Se-harness: AI coding tool unifies instructions across Claude, Gemini, Copilot

Se-harness is a CLI tool that generates unified instruction files for AI coding assistants. Installation is one curl command: curl -fsSL https://raw.githubusercontent.com/manuuuel/seh/main/scripts/install.sh | sh. After that, a user runs seh init --global --agents claude,codex --yes (once per machine) to set up a unified global ruleset at ~/.seh/AGENTS.md with optional symlinks to agent-specific directories (like ~/.claude/CLAUDE.md). Then, per project, seh init --tech typescript --yes detects the project's technology stack and scaffolds AGENTS.md (a short project index) plus .seh/ subdirectories containing project.md (mission and constraints), domain/ modules (architecture and glossary), and stack/ modules (per-technology best practices). Users edit these .seh/ sources and run seh sync to regenerate all agent-specific files in lockstep.

The resolution order is: harness package (if active) → ~/.seh/ (global) → seh bundled core. Global rules apply everywhere; project layers extend them but never override. Generated files (CLAUDE.md, AGENTS.md, GEMINI.md, .github/copilot-instructions.md) must not be hand-edited—seh sync regenerates them from sources. The global AGENTS.md is one self-contained file, led by a forced Craftsmanship principle (small, sharp, elegant code). The project AGENTS.md is the canonical file, a short preamble plus a progressive index linking focused modules.

Harness packages are plain git directories versioned and shared across teams. A harness contains harness.json (package metadata including name, version, and modelTag), a CHANGELOG.md (human-authored reasoning), global/ (global ruleset and agent symlink config), templates/stack/ and templates/project/ (scaffolding templates), projects/ (per-repo overlays), and skills/ (reusable capabilities). Skills can be vendored (committed to the package repo) or referenced (fetched at install time). Each skill has a routing mode: always (invoke every response), when (invoke on matching condition), or optional (agent decides). The routing is stored in harness.json and rendered into a ## Skills section in both the project AGENTS.md and global ~/.seh/AGENTS.md so agents always know which skills to invoke and when.

Supported agents are Claude Code, Codex, Pi, Gemini, OpenCode, GitHub Copilot, and a cross-agent interoperability path (agents). Supported technologies for per-technology modules are JavaScript, TypeScript, Python, Go, C, Rust, and Java; users must select at least one. Key commands include seh init (project setup), seh sync (regenerate from sources), seh check (detect drift), seh link (manage symlinks), seh package init (scaffold a new harness), seh package use (activate a package), seh package install (install artifacts onto the host), seh skills add (add a skill from GitHub), and seh skills update (re-fetch referenced skills). The tool is available for immediate use with a sandboxed demo for experimentation.