~ / rigs / kevo-tz / harness-kit

kevo-tz/harness-kit

Multi-model Claude Code harness: Sonnet works, Opus advises, Haiku verifies, Fable skeptic gates every commit

↗ GitHub ★ 0 no license updated 8d ago personal setup Claude Code
field: AI / ML
stack: Python
share on X
ARCHETYPE
Minimalist
Lean instructions, few tools, tiny context tax. Lets the model think.
CONTEXT TAX · EVERY TURN
~458 tokens
Featherweight · median rig: 2.3k · breakdown
GUARDRAILS
0/5
No committed guardrails · details
Quality 3/100 Guardrails v2

Copy this rig

# review before running: this installs third-party code
$ npx degit kevo-tz/harness-kit/.claude ./rig-harness-kit  # inspect, then merge into .claude/

MCP servers are added to Claude Code at local scope; env vars are shown as YOUR_… placeholders — we never store values. Files are fetched with degit into a separate folder so you can review before merging.

$ curl -fsSL --create-dirs -o .claude/agents/skeptic.md https://raw.githubusercontent.com/kevo-tz/harness-kit/main/.claude/agents/skeptic.md
$ curl -fsSL --create-dirs -o .claude/agents/verifier.md https://raw.githubusercontent.com/kevo-tz/harness-kit/main/.claude/agents/verifier.md

This rig commits no guardrails. Here is the community baseline instead — the deny/ask rules most often found across all 16.2k rigs:

{
  "permissions": {
    "deny": [
      "Read(./.env)",
      "Read(./.env.*)",
      "Read(**/.env)",
      "Read(~/.ssh/**)",
      "Bash(git push --force:*)",
      "Read(**/*.pem)",
      "Read(.env)",
      "Read(~/.aws/**)",
      "Read(**/*.key)",
      "Read(**/.env.*)",
      "Bash(git push --force*)",
      "Bash(rm -rf:*)",
      "Bash(rm -rf *)",
      "Bash(sudo *)",
      "Bash(sudo:*)",
      "Bash(rm -rf /)",
      "Bash(git push -f:*)",
      "Read(.env.*)",
      "Bash(git reset --hard*)",
      "Bash(rm -rf /*)"
    ],
    "ask": [
      "Bash(git push:*)",
      "Bash(git push *)",
      "Bash(gh pr merge *)",
      "Bash(git commit:*)",
      "Bash(rm *)",
      "Bash(gh pr merge:*)",
      "Bash(gh pr create *)",
      "Bash(git reset --hard:*)",
      "Bash(npm publish:*)",
      "Bash(git rebase *)"
    ]
  }
}

Subagents (2)

skeptic
model: opus
Adversarial reviewer. MUST be run on the full pending change before any git commit. Give it the goal of the change and the verifier output. Tries to break the change and answers approved or rejected.
verifier
model: haiku
Cheap checker. Use PROACTIVELY to run tests, linters and builds and report the real output. Never claim tests pass without delegating to this agent.

Similar rigs

copied ✓