DeepSeek Harness rc.8 Is Here: 7 Major Upgrades You Need to Know

DeepSeek Harness v0.1.0-rc.8 is a meaningful developer-preview update that adds a stronger multimodal input path, optional Codex and Claude Code subagents, persistent PowerShell sessions on Windows, faster large-session operations, and better OpenAI-compatible gateway support—but its incompatible SQLite storage change makes a backup essential before upgrading.

Imagine opening a long-running agent workspace on your home server: one session is reviewing a repository, another is delegating a test task to Codex, and a third needs a screenshot attached to a plan. Earlier DeepSeek Harness builds already provided the plugin-based foundation, but several pieces of that workflow still felt separate. rc.8 connects more of them. Images can travel further through the agent interface, external coding agents can be installed as profile-level subagents, Windows terminals can retain state, and large histories are less expensive to fork. The tradeoff is equally practical: this is still a fast-moving developer preview, and storage compatibility cannot be treated casually.

This guide explains what changed in rc.8, which improvements matter in real self-hosted workflows, what “multimodal support” does and does not mean, how to approach a ZimaOS update safely, and whether you should upgrade now or wait.

DeepSeek Harness rc.8 at a glance

DeepSeek Harness v0.1.0-rc.8 was published as a pre-release on August 19, 2026. It remains part of the project’s developer preview, so compatibility-breaking changes are expected. The release is identified by the tag dsh-v0.1.0-rc.8 and commit 141eb6f. The complete change list is available in the official rc.8 release notes.

If you are new to DSH, it helps to understand that Standard, PTC/Code, Minimal, and Creator are different runtime compositions rather than performance tiers. Our guide to DeepSeek Harness modes explains which tool and workflow surface each mode provides.

What changed in DeepSeek Harness rc.8?

Multimodal input now reaches more of the agent workflow

rc.8 expands the path for image-aware work. The DeepSeek model adapter can be configured for native image requests, commands such as /goal and /plan can receive combined text and image input, and the @ menu can reference files and previous sessions. The release also addresses failures caused by oversized images or too much accumulated image payload in a session.

That combination matters more than a simple “upload image” button. A screenshot can become part of a goal definition, a file can be pulled into the current conversation without manually copying its contents, and an existing session can be referenced as working context. For interface debugging, document review, diagram interpretation, and visual planning, those are all useful building blocks.

There is an important capability boundary: rc.8 adds adapter and interface support, but the selected model endpoint must still accept images. The official model configuration guide explains that a custom provider’s declared input types are a claim about the endpoint, not an automatic capability test. If a text-only endpoint is marked as image-capable, the provider can still reject the request. See the official provider configuration documentation.

For text-only DeepSeek routes, a vision bridge remains useful. Our roundup of DeepSeek Harness plugins worth installing covers ModLens and specialized visual toolkits alongside browser, workflow, sidebar, and migration extensions.

Codex and Claude Code become optional Profile Bundles

One of rc.8’s most consequential changes is the ability to install Codex and Claude Code as separate product-subagent bundles. DSH remains the coordinating harness, while each external coding agent is added only to the profile that needs it.

dsh plugin --profile <profile-name> add @deepseek-ai/dsh-subagent-codex
dsh plugin --profile <profile-name> add @deepseek-ai/dsh-subagent-claude-code

The bundles can be installed independently or together. Codex also supports non-interactive permission modes and multiple named instances, allowing a profile to expose separate Codex providers with distinct roles or policies. Installing a bundle does not automatically authenticate the product, select a model, or grant it unlimited permissions; those remain separate deployment decisions.

This turns DSH into more than a single-agent interface. A parent agent can keep the plan and project context while delegating a bounded implementation or review task to a specialized coding agent. The Job Panel and rc.8’s faster subagent report delivery make that delegation easier to supervise.

For exact package commands and lifecycle behavior, use the official DSH CLI reference. Because the native product bundles include platform-specific dependencies, verify the first real delegation after installation instead of treating a completed package command as the only health check.

Windows gets persistent PowerShell sessions

On Windows, rc.8 adds persistent PowerShell support to the PTY terminal and enables it by default in the Minimal preset. A persistent shell can preserve the current directory, environment changes, and command-session state between calls. That is a better match for iterative development than launching a fresh PowerShell process for every command.

The improvement is especially relevant to Windows repositories that rely on PowerShell scripts, Visual Studio tooling, or stateful setup steps. It also narrows the workflow gap between persistent Bash on Linux and native Windows operation.

Custom OpenAI-compatible gateways receive better compatibility controls

Self-hosted users often connect DSH to a company gateway, a local inference server, or a provider that implements only part of the OpenAI request format. rc.8 fixes cases where request-shape differences prevented calls and where reasoning content could be missing from returned results.

This does not make every “OpenAI-compatible” endpoint identical. Gateways can disagree about developer roles, output-token field names, reasoning formats, image parts, and model discovery. DSH exposes compatibility settings for these differences, but the route still needs to match the actual server. Test normal chat, tool calling, reasoning, and image input separately rather than assuming one successful prompt verifies the entire provider.

Parallel research and large sessions get less friction

rc.8 allows web_search to run concurrent queries, improves reportDelivery so a finished subagent can wake its parent task promptly, and reduces the time required to fork sessions with large histories. These changes align with a common research pattern: fan out into several narrow searches, let independent workers investigate separate questions, and then merge their findings into one evidence-backed answer.

The update also improves sidebar search focus, workflow-panel controls, model selection, narrow composer layouts, feedback presentation, local-file retry behavior, and home-directory display. Local dsh web launches now open the browser automatically, while installation downloads have been reduced.

Why the rc.8 SQLite change matters before you upgrade

The most important operational sentence in the rc.8 release notes is not a new UI feature: the SQLite backend now uses a more efficient but incompatible storage structure. The project’s current pre-release storage policy accepts only the schema version implemented by that build rather than promising automatic migration between every developer-preview format.

In practical terms, an application binary and its data should be treated as one rollback unit. If you update the application but keep no copy of the earlier data, reinstalling rc.7 may not restore the previous working state. Backing up only the container image is not enough.

Back up these areas before moving to rc.8

• The persistent DSH application-data directory.
• Profile configuration and installed plugin manifests.
• Credentials or credential references needed to restore providers.
• SQLite databases and retained session data.
• Any workspace content that the agent can modify.

Keep the backup outside the directory that the update process can replace. If the deployment matters, test restoration as well as creation: a backup that has never been restored is still an assumption.

For a home-server deployment, fast persistent storage also improves repository search, package installation, builds, and long-session access. A ZimaBoard 2 can start with a compact always-on DSH environment and expand through SATA or PCIe storage. A ZimaCube 2 is a better fit when the same machine also carries larger repositories, databases, CI containers, backups, and other services.

Should you upgrade from rc.7 to rc.8?

The answer depends on which part of DSH you actually use.

If the main attraction is the broader DSH ecosystem rather than rc.8 itself, start by identifying one missing capability. The DSH plugin guide maps common needs to vision, browser, workflow, TUI, sidebar, and migration options. If you are comparing DSH with a different agent stack, our guide to Hermes plugins for DeepSeek agents provides additional context.

How to update DeepSeek Harness rc.8 on ZimaOS safely

DeepSeek Harness is available through the ZimaOS App Store. The exact update control shown in ZimaOS depends on the app version currently installed and the package version available in the catalog, so confirm that rc.8 is offered before replacing the existing deployment.

1. Record the current version. Confirm whether the existing installation is rc.6, rc.7, or another package build.
2. Stop active agent tasks. Do not update while a session, build, tool call, or subagent is still writing state.
3. Back up persistent application data. Include the database, profiles, configuration, plugins, and session data.
4. Record provider and workspace settings. Keep secrets protected, but make sure you know which credential references and mounts must be restored.
5. Apply the ZimaOS App Store update. Wait for the container to be recreated and become healthy before opening the Web UI.
6. Confirm the running version. Do not assume the new image was activated merely because the download completed.
7. Run a short validation session. Test model output, file access, a harmless command, session persistence, and Web UI reconnection.
8. Test optional features separately. Verify image input against the selected model endpoint and run one real Codex or Claude Code delegation if those bundles are installed.
9. Keep the backup until normal use is verified. Include long-session opening and forking if those are important to your workflow.

For users interested in running a local DeepSeek model separately from the Harness, the ZimaOS DeepSeek R1 deployment guide covers a different layer of the stack. DSH is the agent harness; the model endpoint is the inference layer. Running one does not automatically run the other.

DeepSeek Harness rc.8 known limitations and checks

rc.8 arrived only days after the initial developer-preview builds, so new reports are appearing quickly. The following checks are more useful than treating every open discussion as a universal defect:

• Image support: verify the exact provider and model. Adapter support alone does not make a text-only endpoint multimodal.
• Model catalogs: confirm that the model you expect actually appears after upgrading; catalog dependencies may lag a provider’s newest model list.
• Product bundles: after installing Codex or Claude Code support, perform a real initialization test to catch incomplete native dependency downloads.
• Remote Web UI: use HTTPS for non-local access. Do not expose a workspace-capable agent interface over an unprotected public origin.
• Long sessions: monitor compaction, history loading, and forks instead of assuming a process marked “running” is still making progress.
• Plugin permissions: review profiles after an update. A plugin can affect files, shells, browser sessions, credentials, or network services depending on its role.

DeepSeek’s own documentation still warns that the project is iterating rapidly and will contain compatibility-breaking changes. The official DeepSeek Harness repository should remain the primary source for version status and breaking changes.

DeepSeek Harness rc.8 FAQ

Is DeepSeek Harness rc.8 a stable release?

No. It is marked as a pre-release and the project remains in developer preview. It is suitable for evaluation and carefully isolated self-hosted workflows, but compatibility should not be assumed across release candidates.

Does rc.8 let every DeepSeek model understand images?

No. rc.8 improves image handling and lets the DeepSeek adapter declare native image input, but the chosen endpoint and model must actually accept image parts. Test the provider/model combination before relying on it.

Can DeepSeek Harness rc.8 use Codex and Claude Code?

Yes. They can be installed as separate optional Profile Bundles. Each product still needs its own valid authentication, settings, permissions, and first-run verification.

Do I need to back up before upgrading from rc.7?

Yes. rc.8 changes the SQLite data structure in an incompatible way. Preserve the old application data together with the old application version if rollback matters.

Does rc.8 require a GPU?

No GPU is required when DSH calls a remote model API. GPU and VRAM requirements belong to an optional local inference server and depend on the selected model, quantization, and context size.

What Node.js version does rc.8 use?

The current upstream package declares Node.js ^22.19.0 || >=24.0.0. A packaged ZimaOS application normally carries its own container runtime, so users should follow the app package rather than replacing the host’s Node.js installation.

Can I update DeepSeek Harness through ZimaOS?

Use the ZimaOS App Store update when rc.8 is available for the packaged application. Back up persistent data first, then verify the running version and complete a short functional test after the container is recreated.

Final verdict: rc.8 is worth testing, not rushing

DeepSeek Harness rc.8 moves the project toward a more connected agent workspace. Multimodal inputs can participate in goals and plans, Codex and Claude Code can join as optional subagents, Windows gets a stateful PowerShell path, compatible gateways are easier to configure, and long-session operations receive useful performance work.

The right upgrade strategy is controlled rather than automatic. Back up the full application state, test rc.8 in a separate profile or recoverable deployment, verify the specific model and subagent features you need, and retain a matched application-and-data snapshot for rollback. For ZimaOS users, that approach provides access to the newest DSH capabilities without treating developer-preview data as disposable.

Trademark note: “DeepSeek Harness” is used here to identify and describe compatibility with the DeepSeek project. This article does not imply official endorsement, partnership, or authorization. See the DeepSeek Harness Brand Asset Usage Guidelines.