DeepSeek Harness v0.1.7-rc.2: Scheduling & Dynamic Tools

DeepSeek released dsh-v0.1.7-rc.2 on September 24, 2026. After rc.1's architecture migrations, this release adds scheduled tasks, configurable shortcuts, cache-aware dynamic tools, manual continuation after Auto review, and Desktop background execution. It also concentrates fixes for plugin installation, long conversations, long tool output, and Desktop packaging.

As of September 28, 2026, npm latest and next both point to 0.1.7-rc.2. It is therefore not only an optional preview but also the version installed by the current unversioned npm command.

Release and installation

Pin the release:

npx @deepseek-ai/[email protected] web

GitHub still marks rc.2 as a Pre-release even though it is the npm default. Pin versions for important workspaces and finish running tasks before upgrading.

Scheduled tasks reach Web and Desktop

After scheduled tasks are enabled, users can create and manage reminders, inspect run history, and retain tasks across application restarts. Repetition can be as frequent as once per minute.

Scheduled tasks and time context are disabled by default on Web and Desktop. When time context is enabled, the Agent receives a time update every ten minutes by default; existing custom intervals remain intact.

Support in the release does not mean a fresh installation schedules work automatically. Verify that the feature is enabled, tasks survive a restart, the next-run time is correct, and repeated work does not overlap unexpectedly.

Dynamic tools and KV Cache

Rc.2 can add tools dynamically without invalidating KV Cache when the configured model explicitly declares support.

This can avoid reprocessing a long context when the tool set changes, but it is not universal Provider compatibility. When the model or Provider does not declare support, DSH must use a compatible path. Test by adding a tool in the middle of a long Session and comparing cache reuse, request tokens, and time to first token.

Standard mode also reduces the fixed instruction-token overhead paid on every turn. The benefit compounds across long conversations and tool-heavy workloads.

Manual decisions after Auto review

When Auto review rejects an operation, a user may decide whether to continue if the Session permits manual approval. A review failure is reported separately instead of being conflated with an explicit denial.

Auto review is enabled from Plugin Manager, while Inspector is no longer bundled by default and must be installed separately. Workflows that depended on Inspector should verify that the plugin is present after upgrading.

Automatic approval, automatic denial, review failure, and manual override are distinct audit events. Logs and external monitoring should preserve that distinction.

Desktop: closing the window keeps work running

Tasks continue in the background after the Desktop window is closed. Quitting the application warns about the effect on active tasks. Windows update prompts explain automatic restart more clearly, warn earlier when the app is still running, and bring the updated window to the foreground.

Desktop also gains resumable onboarding that explains available credit and lets users choose intended use and work-detail display. Web and Desktop shortcuts can be viewed, searched, customized, and reset, with current bindings shown in the sidebar.

Upgrade testing must distinguish closing a window from quitting the application: the former should keep tasks alive, while the latter must explain which work will stop.

Long-conversation and long-output fixes

Rc.2 fixes persistent send failures in some long conversations and broken characters in oversized tool output that could cause later requests to fail.

Use a real stress case:

1. return a tool result containing multibyte text, code, and a long payload;
2. continue through several tool calls;
3. verify truncation does not split characters;
4. send another message after truncation;
5. reload the page and continue the same Session.

If the problem remains, record the original tool-result size, retained beginning and end, model context limit, and the turn where failure starts.

Plugin installation and settings recovery

The release fixes persistent plugin-install and settings-save failures after a crash or interrupted installation. Plugin details and settings display more reliably, and startup failures provide clearer errors.

Installation identifies the selected official registry or mirror and combines duplicate choices. When GitHub installation fails, users already using a mirror can choose another method. Browsers are also less likely to autofill saved login passwords into API Key fields.

Run a complete plugin regression: install, enable, configure, execute a tool, disable, re-enable, and uninstall. A successful installation alone does not cover runtime state or configuration persistence.

Account models and API Key tasks are separate

Account-funded and API Key tasks use separate model entries. Signing out confirms and stops active account tasks, while expired authentication prompts a new sign-in. Quota warnings match the current task so API Key users are not directed to top up the signed-in account by mistake.

Account models can be edited directly in Settings. Model switching displays progress and reduces the wait after selection. Before debugging balance or availability, determine whether the task uses an account model or an API Key model, then inspect the matching Provider and model ID.

Default and interface changes

Rc.2 also changes several established entry points:

• Coding Tools controls trajectories, code changes, and mode selection for new tasks;
• the separate mode-selection switch is removed, and new tasks use the saved default;
• archive filtering offers hidden, all, and archived-only states;
• incompatible-plugin warnings appear once per startup;
• approval cards follow the interface language and request an explanation in the user's language;
• a fetched URL can be opened directly from a collapsed card.

These changes do not require data migration, but training material, screenshots, and UI automation may need updates.

Upgrade checklist from rc.1

Rc.2 does not repeat the Session redesign. Its focus is turning the 0.1.7 architecture into a system that can sustain daily work: scheduling, background execution, dynamic tools, review recovery, and long conversations all become more operational.

Related reading

• DeepSeek Harness v0.1.7-rc.1 update
• DeepSeek Harness v0.1.7-alpha.2 update
• DeepSeek Harness Plugin Manager
• DSH Sub-agent usage and billing

References

• Official DeepSeek Harness v0.1.7-rc.2 release
• Full changes from v0.1.7-rc.1 to v0.1.7-rc.2
• @deepseek-ai/dsh 0.1.7-rc.2 on npm

FAQ

Is rc.2 the current npm default?

Yes. Both latest and next point to 0.1.7-rc.2, although GitHub still labels the release as a Pre-release.

Why are scheduled tasks missing after installation?

Scheduled tasks and time context are disabled by default on Web and Desktop and must be enabled in the plugin or feature settings.

Does every dynamic tool addition preserve KV Cache?

No. DSH uses the cache-preserving path only when the configured model explicitly declares support.