Claude Code Harness
A portable, TypeScript-powered guardrail and agent harness for Claude Code.
What this gives you
- 13 guardrail rules (R01–R13) that block dangerous operations:
sudo, rm -rf, force-push, curl | bash, .env leakage, and more. Configurable via harness.config.json.
- 3 general-purpose agents:
worker (implement + self-review + verify + commit), reviewer (read-only multi-angle review), scaffolder (docs + state sync). Plus security-auditor and codex-sync helpers.
- 5 verb skills:
/harness-plan, /harness-work, /harness-review, /harness-release, /harness-setup.
- Zero native dependencies in v0.1.0 (pure JS JSON state store) — works on macOS, Linux, and Windows without
npm install native rebuilds.
Installation
claude plugin marketplace add <owner>/cc-triad-relay
claude plugin install harness@cc-triad-relay --scope project
Or use the bundled helper script from any git checkout of this repo:
# Harness only (default — recommended baseline):
bash /path/to/cc-triad-relay/scripts/install-project.sh
# Harness + Codex companion (opt-in, enables codex-sync / coderabbit-mimic):
bash /path/to/cc-triad-relay/scripts/install-project.sh --with-codex
Or for local development:
claude --plugin-dir /path/to/cc-triad-relay/plugins/harness
Then in your project root:
# Inside Claude Code session
/harness-setup init
This creates a harness.config.json tailored to your project.
Optional companion: openai-codex
The Harness ships stack- and LLM-neutral. Two of its agents —
codex-sync and coderabbit-mimic — shell out to the OpenAI
Codex companion plugin to
run synchronous code review / pseudo-CodeRabbit flows. Without Codex
installed, invoking either of those two agents errors immediately
with a clear, grep-able message (they don't degrade gracefully —
they refuse to proceed). Every other agent, command, guardrail, and
hook works identically with or without Codex. Installing Codex is
therefore optional:
| Plugin installed | What works | What errors on invocation |
|---|
harness only (default) | All 13 guardrails, 13 commands (5 verb + 8 workflow), 4 agents (worker / reviewer / scaffolder / security-auditor), all 12 lifecycle hooks | codex-sync fails fast with ERROR: Codex plugin not found and coderabbit-mimic fails with ERROR: codex-companion.mjs not found. — both hard errors that stop the agent before any work starts. Other agents and commands are unaffected. |
harness + codex | Everything above plus Codex-powered synchronous second-opinion review (codex-sync) and local pseudo-CodeRabbit loop (coderabbit-mimic) | — |
install-project.sh --with-codex flips the opt-in; otherwise run
claude plugin install codex@openai-codex --scope project manually.
Verify with harness doctor, which reports:
codex plugin: detected at ~/.claude/plugins/cache/openai-codex/codex
or
codex plugin: not installed (optional — ...)
Re-install is a no-op if already present (idempotent).
Verifying the install — harness doctor
harness doctor
Surfaces:
- Core build presence + mtime.
- Codex companion presence.
harness.config.json location + parse status.- Resolved project security checklist (from
security.projectChecklistPath).
- Resolved plans file + handoff files.
- Project-local skill directory (
.claude/skills/).
- User-level overlays (
~/.claude/{skills,commands,agents}/) — useful
when diagnosing which layer a skill / command came from.
Quick configuration
harness.config.json:
{
"projectName": "my-app",
"language": "en",
"protectedDirectories": ["training-data", "fixtures"],
"protectedEnvVarNames": ["OPENAI_API_KEY", "AWS_SECRET_ACCESS_KEY"],
"workMode": { "bypassRmRf": false, "bypassGitPush": false },
"tampering": { "severity": "approve" }
}
Documentation
Japanese: 日本語ドキュメント
License
MIT. See LICENSE and NOTICE.
Known notes
