes the reading focus.
• Copy, feedback, and Branch actions stay available without permanently occupying every message row.
• Historical sessions, workspaces, and produced files receive explicit loading, recovery, sorting, and opening behavior.

The result is still DeepSeek Harness, but it is easier to trust during a long run and easier to read after the run ends.

Difference from upstream DeepSeek Harness

Area	Upstream DeepSeek Harness	DeepSeek Harness UX
Primary goal	General-purpose agent harness, plugins, tools, runtime, and official Web UI	A community Web UX edition built on the same harness
Agent execution	Upstream agent loop, model provider, tools, permissions, and sandbox	Kept aligned with upstream; UX work does not redefine the agent strategy
Running-task view	General event and tool presentation	One semantic process surface with visible current progress and expandable technical details
Completed work	Conversation and event history remain available	Process history folds automatically so the final result leads
Long logs	Browser-native nested scrolling and transcript flow	Bounded details own their scrolling; the composer remains anchored without blank-page jumps
Conversation reading	Standard message controls and Markdown rhythm	Hover-only actions, clearer turn separation, tighter long-form typography, and safer answer headings
Sessions and workspaces	Upstream session and workspace capabilities	Recovery and retry states, activity indicators, newest-first presentation, and calmer workspace density
Deliverables	Upstream produced-file pipeline	Clearer deliverable discovery and opening for common formats, including PDF and Web files
Distribution	Official packages and upstream source	Independent source edition; no package is published from this repository under the @deepseek-ai scope

This comparison describes the source baseline used by this repository. Upstream continues to evolve, so some improvements may eventually overlap.

Display assistance does not change the answer

Some stage labels and answer headings are refined by bounded, auxiliary model calls in the Web presentation service. The service reads a limited slice of already-recorded process evidence and returns presentation metadata only.

These calls do not modify the main model's system prompt, user message, tools, reasoning, authored answer, or conversation history. They can add a small amount of presentation-only token use and latency. If the auxiliary path is unavailable, the task continues and the UI uses a local fallback. See Web presentation and its design record.

Which version should I choose?

Choose the official DeepSeek Harness when you want the latest supported release, official package distribution, or primarily use headless and CLI workflows.

Choose DeepSeek Harness UX when the Web UI is your main workspace and you regularly run multi-step tasks where progress clarity, session recovery, compact reading, and deliverable access matter.

Run

This community edition runs from source. Requirements:

• Node.js ^22.19 or >=24
• pnpm 11
• A DeepSeek-compatible API key

git clone https://github.com/ayuanwong/deepseek-harness-ux.git
cd deepseek-harness-ux
pnpm install
pnpm run build
pnpm run dsh -- web --port 3081

Open http://127.0.0.1:3081, then add a model provider under Settings → Models and create a session. Replace 3081 if that port is already in use.

This repository currently ships as a complete source edition. It is not yet a drop-in Fabric patch or a separately published npm plugin for a clean upstream checkout.

Project status

The current edition follows the 2026-08-12 DeepSeek Harness source snapshot plus the UX work documented in this repository. It is suitable for local evaluation and community development, but it does not carry upstream support or compatibility guarantees.

Privacy

Session logs stay local by default. Do not commit .env, .npmrc, API keys, local sessions, build output, or profile data. Review the upstream telemetry settings before enabling a non-default telemetry mode.

Development

Read AGENTS.md, the development guide, and architecture before changing packages.

pnpm run lint
pnpm run build
pnpm run hygiene
pnpm run doc-sync

License and attribution

This repository is derived from DeepSeek Harness and preserves upstream notices. The project uses the BSD 3-Clause license; third-party dependencies and license terms are listed in THIRD_PARTY_NOTICES.md.