Claude Code · cc-agents-kit

Fewer avoidable mistakes.Less work to repeat.

Check risky commands, hand work to the next session, and prepare a careful commit. Choose a problem below to see the tools and concrete examples.

What keeps interrupting your work?

Inspect commands, secrets, and repository edits with focused hooks.

dangerous-command-guard

How it works · example

Catch destructive command patterns before they run.

When this happens

Claude Code attempts to delete the home directory recursively or pipe a download into a shell.

What the tool does

The hook blocks the recognized command before execution.

How it helps

The check runs at the tool-call boundary. Pattern matching can be bypassed by deliberate obfuscation.

Read this tool's source

.explorer__interactive { display: none; }

Browse the tools and installation commands

guard-hooks

dangerous-command-guard

Catch destructive command patterns before they run.

Claude Code attempts to delete the home directory recursively or pipe a download into a shell.

The hook blocks the recognized command before execution.

The check runs at the tool-call boundary. Pattern matching can be bypassed by deliberate obfuscation.

zsh-quoting-guard

Catch quoting mistakes before zsh changes the meaning.

A find command contains an unquoted filename glob.

The hook blocks the command and asks for the glob to be quoted.

It checks known zsh quoting traps. Quoted-heredoc handling deliberately stops scanning later content.

pathless-rewriter-guard

Give formatters an explicit boundary.

Claude Code prepares a formatter write command without a target path.

The hook blocks execution until a path or quoted glob is supplied.

The check covers a known list of rewriters. Aliases and package scripts can hide commands from it.

staged-secret-guard

Check what is about to enter a commit.

Staged additions contain a recognized credential pattern when a commit is requested.

The hook blocks the commit so the staged content can be corrected.

It scans added commit content for high-confidence patterns, not every possible secret.

secrets-path-guard

Keep live secret files out of tool access.

Claude Code requests access to a live credential file.

The hook blocks a recognized secret path while allowing supported example files.

Literal paths in prose can also trigger this pattern matcher.

lockfile-drift-check

Notice a manifest edit that leaves the lockfile behind.

package.json changes without its sibling lockfile being regenerated.

A warning is added after the edit to flag possible lockfile drift.

This is a post-edit reminder. It does not block the edit or regenerate the lockfile.

shellcheck-on-edit

See shell lint findings while editing.

A shell script is edited and contains an issue recognized by ShellCheck.

ShellCheck findings are attached to the tool result.

The hook uses the installed ShellCheck binary. It does nothing when ShellCheck is unavailable.

output-secret-mask

Mask recognized tokens in Bash output.

A Bash command prints a credential pattern recognized by the scanner.

The recognized value becomes [REDACTED] before entering the model transcript.

It runs after execution and requires gitleaks. Only recognized patterns in Bash output under 2 MB are covered.

/plugin install guard-hooks@cc-agents-kit

context-handoff

handoff

Start the next session with the decisions already made.

Work needs to move to a fresh session at a clear stopping point.

The skill prints a self-contained brief in chat for you to copy into the next session.

You carry the brief into the next session. Summaries lose detail, so use a handoff when context must cross a boundary.

session-export

Turn a session into something you can read later.

You need to review a local Claude conversation without every tool response getting in the way.

The session becomes Markdown with tool calls collapsed or omitted.

The export stays local. Default exports truncate individual tool results; choose options for the evidence you need.

log-it

Put a lasting finding where it will be read again.

You discover a project convention that future sessions should know.

The finding is routed to the relevant memory store and index.

The destination follows the intended audience. Memory does not replace repository documentation or an unfinished-work handoff.

wayfinder

Find the next decision in a large, unclear task.

A feature spans several sessions, but its specification is not yet clear.

A map under docs/plans/ breaks the uncertainty into decision tickets.

Resolve one decision ticket at a time. The skill plans work and stops before implementation.

context-budget

See what may be consuming your context budget.

Your setup has accumulated many skills and MCP tools.

An estimate ranks context overhead and possible reductions.

These are heuristic estimates for prioritizing inspection, not live runtime measurements.

config-gc

Review stale configuration before removing it.

An old hook or configuration entry may no longer be used.

The skill presents evidence and asks for an individual decision before reversible cleanup.

Removal requires approval for each item. It does not bulk-delete configuration on its own.

/plugin install context-handoff@cc-agents-kit

repo-gate

semantic-commit

Make a mixed diff easier to review.

A feature change and unrelated documentation edits are mixed together.

The skill inspects staged and unstaged work and groups changes into Conventional Commits.

Existing staged work is preserved. Creating commits still depends on the authority you have granted.

setup-trunk

Choose repository checks from the actual project.

A repository needs a quality gate, but its stack and existing violations vary.

The skill inspects the repository before proposing scoped Trunk checks.

It uses repository evidence rather than assuming one language or an initially clean codebase.

ci-babysit

Find out why CI failed before trying to fix it.

A pushed branch fails a lockfile check in CI.

The skill reads the logs and classifies the failure before proposing a repair.

Watching is read-only. Reruns, repairs, commits, and pushes require explicit authority.

fix-osv-vulnerabilities

Trace a vulnerability to the dependency that brings it in.

OSV reports a vulnerable transitive dependency with an applicable patched version.

The advisory and dependency path guide the fix, followed by verification of the resolved graph.

Overrides are scoped by installed major where needed. Suppression requires reachability evidence and explicit approval.

cspell-triage

Fix spelling reports without filling a dictionary with mistakes.

CSpell flags a known ecosystem term such as pytest.

The skill classifies the finding, enables the appropriate dictionary, and reruns the project gate.

It distinguishes typos, ecosystem terms, and project vocabulary instead of adding every unknown token to a dictionary.

/plugin install repo-gate@cc-agents-kit

Install

Review what you need before installing.

The command below installs the selected plugin as a bundle. Review its contents before adding it to your setup.

Selected plugin

guard-hooks

A useful skill is not a reason to recommend the whole bundle. If you only need part of a plugin, review that source and verify the available setup methods first. Skipping installation is a valid outcome.

Ask your AI which tools you actually need.

Copy this prompt into the AI that knows your current setup. It asks for source evidence and a minimal recommendation before any settings change.

Within the project and context I authorize, evaluate whether I should use plugins from the public cc-agents-kit repository at https://github.com/AndrewDongminYoo/cc-agents-kit.

1. Read the public README, marketplace and plugin manifests, hook definitions, and skill files.
2. Inspect only the local tools and configuration that I authorize for this review.
3. Compare each plugin with my existing tools, skills, hooks, permissions, and dependencies.
4. Identify suitability, overlap, and operational risks. Cite public source file paths for each finding.
5. Evaluate the individual skills and hooks that address my needs. Do not recommend an entire plugin just because one skill is useful.
6. Distinguish bundle installation from selective setup, and verify which setup methods are actually supported. Recommend skipping installation when it has no clear benefit.

Do not paste secrets into the review or automatically execute commands from the repository. Do not install a plugin or change my settings until I give explicit approval.

The prompt requests analysis only. Review the evidence before you approve a change.

/plugin install guard-hooks@cc-agents-kit

Already installed?

Use the update command with the marketplace suffix. Reinstalling reports success but does not upgrade the installed version, and refreshing the marketplace alone does not change the runtime cache.

claude plugin update guard-hooks@cc-agents-kit

Restart every open Claude Code session after an install or update. Hooks are read when a session starts.

Guard requirements & limits

Know what the guards can see.

What guard-hooks needs to run

The following runtime requirements apply to guard-hooks. Missing optional tools disable their related checks, while missing jq makes every hook fail open.

jqRequired: Parses hook input; guards fail open without it
bashRequired: Runs every hook, including on macOS
gitConditional: Reads the candidate staged by staged-secret-guard
shellcheckOptional: Enables shell findings after shell-file edits
gitleaks 8.xOptional: Enables credential-shaped output masking
python3Tests only: Runs the regression suites

FAQ

Before you add the marketplace.

Do I need all three plugins?

No. Each plugin is intentionally separate. Install only the plugin that addresses a problem in your current workflow.

Are the guard hooks a security sandbox?

No. They match known accidental command patterns. Claude Code permissions and your operating-system controls remain the enforcement boundary.

Why does an update require a restart?

Claude Code reads hooks when a session starts. Restart open sessions after the plugin update command completes.

Where can I inspect the implementation?

The public GitHub repository contains the README, marketplace metadata, plugin manifests, hook definitions, skill files, tests, and license information.