Agent Compatibility

Git hooks work everywhere. Skill concepts load automatically for OpenCode.

What Works Where

Feature	OpenCode	Claude Code	Cursor	Devin Desktop	Any Git Agent
Git hooks	Auto	Auto	Auto	Auto	Auto
Manifest gate	Auto	Auto	Auto	Auto	Auto
SOUL.md / AGENTS.md	Auto	Manual	Manual	Manual	Manual
SKILL.md concepts	Auto	Manual	Manual	Manual	Manual
Stack detection	Auto	Auto	Auto	Auto	Auto
i18n	Auto	N/A	N/A	N/A	N/A

Setup per Agent

OpenCode (Recommended)

No extra setup. Skills load automatically via the skill tool.

Just run init-agents in any project. The agent detects your stack, loads matching skills, and enforces the lifecycle automatically.

bash install.sh
cd your-project
init-agents

Claude Code

Copy rules into CLAUDE.md. Install adapter hooks. Skills load via CLAUDE.md reference.

Creates .claude-plugin/ with hooks. Copy SOUL.md and AGENTS.md key principles into your CLAUDE.md.

bash install.sh --agent claude

Cursor

Add rules to .cursorrules. Install adapter hooks. Skills load via .cursorrules reference.

Creates .cursor-plugin/ with hooks. Append SOUL.md and AGENTS.md to .cursorrules.

bash install.sh --agent cursor

Devin Desktop

Install via Kiro adapter. Hooks load automatically from .kiro/hooks/.

Creates .kiro/hooks/ with pre-flight, commit approval, and edit guard hooks.

bash install.sh --agent kiro

Any Git-Based Agent

Copy the git hooks to your project. Works with any agent that uses git.

The hooks enforce: pre-flight checks, skill consultation, build verification, anti-slop, debug tracking, commit approval with mandatory manifest.

cp scripts/git-hooks/pre-commit .git/hooks/pre-commit
cp scripts/git-hooks/commit-msg .git/hooks/commit-msg
chmod +x .git/hooks/pre-commit .git/hooks/commit-msg
cp scripts/commit-approval.sh scripts/

Using Principles in Your Own System

The core principles are agent-agnostic. Adopt them anywhere:

Principle	How to Use
TOOL_GAP — When tools can't verify, report "ship status unknown."	Never fake success when verification tools are unavailable.
Severity Labels — Classify findings: blocking → important → nit → suggestion → learning → praise.	Apply to any code review or quality check.
Error Path Design — Every tool call, gate, and loop needs a failure path.	Design failure paths at build time, not patched post-incident.
Continuation Over Recap — After context loss, resume. Don't waste tokens re-explaining.	Ask "Where were we?" instead of summarizing everything.
Drift Detection — Check docs vs reality regularly.	Verify stats, versions, features, commands, links match.
Manifest Gate — Require a written summary before commit approval.	Agent must write what changed before generating approval token.