Changelog

All notable changes to this project are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.3.0] — unreleased

New labeled-password rule: detects credentials pasted in prose as password: / passwd: / passphrase: values. Lower length floor (8) than the generic catch-all with a modest entropy gate, so short labeled secrets are caught while printenv PWD= dumps and common placeholders (your-password, <redacted>, …) are filtered out.
Rules can now carry an optional deny regex to drop placeholder values that clear the entropy gate.

Report: findings with an unexpected severity are bucketed under low instead of throwing when building the verbose per-severity listing.
FORCE_COLOR parsing now follows Node's documented rule — '', 1, 2, 3, true enable colour; any other value (including 0/false) disables it.

broom sweep — redact in place, equivalent to clean --apply (both work).
Interactive menu when broom is run with no arguments on a terminal, and a confirmation prompt before sweep/clean --apply (skipped with --yes and in non-interactive/CI use). Built on node:readline/promises — still zero runtime dependencies.
--verbose flag: the default report is now a compact summary; --verbose lists every occurrence with rule and line number.
"Did you mean …?" suggestion for unknown commands.

Report redesigned. Instead of one line per finding (thousands of lines on a busy history), the default groups by AI client and shows vulnerable chats and the count of distinct secrets, deduped by value — the same credential echoed many times counts once (the raw occurrence count is shown alongside).
Colour output now respects NO_COLOR / FORCE_COLOR and TTY detection.

Cursor adapter out-of-memory crash. Discovery selected key, value across Cursor's entire ItemTable / cursorDiskKV, loading every blob (embeddings, file caches — often gigabytes) into memory at once and crashing with "heap out of memory" on large databases. Discovery now selects only keys; each row's value is fetched lazily when it's actually scanned.

First public release.

broom scan / broom clean — find and redact secrets in AI coding-assistant transcripts. Dry-run by default; --apply writes, backing up every touched file to ~/.broom/backups/<timestamp>/ first.
Source adapters for Claude Code (~/.claude/**/*.jsonl), Codex (~/.codex/**), and Cursor (state.vscdb SQLite via node:sqlite).
broom sources — list discovered transcript files.
Curated, gitleaks-style ruleset plus an entropy-gated generic catch-all; non-reversible, idempotent «BROOM:<rule>:<sha8>» placeholders.
Allowlist with built-in documentation placeholders and a user file at ~/.broom/allowlist.txt (--allowlist / --no-allowlist).
broom install — installs a Claude Code skill + Stop hook that sweeps transcripts automatically after each turn.
broom proxy — a local redacting proxy that strips secrets from outgoing requests and model responses (text and tool-call arguments) for both streaming and non-streaming Anthropic and OpenAI APIs.
broom proxy --install [--daemon] / --uninstall — wire base-URL env vars into shell profiles and register a login-persistent daemon (launchd on macOS, systemctl --user on Linux); uninstall removes both the daemon and the env vars.
--json, --no-fail, --extra <file> flags; --source filtering.

Requires Node ≥ 22.13 (node:sqlite is unflagged as of 22.13.0).
Zero runtime dependencies.
Extended-thinking blocks are passed through the proxy unredacted to preserve their signatures — see the README Limitations section.