ai-md-stack

Base, standard configuration I run with Claude Code: a single canonical AGENTS.md shared with OpenCode and AbacusAI CLI, plus a small enforcement layer of Claude Code hooks.

Full writeup: My AI setup: Claude, Abacus, and OpenCode.

• Claude Code
• OpenCode
• AbacusAI CLI

ai-md-stack/
├── AGENTS.md                        canonical instruction file
├── CLAUDE.md -> AGENTS.md           symlink, same content
├── settings.example.json            hook registration for Claude Code
├── hooks/                           Claude Code enforcement layer
│   ├── _common.sh
│   ├── mutation-guard.sh
│   ├── commit-guard.sh
│   ├── session-context.sh
│   └── redact-secrets.sh
└── examples/opencode-with-global/   recovers the global AGENTS.md in OpenCode
    ├── opencode.json
    ├── AGENTS.md
    └── README.md

CLAUDE.md is a relative symlink to AGENTS.md (ln -s AGENTS.md CLAUDE.md), so both Claude Code and OpenCode read the same file. Claude Code only recognizes CLAUDE.md natively; OpenCode and AbacusAI CLI read AGENTS.md directly.

This part is shared: AbacusAI CLI only loads AGENTS.md from the project workspace, not from a user-level global path, so its copy has to live in each project root rather than symlinked from a single global location.

The file itself is based on the structure from andrej-karpathy-skills, evolved further through running Claude Code across several different environments.

OpenCode doesn't concatenate local and global instructions the way Claude Code does: once a project has its own AGENTS.md, the global one is ignored entirely. See examples/opencode-with-global/ for the opencode.json workaround.

AGENTS.md/CLAUDE.md is context, not applied configuration: the model tries to follow it, but nothing guarantees it remembers a rule in a long session. This is the enforcement layer for the rules that must hold regardless.

Claude Code only. Doesn't port to OpenCode's plugin system (tool.execute.before/after) or to AbacusAI CLI, which has no documented hook mechanism.

• _common.sh: shared is_command_invoked() helper, so a guard fires on terraform plan but not on jq --arg c 'terraform plan'.
• mutation-guard.sh (PreToolUse): blocks write verbs on aws/gh/kubectl/terraform, with an explicit ALLOW_MUTATION=1 escape hatch.
• commit-guard.sh (PreToolUse): validates git commit subjects against the Conventional Commits rules already documented in AGENTS.md.
• session-context.sh (SessionStart): injects current branch, dirty file count, and last commit into context.
• redact-secrets.sh (PostToolUse): redacts leaked secrets (GitHub/OpenAI/Anthropic tokens, AWS access keys, PEM blocks) from command output, strips known noise, and truncates large output, via updatedToolOutput.

All shellcheck-clean and bash -n validated.

Also running RTK (Rust Token Killer), a CLI proxy that rewrites shell commands before execution to cut token usage. Registered as its own PreToolUse hook, separate from the ones in hooks/.

mkdir -p ~/.claude ~/.config/opencode
cp AGENTS.md ~/.claude/AGENTS.md
ln -s AGENTS.md ~/.claude/CLAUDE.md
ln -s ~/.claude/AGENTS.md ~/.config/opencode/AGENTS.md

cp hooks/*.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/*.sh

Then merge the hooks block from settings.example.json into ~/.claude/settings.json.

Skills follow the same source-of-truth pattern as AGENTS.md. All three tools accept the same format, a directory named after the skill containing a SKILL.md with name/description frontmatter, but each one scans its own user-scoped directory:

• Claude Code: ~/.claude/skills/
• OpenCode: ~/.config/opencode/skills/
• AbacusAI: ~/.abacusai/skills/

Two ways to keep one canonical copy:

1. Version the skill in your regular GIT path (e.g. ~/GIT/<skill-repo> with SKILL.md at the root) and symlink it into all three CLIs.
2. Keep the skill inside one CLI's directory and symlink it into the other two.

Either way, one edit lands in the canonical copy and every CLI reads the same file. The frontmatter name must match the directory name.

Example, humanize-br (github.com/Esl1h/humanize-br), a skill versioned at ~/GIT/humanize-br and symlinked into all three:

ln -s ~/GIT/humanize-br ~/.claude/skills/humanize-br
ln -s ~/GIT/humanize-br ~/.config/opencode/skills/humanize-br
ln -s ~/GIT/humanize-br ~/.abacusai/skills/humanize-br

v-repo-audit uses the other arrangement: the canonical copy lives in OpenCode's directory (~/.config/opencode/skills/v-repo-audit) with symlinks from Claude Code and AbacusAI. Note the difference from AGENTS.md: AbacusAI has no user-global instructions file, but it does read user-global skills, so skills are the reliable way to carry personal behavior into AbacusAI sessions.