harness-engineering

Harness Engineering — remediation recipes

You are improving this repository's AI harness: the guides, sensors, and guardrails that make agent work reliable. The authoritative reference is https://paladini.github.io/harness-score/ — these recipes summarize it.

Ground rules

• Prefer running npx -y harness-score . --json first (and after changes) so improvements are measured, not assumed.
• Fix in points order: the failed checks worth the most points first, unless the user picks specific ones.
• Write harness files that match THIS repository — read the codebase enough to state real commands, real directories, real conventions. Never ship placeholder text like "".
• Keep context lean: AGENTS.md ≤150 lines; scope guidance where the tool supports it; put procedures in skills, not rules.

Recipes by check family

CTX — Context & Guides

Create AGENTS.md with: what the project is (2 sentences), directory layout, exact build/test/lint commands, non-negotiable conventions, and a "do not touch" list.

For guidance scoped to part of the tree, add a nested CLAUDE.md inside that subdirectory — Claude Code loads it automatically when working in that subtree, the same idea as glob-scoped rules in other tools, just directory-scoped instead of pattern-scoped.

SKL — Skills & Commands

For the team's most repeated procedure, create .claude/skills/<name>/SKILL.md with frontmatter name: and a description: written as a trigger ("Use when the user asks to …", ≥40 chars), body = numbered runbook. Add .claude/commands/<verb>.md for human-triggered workflows like /review.

HKS — Hooks & Guardrails

Register hooks in .claude/settings.json's "hooks" key:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [{ "type": "command", "command": "${CLAUDE_PROJECT_DIR}/.claude/hooks/guard-shell.js" }]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [{ "type": "command", "command": "${CLAUDE_PROJECT_DIR}/.claude/hooks/format-on-edit.js" }]
      }
    ]
  }
}

The gate script reads the tool-call JSON from stdin on PreToolUse, tests the command against a destructive-pattern regex (rm -rf on roots, git push --force, DROP TABLE), and — to block it — prints {"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"deny","permissionDecisionReason":"…"}} and exits 0. The feedback script runs on PostToolUse and formats the edited file, best-effort. Commit both scripts. Full examples: https://paladini.github.io/harness-score/guide/guardrails-and-safety

SNS — Sensors

Wire the ecosystem's standard tools: test runner with a test script, linter config, strict type checking (tsconfig "strict": true / mypy / pyright), formatter. Document the commands in AGENTS.md.

CI — CI Feedback

.github/workflows/ci.yml running lint, typecheck, and tests on push/PR. Optionally add the ratchet: npx -y harness-score --min-level <current>. Pre-commit via husky+lint-staged or pre-commit.

HYG — Hygiene & Safety

.gitignore must cover .env and .env.* (keep !.env.example). Remove real env files; replace inlined keys in .mcp.json with ${ENV_VAR} interpolation. Add LICENSE and commit the lockfile.

After fixing

Re-run the scanner, report old level → new level, and if the user wants it to stick, add the --min-level gate to CI so maturity only ratchets up.