Introduction¶
The core philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” In real-world agent or workflow development, the main model sometimes needs to call a more powerful model to obtain advice while strictly adhering to existing permission and approval policies. DSH SuperAdvisor is designed to address this need.
Plugin Overview¶
DSH SuperAdvisor is an independent on-demand advisor tool. It is independently developed and maintained by community members and is an unofficial project. The plugin reuses DSH’s model service, permission service, and approval service. The main model calls the ask_advisor tool, and after the saved approval policy allows it, the advisor returns textual advice, which the main model then validates and acts upon.
Core Features¶
- Independent on-demand advisor tool: Invoked through the
ask_advisortool. - Approval policies: Supports four approval modes:
- Manual approval
- Main-model tag approval
- Independent AI review approval
- Approve all
- Output limits: Supports configurable output limits. New installations inherit the token budget of the selected DSH model service by default and do not impose a local response size limit. Existing installations retain custom limits.
- Security and permissions: The reviewer has no tool permissions, and neither model can grant itself additional permissions. Credential/PII warnings go directly to the fallback process.
Installation and Enablement¶
Because no specific installation command is provided, install it through the DSH UI.
- Find the plugin installation entry in the DSH interface.
- Select DSH SuperAdvisor to install it.
- After installation, go to Settings -> Plugins -> DSH SuperAdvisor.
- Select a configured service and model, or manually enter the model ID, then save the settings.
Typical Usage¶
Approve All¶
In Settings -> Plugins -> DSH SuperAdvisor -> Approval Method, select Approve all (no extra call) and save. This explicitly authorizes all valid advisor requests, including requests marked as requires_human_approval: true, untagged requests, or requests containing local sensitive-content warnings. The result card and audit record identify this policy as always.
- Note: Credential/PII warnings go directly to the fallback process; DSH file/command permissions, host
approval: never, cancellation, and per-send auditing remain in effect.
Independent Model Review¶
In Settings -> Plugins -> DSH SuperAdvisor -> Approval Method, select Independent model review (additional call), then choose the review service. The reviewer is a fixed choice and does not change when the main model changes.
- Workflow: The reviewer has no tools and cannot grant itself additional permissions. If the review passes, the request proceeds directly without waiting for user confirmation; if the review fails, it enters the fallback process. Reviewer calls have independent timeout and output limits (30-second deadline, 1024-token output budget, 8 KiB response binding).
Main-Model Tag¶
In Settings -> Plugins -> DSH SuperAdvisor -> Approval Method, select Main-model tag (no extra call).
- Workflow: The main model includes the
requires_human_approvaltag in the sameask_advisortool call.falseindicates a normal authorized text consultation;trueindicates that manual approval is required or that the consultation is skipped. There is no separate approval reasoning request. The tag reflects the executor’s judgment rather than an independent security review.
Notes¶
- Unofficial project: Independently developed and maintained by community members.
- Version requirements: Version 0.1.11 requires the built source-code version or its package. Version 0.1.8 added verified compatibility for five DSH versions, while older installations remain unchanged.
- Manual mode restrictions: Setting
falsecannot bypass manual mode. - Installation and upgrade: New installations remain in manual mode; upgrades preserve the saved mode.
- Reviewer permissions: The reviewer has no tools, and neither model can grant itself additional permissions.
Conclusion¶
DSH SuperAdvisor provides DSH users with a flexible advisor mechanism that can obtain advice from more powerful models while preserving permissions and security.