Claude Code Harness is a Shell-based development framework designed to impose structured discipline on autonomous AI-assisted coding workflows. The project implements a repeatable Plan→Work→Review cycle that transforms raw agent work into a bounded, auditable process with explicit gates and source-of-truth documentation. Rather than allowing Claude Code agents to drift between chat-based planning, optional testing, and late-stage review, Harness enforces a five-stage operating loop: specification and planning, implementation of approved slices only, independent verification, separate review, and evidence packaging for release.
The core workflow revolves around five verb-based skills that keep the surface minimal and intentional. The /harness-plan command generates spec.md and Plans.md files containing scope, acceptance criteria, dependencies, unknowns, and stop conditions. Users approve or correct this generated contract before execution proceeds, shifting responsibility from hand-writing plans to validating AI-generated ones. The /harness-work command executes approved tasks with test-driven development when required and runs verification to keep work within plan boundaries. The /harness-review command separates review from implementation, treating major findings as blockers. The /harness-release command packages only verified evidence after implementation and review are complete. A /harness-setup command initializes project guidance, command surfaces, hooks, and baseline checks.
Harness supports multiple host tools across different support tiers. Claude Code v2.1+ is the primary supported path through the Claude plugin marketplace. Codex CLI and Codex app are marked internal-compatible, with direct plugin smoke testing tracked separately. OpenCode, Cursor, and GitHub Copilot CLI have internal-compatible or candidate status with explicit runtime parity boundaries. The project includes migration tooling via bin/harness doctor --migration-report, which inventories stale plugin caches, duplicate skills, old symlinks, and memory state without deleting data, allowing existing users to assess impact before upgrading.
According to GitGenius activity tracking, the repository shows median issue and PR response latency of 36.4 hours with a mean of 77.6 hours across 51 tracked items. Primary contributor Chachamaru127 accounts for 104 tracked events, with secondary contributors beomjinkim2000 and SHA888 showing minimal activity. The most active issue labels are bug, codex, and enhancement. The project overlaps with contributors from excalidraw/excalidraw, alirezarezvani/claude-skills, and anthropics/claude-code, indicating cross-pollination within the Claude ecosystem.
The harness treats generated spec and plan files as source-of-truth, preventing silent invention of unobserved data by marking unknowns explicitly. Non-trivial planning can trigger team validation mode, checking spec and plan alignment, memory reuse, product fit, security fit, and practical feasibility through team or sub-agent perspectives. Advanced capabilities include Breezing for larger task lists, Codex companion review through schema-backed second opinions, OpenCode bootstrap for compatible surfaces, and optional harness-mem for cross-session project-scoped memory.
The project is classified across multiple domains including Claude AI, code execution, prompt testing, automation, API integration, output validation, AI harness, workflow management, developer tools, and natural language processing. Documentation includes tool-first onboarding, per-tool install routes, migration checks, plain-language glossaries, skill trigger gates, capability matrices, and evidence pack specifications. The codebase requires Claude Code v2.1+, project repository write access, and optional harness-mem for memory features, but does not require Node.js due to a Go-native guardrail engine.
