bridle-audit 0.1.0

Audit your Claude Code config: dead hooks, duplicate levers, boot-token cost.

bridle-audit

Your Claude Code config grows in silence. Hooks pile up in settings.json — some fire on every single turn — and keep firing long after the script they point at was deleted. The same linter gets wired twice by two different sessions. Skill descriptions are loaded into context at every session start, and nobody counts what that boot costs. Configs only ever grow; nothing prunes them.

bridle-audit is the pruner: one command, zero dependencies, reads your real config and tells you exactly what is dead, doubled, or dormant.

• Dead hooks — commands wired to scripts that no longer exist (fired every turn, for nothing).
• Duplicate levers — the same script wired several times inside one event.
• Dormant cost — how many characters of skill descriptions you pay at every boot.

Install

pipx install bridle-audit

or

pip install bridle-audit

or straight from source:

pipx install git+https://github.com/trinity-organism/bridle-audit

Python >= 3.9. Zero dependencies — standard library only.

Usage

bridle-audit                 # audit $CLAUDE_CONFIG_DIR, falling back to ~/.claude
bridle-audit --config DIR    # audit a specific config directory
bridle-audit --json          # machine-readable output

Exit codes: 0 clean (or no config found), 1 at least one dead hook or duplication, 2 config unreadable — so it drops straight into CI or a shell prompt.

Before / after

Real output (username anonymized). Before — a config that grew for six months:

$ bridle-audit
bridle-audit — installed levers vs the ladder (context < skill < tool < hook < daemon)
hooks source: /Users/you/.claude/settings.json

[per session] SessionStart (1 hook)
  ok    session-banner.sh — /Users/you/.claude/hooks/session-banner.sh

[per turn] PreToolUse (2 hooks)
  ok    format-check.sh — /Users/you/.claude/hooks/format-check.sh
  ok    format-check.sh — /Users/you/.claude/hooks/format-check.sh
  DUP   format-check.sh x2 — same script, 2 voices in one event (one behavior, one lever, one place)

[per turn] Stop (1 hook)
  DEAD  notify.sh — /Users/you/.claude/hooks/notify.sh

dormant cost — 3 skills, ~414 chars of descriptions loaded at every session start

verdict — dead hooks: 1 · duplicate levers: 1 · skills: 3 (~414 chars/boot)
$ echo $?
1

After pruning the dead Stop hook and merging the duplicate into one matcher:

$ bridle-audit
bridle-audit — installed levers vs the ladder (context < skill < tool < hook < daemon)
hooks source: /Users/you/.claude/settings.json

[per session] SessionStart (1 hook)
  ok    session-banner.sh — /Users/you/.claude/hooks/session-banner.sh

[per turn] PreToolUse (1 hook)
  ok    format-check.sh — /Users/you/.claude/hooks/format-check.sh

dormant cost — 3 skills, ~414 chars of descriptions loaded at every session start

verdict — dead hooks: 0 · duplicate levers: 0 · skills: 3 (~414 chars/boot)
$ echo $?
0

And when there is nothing to audit:

$ bridle-audit --config /some/empty/dir
bridle-audit: no settings.json found in /some/empty/dir
Nothing to audit. If your Claude Code config lives elsewhere, point at it with --config DIR or $CLAUDE_CONFIG_DIR.

Hook commands are resolved the way a shell would: quoting (paths with spaces), wrappers (nohup, bash -c, python3, ...), VAR=val prefixes and $HOME/~ are all handled before the target is checked.

Philosophy

A bridle, not a harness rack: one behavior, one lever, one place — and always the cheapest lever that actually steers. A line of context beats a skill, a skill beats a tool, a tool beats a hook, a hook beats a daemon. This tool shows you where your config climbed that ladder without need.

Development

pip install -e ".[dev]"
pytest

The test suite fabricates synthetic configs in temp directories — it never reads your real ~/.claude.

License

MIT