Opus 5.5 Claude Code Plugin: Guard Hooks, Effort Presets, 10 Skills

A deny rule like Bash(rm *) does not stop /bin/rm -rf build or bash -c 'rm -rf build', and Bash(git push *) misses git -C . push: Claude Code's own permissions page lists those gaps. This plugin closes them with a PreToolUse hook that reads commands the way a shell does, down through $(...), sh -c, xargs and sudo, and adds a Stop hook that sends Claude back to work while the tests it just broke are red.

It is the installable version of the WOWHOW guides Opus 5.5 in Claude Code: Effort, Settings, CLAUDE.md Rules and MCP vs CLI vs Skills: the effort levels, routing settings and six CLAUDE.md rules, packaged with its own marketplace file. Claude Code 2.1.295+, Node 18+.

What You Get

• Guard hook: denies rm -rf in any spelling, git push --force, --no-verify and HUSKY=0, curl | sh and bash <(curl ...), and writes to .env*. Asks before git push, deploys (vercel, kubectl, terraform and 25 more), package publishes, git reset --hard and DROP TABLE. Fails closed if it crashes.
• Stop gate: a turn that changed files cannot end while your test command fails. Two retries, then a warning, so an old failure never traps the session.
• Effort presets: /wowhow-opus-55:effort-high <request> runs one request at that level, low to max; the session level returns on the next message.
• 10 skills, a Haiku search agent, 3 settings.json presets (solo-dev, team-safe, ci-readonly) with model, effort and allow/ask/deny rules, plus a merge script that keeps your existing settings.
• Marketplace files and PUBLISH.md for a private repo and team rollout; README, USAGE-GUIDE, VERIFY.md, licence.

The 10 skills

Tested on 16 October 2026

On Claude Code 2.1.299, an end-to-end script installs the plugin in a throwaway config and drives real claude -p sessions against a local fake API: all ten rule ids fired inside real sessions, rm -rf stayed denied in bypassPermissions mode, the stop gate blocked and then let a fixed turn finish, and effort-high sent output_config.effort: "high". 53 of 53 checks passed. The offline suite runs 218 hook inputs (72 deny, 76 ask, 70 look-alikes that must pass) and passes 105 of 105 on macOS and 102 of 102 on Linux with Node 18, 20 and 22. claude plugin validate --strict passes. No paid API call was made.

How to Set It Up

• Unzip somewhere permanent; run scripts/validate.sh: expect 105 passed.
• claude plugin marketplace add <kit>/plugin, then claude plugin install wowhow-opus-55@wowhow --config test_command="pnpm test --run".
• In a session, ask Claude to run rm -rf build: the call is refused with rule D1.
• Optional: merge settings/solo-dev.json into ~/.claude/settings.json with the bundled merge script.

What is not included

• A sandbox. The guard stops the commands an agent usually writes; a script that deletes files on its own still can.
• Model-tested skills. Scripts, frontmatter and listing were tested; how a model follows each skill was not.
• Windows testing. The hooks are Node in exec form; the scripts are bash.
• API usage. Sessions bill as usual; the plugin adds about 700 tokens of always-on context.

Notes

Cost: check how long a Max plan credit lasts on Opus 5.5 versus Sonnet 5.5 with the free Claude Max API Credits Planner. Spend cap: pair it with Claude Code Spend Guard for a hard daily limit; the hooks run side by side.