Stats
Links
Categories
cc-bridle
A "bridle" for Claude Code — guides without blocking.
Claude Code is powerful, but it can silently run rm -rf or use pip install in a uv project. cc-bridle warns before dangerous operations, prevents tech stack misuse, and gives you visibility into tool usage.
Designed to coexist with claude-code-harness — both plugins share hooks without conflict.
Before / After
| Scenario | Without cc-bridle | With cc-bridle |
|---|---|---|
| Destructive command |
|
|
| Wrong package manager |
|
|
Features
Guards (block execution)
| Feature | Behavior |
|---|---|
| 🛡️ env-guard | Auto-detects your project's tech stack (uv / bun / pnpm / poetry and 14 stacks total) and blocks the wrong package manager. Suggests the correct command via redirect-rules.json with capture group support for argument passthrough (e.g. pip install foo → uv add foo). |
| 🔐 git-guard | Applies 4 configurable guards to Claude Code's git operations: force push prevention, direct push to main/master prevention, secret file staging prevention (.env, .key, .pem, etc.), and Conventional Commits enforcement. Each guard is individually togglable in config.json. Your manual git commands are completely unaffected (no .git/hooks/ changes). |
Monitors (warn & record)
| Feature | Behavior |
|---|---|
| 🔴 danger-label | Detects destructive commands (rm -rf, git push --force, DROP TABLE, DROP SCHEMA, mkfs, etc.) via regex and adds 🔴 CRITICAL / 🟡 WARNING labels to Claude's permission dialog. Never blocks (always returns action: 'allow'). Patterns are extensible via global danger-dict.json + per-project danger-dict-extra.json. |
| 📊 stats | Records every tool invocation (tool name, exit code, success/failure, duration) to stats.jsonl. Tracks Skill tool calls by skill name and Agent tool calls by subagent type. /cc-bridle stats shows usage frequency, failure rates, and optimization hints. |
| ⏱️ hook-timer | Measures each hook's execution time and warns to stderr when it exceeds the threshold (default 500ms, configurable in config.json). History is stored in hook-timer.jsonl; use /cc-bridle hook-timer to identify bottlenecks. |
Diagnostics & analysis
| Feature | Behavior |
|---|---|
| 🔍 skill-scan | Recursively scans SKILL.md files for shell injection, eval/exec, credential access (🔴 HIGH), file writes, external network calls (🟡 MEDIUM), and file deletion (🔵 LOW). Exits 1 on HIGH/MEDIUM findings, acting as a CI gate. LOW-only findings exit 0. |
| 🔌 hooks-conflict | Compares hooks.json files across 5 events (PreToolUse / PostToolUse / SessionStart / Setup / UserPromptSubmit) and detects same-matcher collisions. On conflict, proposes 2 resolution strategies: ① priority merge (recommended) ② matcher rename. |
| 🚀 preflight | Parses CLAUDE.md and Plans.md to detect 6 permission types needed for the session (Bash execution, file writes, network, external services, package publishing, tmp file operations). Simultaneously matches all danger-dict patterns against planned commands and reports them as 🔴/🟡. Works with session-probe to auto-generate warmup commands (echo, mktemp, uv --version, etc.) at session start, pre-triggering permission dialogs so they don't interrupt your workflow. |
Plan management
| Feature | Behavior |
|---|---|
| 📋 plan-split | Splits Plans.md ## Phase N: sections into individual docs/plans/phase-N.md files. Skips existing files (never overwrites). Generated files include creation date, purpose, dependencies, task tables with cc:TODO markers, and acceptance criteria checklists. |
| 📝 plan-review | Walks through each task in a phase file one by one, presenting 3 options: OK / Edit / Explain intent. Edits are saved to the phase file in place. Acceptance criteria are also reviewed individually and can be modified on the spot. |
Install
# Plugin manager (recommended)
claude plugin install cc-bridle
# Or manually
git clone https://github.com/your-org/cc-bridle ~/.claude/plugins/cc-bridle
bash ~/.claude/plugins/cc-bridle/scripts/init.sh
For manual install, add "plugins": ["cc-bridle"] to ~/.claude/settings.json.