Introduction¶
In the DeepSeek Harness (DSH) ecosystem, agent development often requires additional judgment logic. The dsh-advisor plugin ports the rpiv-advisor subsystem and provides a zero-parameter advisor() tool for each agent scope. It forwards the complete derived session history and tool manifest to a separately configured advisor model, uses the original rpiv system prompt, and returns plan, correction, or stop signals to help the main model make more accurate decisions.
What It Is¶
- Name: dsh-advisor
- Description: Standalone DSH plugin package for porting the
rpiv-advisorsubsystem. - Maintainer: NexusAgentX
- Category: Model Reasoning
- License: MIT
Core Capabilities¶
- Zero-parameter tool: Registered as a tool in the agent scope without requiring additional parameters to be written for each agent.
- On-demand invocation: The main model decides when to call
advisor(); there is no turn observer or automatic injection mechanism. - History forwarding: Forwards the complete derived session history and tool manifest to the advisor model.
- Review mechanism: Interacts with the advisor model using the original rpiv system prompt.
- Result handling: The advisor model’s output (plan, correction, or stop signal) is returned as a normal tool result.
- Configuration flexibility: Supports three configuration levels: plugin line configuration, dsh Web settings page, and
/advisorcommand.
Installation and Configuration¶
Installation¶
Register and install it using the official command:
dsh plugin --profile web add dsh-advisor # <name> = your profile name
Configuration Notes¶
The plugin is disabled by default. When enabled, provider and model must be configured; otherwise, the tool will not be installed and the reason for being disabled will be reported. Configuration items include:
enabled: Master switch, defaultfalse.provider: Required. Provider routing for the advisor model.model: Required. ID of the advisor model.reasoningEffort: Reasoning effort, defaults to empty (auto), supportsminimal,low,medium,high,xhigh,max,off.systemPrompt: Overrides the original rpiv prompt; defaults to empty (uses built-in).guidance: Executor prompt override, includingpromptSnippetandpromptGuidelines.disabledForModels: Executor blacklist that prevents specific models from seeing theadvisortool.
Configuration is combined across three levels, with later levels overriding earlier ones:
1. Plugin line configuration: Located in cordis.patch.yml, serving as the base layer.
2. dsh Web settings page: Modify it in the Advisor card on the “Plugin Configuration” page; it writes to the advisor namespace and overrides the plugin line configuration.
3. /advisor command: Takes effect only in the current session and does not modify persistent configuration.
Usage¶
Session Control¶
After enabling the plugin, you can control behavior in the session from the command line:
/advisor: Toggle the advisor tool in the current session./advisor on: Enable it in the current session./advisor off: Disable it in the current session./advisor status: View model, effort, tool visibility, and disable reasons.
Tool Invocation¶
When the main model determines that stronger judgment is needed, it actively calls the advisor() tool. The advisor model’s response is free-form text containing a plan, correction, or stop signal. After reading the result, the main model continues execution and reflects key recommendations in the next visible reply.
Notes¶
- Disabled by default: It must be enabled manually after installation.
- Required configuration: When enabled, both
providerandmodelmust be configured, otherwise the tool is unavailable. - Configuration validation: Unknown configuration keys are rejected; legacy configuration keys
immuneTurnsandmaxDeltaMessagesare ignored. - Dependencies: Depends on peerDependencies; ensure the environment satisfies dependency version requirements.
- Control: The main model retains control and decides when to call
advisor(). - Ecosystem note: This plugin is listed in the community catalog (skillhub.cn) and has no affiliation with official DeepSeek.