DeepSeek Harness (DSH) adopts a plugin ecosystem. A single Harness has a single perspective. Before making critical decisions, before declaring risky work complete, or before writing significant new code, a second model can provide useful independent evidence. The plugin turns consultation into bounded tool calls, rather than copy-paste workarounds. Claude answers under behavioral and output contracts, and DSH receives the results as suggestions for weighing rather than obedience.

Core Features

The plugin provides a toolset centered around external expert consultation:

  • Role contracts: Built-in advisor, reviewer, designer, and custom roles with different objectives and JSON contracts. outputKind, prompt, model, fallback, and effort are configurable. The enabled flag can suspend a role without deleting it.
  • Thinking effort: Native --effort (low / medium / high / xhigh / max), with precedence: per-call argument > role > global default.
  • Model fallback: A one-hop retry for model-level errors (such as unrecognized_model, model-not-found); usedFallback is recorded in run metadata.
  • Agent tools:
    • consult_expert: A single role and a single question, with an optional structured brief.
    • consult_panel: Parallel calls across different roles on the same brief, not majority voting.
    • consult_roles: Live list of roles, including outputKind, role-level model/effort, and global defaults.
  • Auto consult:
    • Composer-seat toggle: The Expert Consult switch on the permission-control line; checked roles add selection conditions and lifecycle reminders to the session.
    • Policy section: The model requests the names of checked roles and their selection conditions, with mode off | remind | hard-remind (default remind).
    • Lifecycle nudges: Reminders triggered before file writes or before task completion.
    • Budget: capPerRole (default 3), limiting the number of real consult_* calls per role per session.
  • Settings workspace: One tab per Harness CLI; hot-apply support is available, so role edits can directly affect the next model step without a restart.
  • Connectivity test: A real end-to-end consultation test (CLI + login + proxy), including turns, duration, cost, and fallback flags.
  • Defense in depth: Workspace-read-only CLI tools; strict MCP isolation; permission locking; bounded output/turns/time; explicit marking of untrusted evidence.
  • Fully bilingual UI: Full Chinese/English support, including built-in role descriptions, backend notes, and validation messages.

Installation

Install the plugin with the official command line:

dsh plugin --profile web add dsh-capability-optimizer

After installation, restart dsh web or your chosen profile.

Prerequisites:
* The claude CLI (npm i -g @anthropic-ai/claude-code) must be on the PATH.
* You must be signed in to a Claude account.

Typical Usage

Invoke the agent tool directly in conversation. For example:

“consult the reviewer on this diff before we call it done”

The Agent will automatically select a role and call consult_expert.

For structured tasks, you can send a Structured brief:

  • objective (Objective)
  • successCriteria (Success criteria)
  • constraints (Constraints)
  • currentAttempt (Current attempt)
  • artifacts (Artifacts)
  • verification (Verification method)
  • unknowns (Unknowns)

The response returned by Claude includes model, turns, duration, cost, and protocol metadata.

Tool Workspace read/write Purpose
consult_expert read* A single role and a single question, optional brief or context
consult_panel read* Parallel consultation across multiple roles
consult_roles read Live role list and configuration

* Workspace is read-only; each call consumes Claude subscription quota.

Applicable Scenarios and Cautions

Applicable scenarios: Scenarios requiring a second-model perspective at critical decision points, code review, or design phases.

Cautions:
* Permission isolation: Agent tools are read-only to the workspace and do not modify files, but they consume Claude subscription quota.
* Instruction passing and mandatory calls: Nudges ensure instructions are passed, rather than forcing tool calls. If the model refuses, it must explain the reason in one line.
* Status display: The status bar uses polling and server-side truncated previews, not token streaming on all hosts.
* Phase limitation: The current phase (Phase 1) supports only the Claude Code CLI. Other backends (codex, zcode, etc.) are planned for later support.

Summary

dsh-capability-optimizer integrates external expert consultation into DSH’s toolchain. Through role contracts and structured briefs, it provides controllable reference answers and helps the Agent perform independent validation before important decisions.

Plugin directory page
GitHub repository