claude-code-harness

Why we included this project

Claude Code Harness is a plugin that replaces the usual 'ask the agent to code' habit with a fixed delivery loop: write the spec, implement only the slice you approve, verify, review independently, and package the evidence. It works with Claude Code, Codex CLI, Cursor, and Grok, and it writes plans to spec.md and Plans.md instead of letting them live in chat, so the agent only implements what you sign off on. The safety layer is the interesting part: a Go engine checks every tool call before it runs, with a non-overridable floor covering billing, network egress, secret reads, and destructive operations, plus tunable guardrails like blocking direct pushes to main. Approvals are collected up front at plan time with expiry and scope limits, and every blocked action is logged. Teams that have been burned by agent drift, skipped tests, or post-merge review will appreciate the enforced gates and the independent review step, and the three HTML decision surfaces let a non-engineer sponsor follow progress without reading code.