DeepSeek Harness 一周更新汇总：v0.1.5 正式定档，v0.1.6 带着八个新入口杀到DSH Weekly Update Roundup: v0.1.5 Locks In, v0.1.6 Arrives With Eight New Entry Points

The release history of the deepseek-ai/deepseek-harness repository on GitHub shows 7 versions shipped in the 8 days from September 8 to September 15:

DSH updated densely this week: v0.1.5-rc.2 was finalized on September 10, the first time the model and the shell "evolved together"; on September 15, v0.1.6-alpha.1 handed the Agent a pile of new "hands," and two days later alpha.2 fitted users with a "console"; and hidden in the source is an official desktop app already in internal testing. Seven versions in one week—a pace that looks like a rush, when in fact every step is traceable.

First, this week's release cadence

The release history of the deepseek-ai/deepseek-harness repository on GitHub shows 7 versions shipped in the 8 days from September 8 to September 15:

• September 8: v0.1.5-alpha.1
• September 9: v0.1.5-alpha.2
• September 10: v0.1.5-rc.1 and v0.1.5-rc.2 (two in one day)
• September 15: v0.1.6-alpha.1
• September 17: v0.1.6-alpha.2

rc.1 and rc.2 landed back to back on the same day with a clear division of labor: rc.1 collected the model adaptations and UI capabilities from the alpha phase (V4.1 Flash became the default model for new sessions), while rc.2 added just one experience improvement—feedback submission now uses a confirmation dialog and preserves what you've filled in if it fails. The former is the "content freeze for the stable release," the latter the final polish before shipping, and two days later 0.1.5 was locked in.

According to releasealert, the repository has published 17 versions since August 17, averaging one every 1 day 18 hours. That cadence is unusual for the 0.1.x stage.

v0.1.5: the model and the shell start to "evolve together"

The core of v0.1.5 is deep adaptation for DeepSeek V4.1 Flash. In the official announcement's words: the model was specially trained and optimized across different Harness configurations. That sentence deserves unpacking—according to media breakdowns, training covered three Harness configurations:

• Standard mode: the full set of tools, Skills, files, Shell, and task planning—the complete form you use day to day
• PTC mode: programmatic tool calling. The model doesn't "call one tool, wait for the result, then think again"; it generates a block of code at once and orchestrates multiple rounds of tool calls to run in batches
• Minimal mode: the shell stripped down as far as possible, leaving only the terminal and file editing, to see how much the model can do on its own

It used to be that you trained the model first, then wrapped an Agent framework around it to see whether it worked; now the shape of the Harness is taken into account during model training, and the two sides co-evolve. That also lines up with DeepSeek's long-standing framing: "Agent = Model + Harness."

v0.1.5 also has a very engineering-flavored detail: preserve the KV Cache as much as possible when updating the system prompt. In long-horizon Agent tasks, changing a rule midway doesn't force a large stretch of downstream context to be recomputed. Ordinary chat won't notice, but for users running long tasks it's a real speedup.

The UI side gained a batch of capabilities at the same time: the Web interface supports uploading images and PDFs, which the model reads on demand through dedicated file tools; the right sidebar natively previews Markdown, HTML, PDF, images, and common code formats; and plugin authors got standardized extension points—the left sidebar can register global entry points, while the right sidebar supports custom tabs with multiple tabs, split screen, and full screen.

Also open-sourced alongside is experimental Agent Teams: parent and child Agents communicate in both directions, the main Agent can assign different models and reasoning efforts to different subagents, and the main Agent aggregates the results.

v0.1.6-alpha.1: eight new entry points, and more "hands" for the Agent

v0.1.6-alpha.1 on September 15 differs from v0.1.5-rc.2 by about 800 commits (including features, fixes, tests, and merge records—not the same as 800 features). Eight highlights:

• Browser Use: integrates three experimental backends—Playwright MCP, Chrome DevTools MCP, and Stagehand—so the Agent can perform real web actions like clicking, typing, and navigating. It supports launching a new browser or attaching to an existing one (attach requires configuring a debug endpoint before you can use existing tabs and login state)
• Computer Use: two experimental paths, Cua Driver's MCP and native, for local machine operations and screenshots, requiring explicit configuration
• SSH remote workspace: the local Harness keeps model calls, permission management, and Session persistence, while using the remote machine's files, processes, and execution environment over SSH. The initial scope is Headless and custom-configuration combinations in POSIX environments, and it requires preparing a remote helper program
• Web Terminal: a new terminal in the web sidebar, with multiple tabs, shell selection, and restoration after a page refresh
• Headless enhancements: new stdin task intake, --session-id to continue an existing session, and --json to emit run events line by line—external programs can finally "understand" what DSH is doing
• MCP resource capabilities: support for resource discovery, resource reading, and URI Templates, distinguishing "performing an action" from "reading content"
• Archived session recovery: a new list of archived sessions in settings, viewable and restorable
• Experimental Auto Review: tool calls are automatically reviewed by the current Agent's model before execution and allowed or denied. Note that every review generates an additional model request

A few boundaries need to be stated clearly. Browser Use and Computer Use are both optional components that are disabled by default; the "preserves login state" claim for attach mode comes with preconditions—not just any browser can be taken over. Auto Review approves calls that then execute with full access, with no file sandbox, and the model may also wrongly allow or wrongly deny, so there is no question of an unattended safety guarantee. Canceling a task is also not the same as undoing an action: clicks and keystrokes already sent out will not automatically revert. In the JSON event stream, body text and reasoning are emitted after a step is submitted, not token by token in real time, so automation programs should check the exit code and the reason the turn ended—seeing final does not mean success.

There are also several configuration-level changes that will affect existing environments: the DeepSeek adapter now defaults to the Messages protocol, so users who manually overrode the old official base URL need to adjust it as documented; Ralph is off by default, the built-in E2B backend is removed, and the PTC package and services are unified under the ptc-runtime series; experimental Team mode uniformly uses spawn_teammate, and the default teammate creation limit rises from 8 to 16; Node PTC now executes in a separate process with an empty process.env, so PTC code that depends on environment variables needs to be rechecked.

v0.1.6-alpha.2: fitting users with a "console"

alpha.2 on September 17 is a "plugins + interface" themed release, with 6 groups and 33 items in the official release notes. If alpha.1 let the Agent operate browsers and computers, this version is about letting people keep them under control:

• Plugin management page: install, change configuration, and start/stop in real time—the plugin ecosystem moves from the command line into the interface
• File change cards: at the end of each session turn, list which files were changed, with per-file diff review in the sidebar
• Sidebar preview of Office files: Word, Excel, and PowerPoint can all be viewed directly
• Sidebar browser mode: open a specified URL in the sidebar
• Open Subagent sessions and preview the submission plan in the sidebar

The experience-side upgrades are dense but small: sidebar layout persistence and terminals that stay connected after a refresh; file preview and terminal available before the first message is sent; context usage moved to the bottom of the input box, expandable for details; faster CLI and Web startup. Among the fixes, two are worth naming: the repair of Messages API URL concatenation (follow-up to the earlier protocol switch) and the long-standing issue of console windows flashing when running PTC and Shell commands on Windows.

Developers should note three breaking changes: plugin dependencies are now resolved at runtime and support runtime uninstall, so loading and unloading logic needs rechecking; creation mode removes Cordis dynamic tool definition in favor of installing persistent plugins through Plugin Manager; and client Sessions support multiple coexisting instances, with related APIs and slots changed. In addition, the default model list drops V4 Flash and V4 Flash Vision Exp—combined with V4.1 Flash becoming the default in rc.1, the model line is clearly converging. The Subagent chain also gained default limits: at most 8 subagents at once and a delegation depth of 1, adjustable in settings.

An easter egg in the source: the official desktop app is already in internal testing

Digging through the repository source, the apps directory now contains desktop and desktop-host alongside cli and web. The official desktop app isn't a rumor—it's already a reality in the repository.

The README is candid: DeepSeek Harness Desktop is a thin Electron wrapper that fully reuses the Web app's behavior and only handles platform adaptation. A few key details:

• Platform support covers macOS (arm64 and x64) and Windows x64; Linux is explicitly not a release target
• A bundled standalone runtime: it ships its own Python, Node.js, and pnpm distributions, with Python preloaded with numpy, pandas, python-docx, python-pptx, openpyxl, Pillow, lxml, and XlsxWriter, installed offline into the user directory on first use—Office document handling that works out of the box
• Three skills registered by default: office-docx, office-pptx, office-xlsx
• It shares sessions, settings, credentials, and workspaces with the CLI, but executable state is fully isolated, with the desktop getting a profile of its own
• Auto-update runs on the Nightly channel, checking asynchronously every 10 minutes, with support for forced-update policies
• Windows uses EV certificate signing plus an NSIS installer; macOS gets notarization with dual DMG and ZIP channels

Judging by the commit history, desktop-related commits landed densely from September 11 to 17: icon design, Windows installer packaging, and a version number already advanced to match 0.1.6-alpha.2. The README uses phrases like "early internal test build" and "internal-test packaging," with the update channel fixed to Nightly—so the stage is clear: the official desktop app is in internal testing, not yet publicly released. For users who don't want to set up an environment, this may matter more than any single feature.

A switch worth a look before upgrading

In v0.1.6, a session log contributor is enabled by default in the DeepSeek adapter request path, sending not-yet-acknowledged session events through the dsh_session_log field. The content may include message text, tool arguments and results, workspace paths, and feedback.

The official explanation is that this field is not added to the model's messages, does not consume model input tokens, and does not change the cache prefix visible to the model; but the content is still sent out with the HTTP request. Not consuming tokens and staying on your own machine are two different things.

Users who don't want it enabled can turn it off explicitly: create privacy.patch.yml and write in the corresponding entry, first use --dump-config to confirm the configuration takes effect, then start with the same patch. Also note that this logging chain is independent of OTel telemetry—turning off one does not turn off the other.

How to read this week's updates

Seven versions in a week, and three threads are clear. Vertically, the model and the shell moved from "iterating separately" to "training together," with V4.1 Flash's dedicated optimization for three Harness configurations as the landmark step. Horizontally, the Agent's execution entry points expanded from files and commands to browsers, the desktop, and remote servers, while Headless hands that whole capability set to scripts and pipelines. The third thread is hidden in the source—the desktop app, the plugin management page, and Office preview all point the same direction: what DSH wants to be isn't a command-line pipe but a complete development workbench.

The v0.1.6 series is still an Alpha prerelease, and "added to the feature list" and "mature enough for daily use" are two different stages. A few practical ordering suggestions: start with low-risk items like Web Terminal, archived session recovery, and Headless; validate browser and local machine operations first with a separate test browser and a disposable test project; leave Auto Review for last, after you've seen its permissions and costs clearly. Plugin authors should check the two breaking changes—dependency resolution and the Session API—before upgrading. Back up DSH Home before upgrading, and keep a way back for important environments.

Safety note: DSH-related capabilities are for authorized scenarios only. When browser login state, remote servers, and local desktop operations are involved, be sure you have the appropriate permissions, and do not use them for any illegal purpose.

This article's information is current as of September 19, 2026. Version information is based on GitHub's official release records (the v0.1.5-rc.1 / rc.2 and v0.1.6-alpha.1 / alpha.2 release notes and the repository source); for individual details, defer to the official release notes.