DeepSeek Harness v0.1.5-rc.1: V4.1 Flash & Session V3

DeepSeek released dsh-v0.1.5-rc.1 on September 10, 2026. It is the first release candidate in the 0.1.5 line and collects the major user- and developer-facing changes since v0.1.2-rc.1. npm currently assigns it to latest, so an unpinned npx @deepseek-ai/dsh web installs this version.

Release and npm status

Install the npm default:

npx @deepseek-ai/dsh web

Pin the version when reproducibility matters:

npx @deepseek-ai/[email protected] web

V4.1 Flash becomes the default for new sessions

The DeepSeek adapter adds DeepSeek-V41-Flash under the model ID deepseek-flash. It supports text, images, and updates to system prompts in conversation history. New sessions select it by default unless the configuration explicitly names another model.

Existing automation should check the model ID it sends, the selected provider, and any assumptions tied to the previous V4 Flash entry. See the V4.1 Flash Harness configuration guide for a focused migration checklist.

The Web UI handles files as first-class inputs and outputs

Web can upload arbitrary file types, keep background-upload progress visible across conversation switches, and let the model read the saved file through its existing tools. The right Sidebar adds tabs, split panes, fullscreen, and previews for Markdown, code, HTML, PDF, and images.

Agents can also mark files as deliverables. Users can preview them in the Sidebar, open them with the default application, or reveal them in the file manager. The former Detail panel is removed, so Web plugins that placed content there need to migrate.

Subagents are easier to steer

Continuable subagents now accept queued messages. A queued item can be edited, deleted, steered individually or together with the queue, and the subagent can be stopped. Agent Team messages use consistent steer semantics and retain sender attribution and ordering during cross-agent delivery and cold recovery.

For long-running work, this changes the practical workflow: a user can adjust an active delegation without discarding the full subagent session.

Session V3 is the main migration boundary

Supported older logs are migrated into new files while the originals are preserved. An upgraded Session cannot be read by an older build. The persistence API is now owned by lifecycle-scoped SessionHandles, agentLoop.create() is asynchronous, and a lock limits a Session to one process.

Before upgrading a custom integration:

1. Back up Session files and test migration on a copy.
2. Update custom log readers for V3.
3. Replace direct persistence assumptions with SessionHandle lifecycle handling.
4. Await agentLoop.create().
5. Test concurrent-open behavior against the new Session lock.

Plugin and tool defaults changed

• SDK, Headless, and ACP use read, write, and edit for file editing by default.
• Web minimal and Python sdk-minimal provide only a persistent shell by default; str_replace_editor requires explicit opt-in.
• Agent plugins no longer receive ctx.agent; callers pass the Agent explicitly.
• Inbox is now a type-only interface, with pending messages accessed through agent.inbox.
• Web plugins register global panels through sidebar.panellist and main; the former conversation slot moves under main.
• Experimental Agent Teams packages are installable from npm but remain opt-in.

Other changes worth testing

The release also adds proxy-environment support for outbound requests, richer model discovery, an “Open in” Workspace action, independent feedback submission, long-session performance improvements, Windows fixes, MCP pagination-loop protection, and better rendering for images returned by tools.

Upgrade recommendation

Related reading

• DeepSeek Harness complete setup guide
• DeepSeek Harness v0.1.5-rc.2 update
• DeepSeek V4.1 Flash in Harness
• DeepSeek Harness Agent Teams guide

FAQ

Which version does the unpinned command install?

As of September 11, npm latest is 0.1.5-rc.1. npm next points to 0.1.5-rc.2.

Is rc.2 included in this article?

No. Each release has its own page. rc.2 is a smaller follow-up focused on feedback submission and delivered-file presentation.