DeepSeek Harness v0.1.2-alpha.1: Agent Teams, PTC & Breaking Changes

Follow-up: DeepSeek released v0.1.2-alpha.2 on August 30, 2026 and published it through npm's alpha channel; latest and next remain on rc.2. Anyone starting a new test of the 0.1.2 development line should read the alpha.2 notes first.

DeepSeek Harness published dsh-v0.1.2-alpha.1 as an official GitHub pre-release on August 27, 2026, with the tag fixed at commit cd5ef81. As of August 30, release notes and source assets are available, but npm's latest and next tags still point to 0.1.1-rc.2; there is no installable @deepseek-ai/[email protected] package.

Compared with v0.1.1-rc.2, this development line contains broad architectural work. The meaningful story is not the commit count but DSH's movement from an early composable preview toward a more complete agent runtime.

Bottom line

v0.1.2-alpha.1 is an official GitHub pre-release, not a stable release or the current npm version. It expands multi-agent operation, remote control, SDKs, session persistence and the Web UI while changing interfaces used by some plugins. General users should stay on npm's rc.2; plugin authors can pin the alpha.1 tag for isolated compatibility testing.

Major changes

Experimental Agent Teams

The source includes Web and CLI profiles for Agent Teams. Persistent teammates, shared tasks and messaging are moving toward a launchable experimental composition. This is not yet a stable default, and lifecycle, recovery and long-running scheduling still need careful testing.

Code Mode becomes PTC

The user-facing code-mode terminology moves to PTC. PTC lets a model generate a program that combines multiple tool calls, reducing model–tool round trips. Older sessions can retain the previous name, so compatibility involves more than a string replacement.

SDK, ACP and model routing

• More independent SDK and ACP application bundles appear.
• TypeScript clients can launch through DSH profiles.
• The GitHub pre-release adds a Windows x64 Python SDK runtime; verify the actual wheel available from PyPI before installing.
• Subagents can receive provider, model, reasoning-effort and maximum-output configuration; Claude Code and Codex subagents can use configured models too.
• ACP completes standard session controls, model settings, MCP, permissions and cancellation.

Together these changes position DSH as a controllable harness runtime, not only a Web chat surface.

GitHub webhooks

New code can create workspace sessions from GitHub events. That opens a path to issue triage, review and automated repair, while making authentication, input validation and event deduplication more important.

Session and client restructuring

Client modules, remote controllers, session migration and per-session projection caches have changed. The intended benefits include cleaner remote APIs, smaller storage and more reliable cold reads. Plugins tied to old client modules, ApiProxy or old DOM structure may break.

Web security and transport

Network access to the Web interface now requires the one-time token in the launch URL, and the gateway sends WebSocket heartbeats. The release notes also enable public WebFetch by default with SSRF protection and no per-request approval for public destinations. These defaults do not replace reverse-proxy authentication, least privilege or third-party code review.

Tokens, logs and privacy controls

The conversation stream can expose exact token usage for each completed answer. The official DeepSeek adapter now sends enabled plugin package names and versions by default, with a switch to disable that behavior. Incremental Session-log upload is a separate opt-in feature and remains disabled by default. Operators should document the data boundary and retention policy before enabling it.

Session and interface improvements

• A message entered during a running turn can be queued for delivery.
• Session records use less disk, and truncated-log repair now names the affected session.
• Image compression and upload continue in the background, while Trajectory can render images from users, assistants and tool results.
• The conversation stream adds turn navigation, width and font controls, and exact token accounting.

Plugin-author checklist

1. Check imports from moved or removed client modules.
2. Find dependencies on ApiProxy or old RPC paths.
3. Audit fixed DOM selectors used for UI injection.
4. Test dependency resolution without assuming global node_modules access.
5. Verify old session migration and persistent state.
6. Adapt to one-time tokens and the new remote controllers.

Should you upgrade now?

Related reading

• DeepSeek Harness v0.1.1-rc.2 update
• How Agent Teams work
• DSH agent presets versus plugins
• DeepSeek Harness security model

FAQ

Is v0.1.2-alpha.1 officially released?

It is officially published as a GitHub pre-release with release notes and source assets. It is not stable and, as of August 30, 2026, it is not published to npm. npx @deepseek-ai/dsh web still resolves to 0.1.1-rc.2.

Why cover a pre-release that is not on npm?

The official release documents the next architectural direction and gives plugin authors an early compatibility signal. An alpha can still change and is not a stability promise.