dsh-capability-optimizer

External-expert consultation for DeepSeek Harness: the agent headlessly invokes the Claude Code CLI through explicit advisor, reviewer, designer, or custom role contracts, then weighs the structured replies as reference answers.

English | 中文

Why this exists

A single harness has one perspective. At a consequential decision, before declaring risky work done, or before significant new code, a second model can supply useful independent evidence. That is a hypothesis to measure, not a guaranteed quality gain: persona research does not show that merely calling a model an expert reliably improves accuracy. This plugin makes consultation a bounded tool call instead of a copy-paste detour. Claude answers under a behavioral and output contract, and DSH receives the result as advice to weigh, not obey.

Phase 1 speaks only to the Claude Code CLI. The settings schema (v2, one workspace per harness), the UI tab catalog, and the runner seam are already multi-backend: codex, zcode, kimi-code, pi, opencode, and omp each land later as a runner behind the same three tools.

Features

🎭 Role contracts	Built-in advisor / reviewer / designer with different objectives and JSON contracts, or your own (outputKind, prompt, model, fallback, effort). enabled parks a role without deleting it
🧠 Thinking effort	Native --effort (low / medium / high / xhigh / max) at three levels: per-call argument > role > global default
🔄 Model fallback	One-hop retry on model-level errors (unrecognized_model, model-not-found, …) with usedFallback recorded in run metadata
🤖 Agent tools	consult_expert (one role, one question) · consult_panel (up to N roles in parallel, one wall-clock wait) · consult_roles (live roster)
🎛 Auto consult	Composer-seat toggle (permissions row) picks roles per session; a policy section rides the system prompt and lifecycle nudges fire at write/finish anchors, budgeted per role per session
🖥 Settings workspace	One tab per harness CLI; saves hot-apply — role edits reach the agent's next model step without a dsh restart
🔬 Connectivity test	One real consultation end-to-end (CLI + login + proxy) with turns, duration, cost, and fallback marker
🛡 Defense in depth	Read-only CLI tools, strict MCP isolation when supported, permission pinning, typed schemas, bounded output/turns/time, and explicit untrusted-evidence labels
🌐 Fully bilingual	Every UI string — including built-in role descriptions, reserved-backend notes, and validation messages — follows the UI language (zh/en); agent tooling keeps stable English identifiers

Install

# from npm (recommended)
dsh plugin --profile web add dsh-capability-optimizer

# or straight from the GitHub repo
dsh plugin --profile web add github:hezhongtang/dsh-capability-optimizer

Restart dsh web (or your profile of choice). Works in any profile — web, tui, headless — because the tools are host-side agent tools.

Requirements: the claude CLI (npm i -g @anthropic-ai/claude-code) on PATH, logged in.

Usage

Ask your agent:

"consult the reviewer on this diff before we call it done"

…