claudemd-cli 0.17.1 → 0.17.3

package/CHANGELOG.md

@@ -8,6 +8,133 @@ 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).
 
+## [0.17.3] - 2026-05-14
+
+**Patch — fix: CRITICAL — `pre-bash-safety-check.sh` multi-line CMD §8 SAFETY bypass closure. Multi-line bash commands containing `rm -rf $UNSAFE_VAR` on any line other than the first silently passed the hook; the matching multi-line `npx pkg@PIN` case wrongly denied as unpinned. One root cause, two opposite-direction defects.**
+
+### Background
+
+v0.17.2 shipped the bare-`rm -rf $HOME` whitelist closure (Bug 5 / Steam-disaster class). End-to-end dogfood Round 4 — running the *just-shipped* hook against realistic multi-line bash scripts — surfaced that the §8 SAFETY hook's pattern extraction was completely broken for any multi-line CMD where the rm/npx call wasn't on line 1. The fix landed in v0.17.2 reduced bare-`$HOME` to a denied case for *single-line* commands but the hook was systematically failing to fire on multi-line ones.
+
+Minimal repro for the CRITICAL leg (pre-fix):
+
+```
+CMD = "TMP=$(mktemp -d)\nrm -rf $UNSAFE_VAR"
+→ hook output: <empty> (no decision, allow)
+```
+
+The single-line equivalent `rm -rf $UNSAFE_VAR` correctly denied. The multi-line form bypassed §8 SAFETY entirely — the exact spec rule the hook exists to enforce. An agent that issues a heredoc-style bash command with any setup line before the destructive call would slip through.
+
+Reverse leg (false-DENY, also pre-fix):
+
+```
+CMD = "TMP=$(mktemp -d)\nnpx prettier@3.0.0 --check ."
+→ deny: "npx unpinned package: TMP=$(mktemp"
+```
+
+Pinned `prettier@3.0.0` flagged as `TMP=$(mktemp` unpinned. Innocent scripts denied; users would either disable the hook or wrap every npx call in a one-liner — eroding §8 trust.
+
+### Root cause
+
+`pre-bash-safety-check.sh` extracted rm-target and npx-package via per-line sed:
+
+```bash
+rm_tail=$(echo "$SANITIZED_CMD" | sed -E "s/.*${RM_FLAG_REGEX}//" | head -n1)
+npx_tail=$(echo "$SANITIZED_CMD" | sed -E "s/.*${NPX_REGEX}//" | ...)
+```
+
+`sed -E` processes line-by-line. Lines without `rm`/`npx` passed through unchanged; only the rm/npx line had its prefix stripped. Downstream:
+
+- `head -n1` (rm path): always returned line 1 regardless of where the rm sat. If rm was on line ≥2, line 1 had no rm content → `rm_target` empty → deny path never fired. **§8 SAFETY bypass.**
+- `for tok in $npx_tail` (npx path): bash word-splitting iterated tokens from ALL lines. Line 1's first token (typically `TMP=$(mktemp`) became `pkg_token` → flagged as unpinned. **False deny on innocent scripts.**
+
+Sanitize already stripped heredoc bodies / line comments / quoted bodies, so the remaining newlines were between *independent command lines* — safe to flatten.
+
+### What changed
+
+- `[fix CRITICAL]` **`hooks/pre-bash-safety-check.sh`** — new `SANITIZED_CMD_FLAT=$(printf '%s' "$SANITIZED_CMD" | tr '\n' ' ')` computed once after sanitize. Both extraction passes (`rm_tail` and `npx_tail`) now read `SANITIZED_CMD_FLAT` instead of `SANITIZED_CMD`. The `head -n1` on the rm extraction is removed (no longer needed; whole flat string contains all targets after the sed strip).
+
+- `[test]` **`tests/fixtures/bash-safety/corpus.tsv`** +7 multi-line cases (`__NL__` LF marker per corpus convention):
+  - `deny`: multi-line `rm -rf $UNSAFE` (was false-ALLOW — the CRITICAL leg).
+  - `deny`: multi-line bare `rm -rf $HOME` on line 2 (v0.17.2 Bug 5 cross-check — confirms the prior fix is intact when rm is on a later line).
+  - `pass`: multi-line `rm -rf $HOME/cache` on line 2 (whitelist subpath survives multi-line).
+  - `pass`: multi-line `npx prettier@3.0.0` on line 2 (was false-DENY).
+  - `pass`: multi-line `npx ./node_modules/.bin/foo` on line 2 (local path survives).
+  - `deny`: multi-line `npx prettier` unpinned on line 2 (real catch preserved).
+  - `deny`: `rm -rf $UNSAFE` on line 3 of 5 (cross-checks the head -n1 removal — earlier lines no longer hide a deeper rm).
+
+  Corpus 69 → 76. `bash tests/hooks/pre-bash-safety.test.sh` 76/76.
+
+- `[fix LOW]` **`commands/claudemd-rules.md`** — `--verbose` was documented as if it were a script flag (`$ARGS --verbose`), but `scripts/hard-rules-audit.js` rejects it as `Unknown argument`. The command's intent was for `--verbose` to be an agent presentation directive (show full `rules` array in output), not a script flag. A user typing `/claudemd-rules --verbose` would have `CLAUDEMD_RULES_DAYS="--verbose"` set as env, then the script would crash with `--days requires a positive integer (got '--verbose')`. The md now documents a 4-row parsing table (`""` / `90` / `--verbose` / `90 --verbose` → env + agent output) that the LLM follows to split numeric and presentation tokens cleanly.
+
+### Why patch (not minor)
+
+Restores documented §8 SAFETY behavior — the hook was supposed to deny `rm -rf $VAR without validating VAR` per spec §8 (immutable), but multi-line CMDs bypassed it. CHANGELOG `fix:` not `change:` or `feat:`. No new hook surface, no new flag, no new spec rule. The corpus addition is regression-anchor only.
+
+Severity disclosure: the CRITICAL leg means any user on plugin v0.17.2 or earlier had a §8 SAFETY hole for multi-line bash CMDs that included `rm -rf $UNSAFE_VAR` on lines ≥2. We have no telemetry indicating active exploitation, but the spec-coverage gap is real — recommend immediate update.
+
+### Tests
+
+- `bash tests/run-all.sh`: 411 node-test + 2 integration suites pass.
+- `bash tests/hooks/pre-bash-safety.test.sh`: 76/76 (was 69/69 after v0.17.2; +7 multi-line cases this release).
+- Manual end-to-end: 6-scenario rm/npx matrix (multi-line + single-line × bare/subpath/glob/pinned/local/unpinned) verified post-fix; v0.17.2's bare-`$HOME` denials still fire correctly when `$HOME` sits on line ≥2.
+
+### Operator notes
+
+- Update path: plugin marketplace update + `/reload-plugins`. Installed plugin picks up the new hook body via `${CLAUDE_PLUGIN_ROOT}` expansion.
+- **If you have v0.17.2 or earlier, agent-issued multi-line bash with `rm -rf $UNSAFE` was silently allowed**. Audit your `~/.claude/logs/claudemd.jsonl` for the absence of `§8-rm-rf-var` rows on sessions where you expect the hook should have fired.
+- The per-cmd escape hatch `[allow-rm-rf-var]` / `[allow-npx-unpinned]` continues to work in multi-line CMDs (already operates on raw `$CMD`, not sanitized).
+
+## [0.17.2] - 2026-05-14
+
+**Patch — fix: 6-bug end-to-end dogfood pass. §8 SAFETY `rm -rf $VAR` whitelist closure (closes bare-`$HOME` Steam-disaster class); `transcript-vocab-scan` multi-paragraph false-negative; CLI `lint` whitespace-in-positional misclassified as path; CLI `audit` silent-OK on non-JSONL files; manifest `spec_version` drift v6.11.12 → v6.11.16; `update.js` raw Node stack trace on bogus env value.**
+
+### Background
+
+End-to-end agent dogfood across 3 rounds: real user paths through bin CLI, all 17 hooks, install/update/uninstall flows. 411 unit + 2 integration tests as baseline; added 18 regression tests across the 6 fixes. Highest-severity finding: a CRITICAL whitelist gap in `pre-bash-safety-check.sh` that allowed bare `rm -rf $HOME` (and `$PWD`/`$TMPDIR`/`$OLDPWD`) to pass without a `[allow-rm-rf-var]` token. The whitelist was supposed to certify the variable is shell-typed; in practice it also certified the *bare* expansion, which is exactly the Steam-disaster shape (Valve/steam-for-linux#3671: `rm -rf "$STEAM_ROOT/"*` with empty STEAM_ROOT wiped entire home dirs). Spec §8 already forbids `rm -rf $VAR without validating VAR` — the hook was simply not enforcing the spec it was supposed to enforce.
+
+### What changed
+
+- `[fix CRITICAL]` **`hooks/pre-bash-safety-check.sh`** — whitelisted vars (HOME/PWD/OLDPWD/TMPDIR) now require ≥1 non-`/` character in the literal-path residue (the rm-target with all `$VAR` expansions + quotes stripped). `rm -rf $HOME` / `rm -rf "$HOME"` / `rm -rf ${HOME}` / `rm -rf $HOME/` all DENY post-fix. `rm -rf $HOME/cache` / `rm -rf $HOME/*` / `rm -rf "$HOME/sub"` continue to ALLOW — subpath-bounded targets retain the prior behavior. `BASH_SAFETY_INDIRECT_CALL=1` path (`bash -c '...'` unwrapping) inherits the same check. `tests/fixtures/bash-safety/corpus.tsv` +10 rows: 7 new `deny` (bare-var shapes incl. trailing-slash) + 3 new `pass` (glob / quoted / braced subpath) — corpus-driven test goes 62 → 69.
+
+- `[fix HIGH]` **`hooks/transcript-vocab-scan.sh`** — `jq` per-text-block `gsub("[\\r\\n]+"; " ")` before the outer `join(" ")` collapses internal newlines so the whole assistant turn is one scan-friendly line. Pre-fix, an agent turn like `"I significantly improved X.\n\nNext step is Y."` extracted as multi-line text; downstream `tail -n 1` then picked only "Next step is Y." and the §10-V hit in the first paragraph was silently dropped. The hook's docstring comment claimed `join(" ")` made each turn one line, but that only joined CONTENT BLOCKS — embedded `\n` inside a single `.text` block survived. `tests/hooks/transcript-vocab-scan.test.sh` 8 → 10 (+2 multi-paragraph anchors: first-para-only and last-para-only banned word both caught).
+
+- `[fix MED]` **`bin/claudemd-lint.js`** — `lint` positional argument path-shape heuristic now requires no whitespace. Pre-fix, `claudemd-cli lint "Fixed crash in scripts/audit.js:42 (12/12 tests pass)"` exit 2 "file not found" because the heuristic only looked for `/`. Real paths are token-shaped; whitespace-containing positionals are inline sentences with file:line citations. `tests/scripts/lint-cli.test.js` +1 test (`sentence with /file:line citation` → text-scan + banned-vocab variant).
+
+- `[fix MED]` **`bin/claudemd-lint.js`** — `audit` subcommand pre-flight check: a non-empty JSONL file with zero parseable JSON rows exits 2 with `"audit: no parseable JSON rows in <path> (expected JSONL transcript with one JSON object per line)"`. Pre-fix, pointing audit at a non-JSONL file (plain log, CSV, corrupted transcript) silently exited 0 with `"OK: no §10-V hits across 0 assistant turn(s)"` — CI hooks would falsely greenlight a wrong-format input. Same silent-success class as v0.9.14 / v0.9.21 lint fall-through. `parseTranscript`'s documented per-row silent-skip contract is preserved: the guard fires ONLY when 100% of non-empty lines fail to parse. `tests/scripts/lint-cli.test.js` +3 tests (non-JSONL, empty file degenerate-OK, partial-corruption preserves silent-skip).
+
+- `[fix LOW]` **`scripts/update.js`** — `.then().catch()` wrapper translates env-shape errors (unknown `CLAUDEMD_UPDATE_CHOICE` like `YOLO`) into a one-line stderr + exit 1, mirroring the validation-error contract used by audit.js / sparkline.js. Pre-fix, an unknown choice surfaced as a 5-line Node promise-rejection stack trace (`Error: unknown choice: YOLO\n    at update (file:.../update.js:41:11)...`) + exit 1 — same exit code but unreadable for users typo-ing the env var. `tests/scripts/update.test.js` +1 test (assert exit 1 + clean stderr + no `at update (file:...` stack lines).
+
+- `[fix data]` **`spec/hard-rules.json`** — `spec_version` synced `v6.11.12` → `v6.11.16`. Four prior patch releases (v6.11.13–v6.11.16, all compression/wording-only with `§13.2 budget cost: 0`) did not add or remove HARD rules, so the manifest was never bumped. But `/claudemd-rules` and `safety-coverage-audit` both surface this field at the top of their output — users saw "Spec v6.11.12" against a v6.11.16 spec on disk. Manifest is now bumped with every spec H1 change. `tests/scripts/hard-rules-drift.test.js` +1 test (`hard-rules-7`) asserts `manifest.spec_version === spec/CLAUDE.md H1 version` — future H1 bumps that miss the manifest sync will fail CI before reaching users.
+
+### Why patch (not minor)
+
+All 6 changes are `fix:` per CHANGELOG convention — each restores intended/documented behavior:
+
+- §8 SAFETY explicitly forbids `rm -rf $VAR without validating VAR`; bare `$HOME` falls under that rule. The whitelist was an over-permissive shortcut, not the documented intent.
+- §10-V transcript scanning was documented to scan agent assistant text; truncating to the last line of the last turn was a parser implementation bug, not the documented behavior.
+- CLI `lint` whitespace heuristic and `audit` non-JSONL exit 2 both close silent-success / spurious-error variants of the same parser-discipline class fixed in v0.9.14 / v0.9.16 / v0.9.21.
+- `update.js` clean error contract matches the rest of the script suite.
+- `hard-rules.json` `spec_version` sync is a data fix, no behavior change for hook consumers.
+
+No LLM-visible metadata change. No new HARD rules. No new hook surface. No public CLI flag added or removed. Existing `[allow-rm-rf-var]` per-cmd escape hatch unchanged — users with intentional `rm -rf $HOME` patterns (rare) can still bypass with the token.
+
+### Tests
+
+- `bash tests/run-all.sh`: 411 node-test + 2 integration suites pass. Test count unchanged (the 6 new node-test cases offset 0 deletions; corpus + bash hook test counts grew internally: corpus 62 → 69, transcript-vocab-scan 8 → 10).
+- `bash tests/hooks/pre-bash-safety.test.sh`: 69/69 (was 62/62 + 10 corpus rows wired in).
+- `bash tests/hooks/transcript-vocab-scan.test.sh`: 10/10 (was 8/8).
+- Manual end-to-end:
+  - 12 `rm -rf` shape matrix verified: bare `$HOME`/`$PWD`/`$TMPDIR`/`$OLDPWD` + trailing slash all DENY; subpath/glob/quoted/braced all ALLOW.
+  - Real CC transcript audit on 30 recent `~/.claude/projects/-mnt-data-ssd-dev-projects-claudemd/*.jsonl`: 13 flagged with §10-V hits (signal validated against production prose).
+
+### Operator notes
+
+- **Breaking change risk: low.** `rm -rf $HOME` (bare, no subpath) is now denied — but no real workflow runs that intentionally; it's a footgun. Workflows using `rm -rf $HOME/<subpath>` (the normal shape) are unaffected.
+- **Update path:** plugin marketplace update + `/reload-plugins`; the `${CLAUDE_PLUGIN_ROOT}` expansion in hook registration means installed plugin picks up the new hook bodies automatically (`reference_plugin_root_hook_expansion.md`). No manual re-install needed.
+- **If `rm -rf $HOME` is denied for a legitimate use case** (rare — wholesale home wipe in container build, etc.), the existing `[allow-rm-rf-var]` per-cmd escape hatch still works: `rm -rf $HOME [allow-rm-rf-var]`.
+- **`/claudemd-rules` output:** "Spec v6.11.16" header replaces stale "Spec v6.11.12" immediately on first run post-update.
+
 ## [0.17.1] - 2026-05-14
 
 **Patch — fix: `memory-read-check.sh` tag-match phase now sanitizes quoted bodies (and heredoc bodies / line comments) before tag scan; closes the FP class where descriptive text inside `--title "..."` / `-m "..."` / `'release/...'` triggered §11 deny on incidental keyword matches.**

package/bin/claudemd-lint.js

@@ -197,7 +197,13 @@ function lintCmd(rawArgs) {
     // pre-commit-hook ergonomic where `--file` is explicit.
     if (positional.length === 1) {
       const arg = positional[0];
-      const looksLikePath = arg.includes('/') || arg === '.' || arg === '..';
+      // Path-shape heuristic: contains `/`, or is `.`/`..`. Whitespace disqualifies
+      // — `lint "Fixed crash in scripts/audit.js:42 (12/12 tests pass)"` is one
+      // quoted positional whose `/` came from a file:line citation inside a
+      // sentence, not a literal file path. Without this guard, the auto-detect
+      // branch saw `/` and exited 2 with "file not found", forcing users to
+      // either omit citations from inline text or explicitly switch to --stdin.
+      const looksLikePath = (arg.includes('/') || arg === '.' || arg === '..') && !/\s/.test(arg);
       try {
         const st = fs.statSync(arg);
         if (st.isFile()) {
@@ -287,6 +293,25 @@ function auditCmd(rawArgs) {
   }
 
   const jsonl = fs.readFileSync(transcriptPath, 'utf8');
+
+  // Silent-success guard: parseTranscript intentionally skips unparseable rows
+  // (matches transcript-vocab-scan.sh). But if the WHOLE file fails to parse
+  // — user pointed audit at a CSV, plain log, or corrupted JSONL — `turns.length`
+  // is 0 and the CLI happily prints "OK: 0 assistant turn(s)" exit 0. Same
+  // silent-OK family as v0.9.14 / v0.9.21. Pre-flight: a non-empty file that
+  // yields zero parseable JSON rows is malformed, not clean.
+  const nonEmptyLines = jsonl.split('\n').filter(l => l.trim().length > 0);
+  if (nonEmptyLines.length > 0) {
+    let parsedAny = false;
+    for (const l of nonEmptyLines) {
+      try { JSON.parse(l); parsedAny = true; break; } catch { /* keep scanning */ }
+    }
+    if (!parsedAny) {
+      process.stderr.write(`audit: no parseable JSON rows in ${transcriptPath} (expected JSONL transcript with one JSON object per line)\n`);
+      process.exit(2);
+    }
+  }
+
   const turns = parseTranscript(jsonl);
   const patterns = readPatterns();
   const annotated = turns.map(t => ({

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claudemd-cli",
-  "version": "0.17.1",
+  "version": "0.17.3",
   "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": {