DeepSeek Harness

Summary

DeepSeek Harness is an open-source agent harness that operates by treating all components as plugins, running as a local service via an npm command or directly from source after installation. It is intended for those working with coding agents, and its architecture is described using Cordis design concepts. Development activity was noted by the existence of contributor and documentation materials.

DeepSeek Harness: Everything is a Plugin.

Command line

Values are placeholders taken from each option’s declared default. Nothing is executed here — the output shown is a recording of a run that already happened.

Release history

1. dsh-v0.1.2-alpha.2 Aug 30, 2026 · issue 011

   DeepSeek Harness adds active reminders and schedules in session views, collapsed turn tails, and split usage/time stat pills in chat UI.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch dsh-v0.1.2-alpha.2 https://github.com/deepseek-ai/deepseek-harness.git
   # already have the repo? check out this version:
   $ git checkout dsh-v0.1.2-alpha.2

   • ›Surfaces active schedules in session views via the web UI.
   • ›Lists active reminders in the session header.
   • ›Collapses the turn tail into one clickable meta line in the chat UI.
   • ›Splits turn stats into separate usage and time pills, each opening a detail dialog, in the chat UI.
   • ›Reshapes the usage trigger as an icon-row pill in the chat UI.
   +1 moreshow less

   • ›Assigns CI coverage partitions by recorded file duration for faster targeted test runs.
2. dsh-v0.1.2-alpha.1 Aug 27, 2026 · issue 009

   DeepSeek Harness v0.1.2-alpha.1 adds optimistic message echo, provider-card extension slots, and reduced session persistence size.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch dsh-v0.1.2-alpha.1 https://github.com/deepseek-ai/deepseek-harness.git
   # already have the repo? check out this version:
   $ git checkout dsh-v0.1.2-alpha.1

   • ›Adds provider-card and footer extension slots in ui-settings-models, enabling third-party UI extensions in the models settings panel.
   • ›Reduces session persistence storage size (dsh-session, PR #3048), lowering disk footprint for long-running sessions.
   • ›Introduces optimistic submit echo in ui-conversation, rendering user messages instantly at the tail of the chat stream before server confirmation.
   • ›Migrates browser control in subagent to the Remote architecture, enabling browser operations to be routed through the Remote BFF layer.
   • ›Exposes directory picking through Remote in workspace-controller, making directory-picker operations available over the Remote RPC gateway.
   +1 moreshow less

   • ›Polishes the input trigger menu presentation in the web UI.
   └──▷ BREAKING ON UPGRADE

   • !The apiproxy package removes directory-picker RPCs (commit 6e40876 — refactor(apiproxy)!: remove directory-picker RPCs); callers that relied on those RPC endpoints must migrate to the Remote-based directory-picker.
3. dsh-v0.1.1-rc.2 Aug 21, 2026 · issue 004

   DeepSeek Harness rc.2 unifies the image request pipeline, adds Files API upload reuse, and surfaces downscaled dimension metadata from read_image.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch dsh-v0.1.1-rc.2 https://github.com/deepseek-ai/deepseek-harness.git
   # already have the repo? check out this version:
   $ git checkout dsh-v0.1.1-rc.2

   • ›Adds read_image dimension reporting: the tool now reports downscaled dimensions and the coordinate scale factor when an image is resized before being sent to the model.
   • ›Introduces file-store with MAX_CHAT_IMAGE_BYTES (32 MiB) cap, upload deduplication, expiry-based invalidation, and quota-recovery batch cleanup for the DeepSeek Files API.
   • ›Unifies the 'master' and Files request image pipeline via offloadRequestImagesWithPolicy, consolidating how inline base64 and Files-API-referenced images are prepared for chat requests.
   • ›Expands the source upload envelope, allowing a broader set of image sources to be submitted through the unified pipeline.
   • ›Stores a deterministic canonical image encoding (versioned as REQUEST_IMAGE_TRANSFORM_VERSION = 'request-image-v4') with preferred quality levels of 85 and 80, enabling stable cache and upload-index identity across requests.
   +1 moreshow less

   • ›Decouples files and stream timeouts in the DeepSeek adapter, using separate deadline and idleWatchdog/timeoutOf controls so a slow Files upload no longer races the stream timeout.
4. dsh-v0.1.1-rc.1 Aug 21, 2026 · issue 004

   Adds vision model support, multi-line ask_user_question answers, smart markdown table sizing, and a structured index injection table.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch dsh-v0.1.1-rc.1 https://github.com/deepseek-ai/deepseek-harness.git
   # already have the repo? check out this version:
   $ git checkout dsh-v0.1.1-rc.1

   • ›Publishes the DeepSeek vision model via the llm-deepseek package, adding vision capability to the LLM layer.
   • ›Enables multi-line responses to ask_user_question in the web UI, allowing agents to receive longer free-text answers from users.
   • ›Sizes markdown tables by column count in the web UI, automatically widening wide tables past the standard column boundary.
   • ›Reveals wide-table scrollbars on hover in the web UI rather than painting them persistently, reducing visual clutter.
   • ›Adds a structured index injection table and client boot seams via the webserver package.
   +1 moreshow less

   • ›Preserves near-full cache-hit precision in the web layer.
5. dsh-v0.1.0-rc.8 Aug 19, 2026 · issue 004

   Adds durable Agent Teams runtime, multimodal DeepSeek LLM requests, bulk model-picker selection, bounded multi-query web search, and MCP support in the packaged Python runtime.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch dsh-v0.1.0-rc.8 https://github.com/deepseek-ai/deepseek-harness.git
   # already have the repo? check out this version:
   $ git checkout dsh-v0.1.0-rc.8

   • ›Adds durable Agent Teams runtime (feat(team)) enabling persistent multi-agent coordination across sessions.
   • ›Supports multimodal requests in the llm-deepseek provider, allowing image content alongside text in DeepSeek LLM calls.
   • ›Adds bounded multi-query web search, letting agents issue multiple search queries in a single web tool call with enforced query-count limits.
   • ›Adds MCP (Model Context Protocol) support in the packaged Python runtime.
   • ›Adds bulk model selection to the web UI model picker, enabling users to select multiple models at once.
   +3 moreshow less

   • ›Adds POSIX home-path abbreviation connection handling in the ui-tool layer.
   • ›Delivers subagent reports at the next agent step, improving subagent runtime report timing.
   • ›Adds file-open failure handling in the chat view, surfacing errors inline when attachments cannot be opened.