The core design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When developing agents that require high-precision reasoning or strict process control, developers often want to force the model to use maximum reasoning depth within a session and uniformly apply specific orchestration strategies. The dsh-ultracode plugin is designed for this purpose: it adds an “Ultra Tier” to DSH. With a single toggle, the plugin forces all model requests to lock onto the maximum reasoning tier declared by the serving model and activates a set of resident orchestration strategies.

Plugin Positioning

  • Name: dsh-ultracode
  • Maintainer: BackMountainBird
  • License: MIT
  • Core Value: A single toggle that forces maximum reasoning depth and activates orchestration strategies.

Core Features

The plugin mainly provides the following capabilities:

  1. Session-level Effort Tier: When enabled, the plugin intercepts requests and forces the reasoning effort level of each Agent request to the maximum value declared by the serving model (by default auto, meaning automatically select the maximum value). If a specific value is configured, that value is used directly.
  2. Orchestration Policies: During an active session, a prompt section called ultra:policy is injected into the system prompt. For top-level sessions, this includes the workflow enabled by default, adversarial validation, context cleansing, and a “narrow wave” retry strategy under provider limits.
  3. UI Interaction: An “ULTRA” chip appears in the Composer toolbar of the web interface. When active, the Composer card displays a rainbow border (static when reduced motion preference is enabled).
  4. Selector Consistency: When switching modes, the plugin re-syncs the native effort tier selector through the public model API, ensuring that the displayed state is consistent with the actual forced request state.
  5. Delegation Awareness: Delegated child agents inherit the “effective state” through the delegation chain. They receive the effort tier lock and the “working policy” (childSection), but do not inherit the top-level session’s “workflow instructions.” Branches (Forks) inherit through a seed prefix and become frozen once generated.

Installation

Before using this plugin, ensure that DeepSeek Harness is installed and satisfies the version requirement (@deepseek-ai/dsh-* >= 0.1.0-rc.6).

Install via npm:

dsh plugin --profile <name> add dsh-ultracode

Typical Usage

After installation, you can switch modes via commands or the Composer interface:

  • Turn on mode:
    Enter /ultra in the conversation. The model will then be forced to use maximum reasoning depth, and orchestration policies will be enabled (for example, auditing dependency points).
  • Turn off mode:
    Enter /ultra off or click the ULTRA chip in Composer again.
  • UI toggle:
    Directly toggle the ULTRA chip in Composer. This operation triggers Picker synchronization.

How It Works

The plugin uses Harness’s event system for state management and injection:

  1. State Management: State is managed by the Harness-owned command/run event. The plugin itself does not define custom session events. This means the state is persistent and requires no extra handling during recovery or replay.
  2. Policy Injection: When the tier is active, ultra:policy or childSection is rendered into the request. Top-level sessions use section (containing orchestration policies), while sub-sessions use childSection (containing only the working policy).
  3. Request Interception: The plugin uses a front-placed agent/request waterfall listener to replace the reasoning tier of the request, ensuring that regardless of how the model selector is set, the actual requested effort is locked.
  4. Inheritance Mechanism: Child agents resolve the parent session via parentSession to obtain the locked value and working policy. Generated branches (Forks) inherit only through seed prefixes and cannot receive subsequent tier changes.

Notes and Boundaries

  • Version Dependency: Must use @deepseek-ai/dsh-* >= 0.1.0-rc.6.
  • Inheritance Rules: Inheritance only adds content. Child agents do not have a user command surface and cannot exit Ultra Tier via a child agent.
  • Display vs. Truth: Picker synchronization is for display consistency only; the pin on the request side is the source of truth. Manually changing the Picker in Ultra mode will reset it on the next toggle, but the current request remains locked.
  • Mode Isolation: Ultra mode does not touch Plan Mode. The two are independently recorded states.
  • Provider Limits: The default policy requires the model to perform “narrow wave” retries under provider concurrency limits, but the actual fan-out capability depends on the provider itself.

Summary

dsh-ultracode provides DeepSeek Harness with a standardized “deep reasoning + orchestration” solution. Through the plugin mechanism, it solves developers’ need to force control of per-session reasoning depth and policy execution without modifying the core framework.