Web0926/dsh-llm-verifier

Web0926★ 3TypeScriptLast synced: 2026-08-21

Open on GitHub

DeepSeek Harness plugin that validates and ranks 3/5 independent coding-agent patches before approval-gated apply.

README excerpt

dsh-llm-verifier Generate several coding-agent patches, reject the ones that fail your tests, and let an LLM verifier rank the rest before you decide whether to apply the winner. 简体中文 dsh-llm-verifier is a developer-preview plugin for DeepSeek Harness. It runs 3 or 5 independent coding candidates in detached Git worktrees, validates every candidate against project tests, ranks the passing patches with llm-verifier , and keeps the original checkout unchanged until a separate approval applies the selected patch. [!WARNING] This public version is for trusted repositories only . Candidate generation uses detached Git worktrees and the DeepSeek Harness workspace-write permission mode, but validation commands still execute code from the target repository on the host. Read the security boundaries before using it. Why this exists A single coding-agent run can produce a plausible patch, yet another independent run may find a simpler fix, add better tests, or avoid a subtle regression. Running several candidates manually creates a new problem: comparing them consistently. This plugin turns that process into one approval-gated workflow: 1. Give every candidate the same coding task. 2. Keep ca…

View full README on GitHub →
Agentsagent-evaluationai-agentscoding-agentdeepseek-harnessdeveloper-toolsdsh-plugingit-worktreellm-as-a-verifier

Category