claudemd-cli 0.23.21 → 0.23.23

package/CHANGELOG.md

@@ -8,6 +8,37 @@ All notable changes to the `claudemd` plugin. This changelog tracks plugin artif
 - **Canonical spec version source**: `spec/CLAUDE.md` top-line title (`# AI-CODING-SPEC vX.Y.Z — Core`) + `spec/CLAUDE-changelog.md` top `##` entry.
 - **Plugin semver vs spec semver** are independent: plugin patch (0.2.0 → 0.2.1) may ship when spec is unchanged (this release); plugin minor (0.1.9 → 0.2.0) ships when spec minor updates (v0.2.0 shipped spec v6.10.0).
 
+## [0.23.23] - 2026-07-03
+
+**Patch — a self-audit found a doctor↔audit telemetry-parity gap, then an adversarial §8 false-negative sweep (hardened by a code review) closed four "execute scripts of unknown origin" bypass classes in the pre-bash-safety hook.** Spec content unchanged (stays v6.14.1).
+
+### Fixed — `doctor`: rule-usage counted test-session/probe rows that `audit.js` filters
+
+- `scripts/doctor.js`: the `rule-usage` health check derived its deny/bypass counts (and the §0.1 demote verdict downstream of them) from the raw `readHits` output, while the sibling `audit.js` strips manual-probe / sentinel-session rows via `excludeTestSessions`. On the same 30-day window doctor over-reported — `§7-ship-baseline` deny 17 vs audit's 9, `§8-rm-rf-var` 121 vs 101. Both rule-usage consumers (the `groupBySection` count + the demote-candidate token-breakdown loop) now filter identically; the `fail-open` check stays on raw hits by design (its rows are `session_id:null`). Doctor now matches audit.js across every section. Tests: `doctor.test.js` +2 (sentinel-exclusion count + token-breakdown).
+
+### Fixed — §8: four "execute scripts of unknown origin" false negatives in `pre-bash-safety`
+
+A 2026-07-03 adversarial audit (feed dangerous commands to the detector, observe allow/deny) plus a `superpowers` code review found the hook enforced only the literal forms of §8's unknown-origin / NPX rules. Four classes now covered — each adds denials only, verified against a false-positive control set:
+
+- **Fetch-execute runner family**: the NPX gate matched only literal `npx`; `pnpm dlx` / `yarn dlx` / `bunx` / `npm exec` of an unpinned unknown package bypassed it (identical fetch-execute — `npx` is a shortcut for `npm exec`). `NPX_REGEX` extended to the family, reusing the existing pinned/local/lockfile resolution (already reads pnpm-lock.yaml / yarn.lock). `npm install` / `pnpm install` / `yarn add` stay excluded (the regex requires the `exec` / `dlx` subcommand).
+- **`curl … | sh`** (new Pattern 3): no detector existed for piping or `<()`-substituting a network fetch into a shell. Fires when `curl`/`wget` in command position feeds `sh`/`bash`/`zsh`/`dash`/`ksh`/`ash` (optionally via `sudo`), matched per pipeline segment on the sanitized command. Local/literal sources (`cat x.sh | sh`), non-shell sinks (`| jq`), download-only (`curl -o`), and prose-in-quotes stay allowed. New `§8-curl-sh` telemetry section + `[allow-curl-sh]` bypass token.
+- **rm behind wrappers**: `sudo` / `doas` / `timeout N` / `nice -nN` / `stdbuf` / `ionice` / `chrt` before `rm -rf $VAR` bypassed the segment-start `rm` check. The wrapper-strip loop now handles them (arg-less + flag-bearing, consuming a wrapper's options and numeric/duration args, stopping at the first command word). Stripping only removes prefixes, so a non-rm command or a safe rm (literal path / `$HOME` subpath) never false-denies.
+- **Code review follow-up** (folded in): `sudo -E/-i/-H/-n rm` (sudo is flag-bearing, not arg-less), `npm exec`, and brace-group `{ curl … | sh; }` were caught adversarially after the initial fixes. Documented residuals: option-with-argument wrapper forms (`sudo -u svc rm`, `timeout -s KILL 5 rm`), `eval "$(curl)"`, `find -delete`.
+
+Hook tests +50 (206 in `pre-bash-safety.test.sh`); full suite 523 pass. `docs/RULE-HITS-SCHEMA.md` synced (new `§8-curl-sh` section + `allow-curl-sh` token). Findings + deferred residuals recorded in `tasks/s8-false-negative-audit-2026-07-03.md` and `tasks/project-audit-findings-2026-07-03.md`.
+
+## [0.23.22] - 2026-07-02
+
+**Patch — follow-up to v0.23.21 (code review): sync the two remaining dedup-key reference docs + harden the `extra` key against future key-order variance.** Spec content unchanged (stays v6.14.1).
+
+### Fixed — two dedup-key reference docs still taught the pre-v0.23.21 4-field model
+
+- v0.23.21 extended the `uniqueInvocations` dedup key to `(ts, hook, session_id, tool_use_id, event, extra)` and synced the parse-fn comment + `commands/claudemd-audit.md`, but two other canonical descriptions were left on the old 4-field model — and `scripts/audit.js:70` cross-references the first as authoritative, routing a maintainer straight into the pre-fix "byte-identical multi-emit looks like a bug" misread. Synced both: `hooks/lib/rule-hits.sh:21-25` (the `TOOL_USE_ID` param doc — ships with the plugin) and `docs/RULE-HITS-SCHEMA.md:16` (the `tool_use_id` field row — tracked via the `docs/*` + `!docs/<file>` gitignore allowlist) now carry the 6-field key + the multi-emit "confirm against the source command" caveat. Found by a `superpowers:requesting-code-review` pass on v0.23.21.
+
+### Changed — canonicalize `extra` before the dedup key (defensive hardening)
+
+- `scripts/lib/rule-hits-parse.js`: the key serialized `extra` with bare `JSON.stringify`, so double-fire detection silently depended on stable key order. Unreachable for `_real` today (every hook emits `extra` from fixed templates / single-key objects; the only multi-key extra — mem-audit `{missing,drift}` — carries a null `tool_use_id` → lands in `_legacy`), but a future multi-key extra on a `tool_use_id`-bearing hook built from an unordered source (`declare -A`) would let a genuine double-fire evade. New `stableExtraKey()` sorts top-level keys before serializing — behavior-preserving on current data (live-log `pre-bash-safety` `duplicate_rows_real` stays 34). Two direct unit tests added in `tests/scripts/rule-hits-parse.test.js` (reordered-key double-fire still collides → 1 unique / 1 `_real`; distinct-extra multi-emit stays separate → 2 / 0); `uniqueInvocations` previously had no unit-level coverage, only via `audit()`. Also noted: the v0.23.21 key extension incidentally narrows `duplicate_rows_legacy` (distinct event/extra rows at a null `tool_use_id` now count as unique — more correct; `_legacy` is documented non-gating noise). Renamed the v0.9.34 audit test title to state what it verifies (byte-identical collapse) instead of over-claiming the full key. Full suite 519 → 521 pass.
+
 ## [0.23.21] - 2026-07-02
 
 **Patch — audit `duplicate_rows_real` over-counted multi-emit hooks as registration double-fires; found by a 30-day /claudemd-audit self-review.** Spec content unchanged (stays v6.14.1).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claudemd-cli",
-  "version": "0.23.21",
+  "version": "0.23.23",
   "description": "Standalone CLI for §10-V banned-vocab + transcript scanning. Companion to the claudemd Claude Code plugin (github.com/sdsrss/claudemd) for use in git pre-commit hooks, GitHub Actions, and other agents.",
   "type": "module",
   "bin": {