diff --git a/.githooks/pre-commit b/.githooks/pre-commit
index 27e96f6..4da4819 100755
--- a/.githooks/pre-commit
+++ b/.githooks/pre-commit
@@ -18,3 +18,14 @@ if ! bash "$repo_root/utils/sync.sh" --check; then
   echo "            run \`utils/sync.sh\` and \`git add\` the result, then retry." >&2
   exit 1
 fi
+
+# A PATH-injected mock must never delegate with `command <tool>` — `command`
+# bypasses functions and aliases but not PATH lookup, so the mock re-executes
+# itself without bound. This is the cheap grep only; the full shellcheck pass
+# (~8s) stays in CI and `lint-shell.sh` rather than on every commit.
+if grep -rn --include='*.sh' -E '\)\s*command (git|gh|tea) ' "$repo_root/tests/" 2>/dev/null; then
+  echo ""
+  echo "pre-commit: mock delegates via \`command <tool>\` — this recurses through PATH." >&2
+  echo "            use write_mock_git from tests/lib/mock-git.sh instead." >&2
+  exit 1
+fi
diff --git a/.github/scripts/check-ci-parity.sh b/.github/scripts/check-ci-parity.sh
new file mode 100755
index 0000000..6b61fcd
--- /dev/null
+++ b/.github/scripts/check-ci-parity.sh
@@ -0,0 +1,59 @@
+#!/usr/bin/env bash
+# check-ci-parity.sh — assert validate-all.sh runs everything CI runs.
+#
+# validate-all.sh is documented as the single local equivalent of CI. That claim
+# is only true if the two lists agree, and a comment asking humans to keep them
+# in step is exactly the kind of thing that silently rots — CLAUDE.md claimed
+# "four checks" while CI ran five, and the missing one had no local equivalent
+# at all. This turns that convention into a check.
+#
+# Compares the check commands in .github/workflows/ci.yml against the
+# run_check lines in validate-all.sh. Setup steps (installing jq, rumdl) are
+# not checks and are ignored.
+#
+# Usage: bash .github/scripts/check-ci-parity.sh
+
+set -uo pipefail
+
+cd "$(dirname "${BASH_SOURCE[0]}")/../.."
+
+WORKFLOW=".github/workflows/ci.yml"
+VALIDATE_ALL=".github/scripts/validate-all.sh"
+
+for f in "$WORKFLOW" "$VALIDATE_ALL"; do
+  [[ -f "$f" ]] || {
+    echo "missing $f" >&2
+    exit 1
+  }
+done
+
+# Check commands are single-line `run:` steps invoking a script or linter.
+# Anything else (multi-line `run: |` install blocks) is setup, not a check.
+mapfile -t ci_cmds < <(
+  grep -oE '^\s+run: (bash [^ ]+\.sh|rumdl check \.)$' "$WORKFLOW" |
+    sed -E 's/^\s+run: //' | sort -u
+)
+
+if [[ ${#ci_cmds[@]} -eq 0 ]]; then
+  echo "no check commands found in $WORKFLOW — parsing likely broke" >&2
+  exit 1
+fi
+
+missing=0
+for cmd in "${ci_cmds[@]}"; do
+  if grep -qF -- "$cmd" "$VALIDATE_ALL"; then
+    printf '  ok      %s\n' "$cmd"
+  else
+    printf '  MISSING %s\n' "$cmd"
+    missing=1
+  fi
+done
+
+echo ""
+if [[ $missing -eq 0 ]]; then
+  echo "CI parity: validate-all.sh covers all ${#ci_cmds[@]} CI check(s)."
+else
+  echo "CI parity FAILED — $VALIDATE_ALL does not run every CI check." >&2
+  echo "Add the missing command(s) to $VALIDATE_ALL so a local run matches CI." >&2
+  exit 1
+fi
diff --git a/.github/scripts/validate-all.sh b/.github/scripts/validate-all.sh
index fd1be49..506b717 100755
--- a/.github/scripts/validate-all.sh
+++ b/.github/scripts/validate-all.sh
@@ -35,6 +35,7 @@ run_check() {
   echo ""
 }
 
+run_check "ci parity" bash .github/scripts/check-ci-parity.sh
 run_check "plugin tests" bash test.sh
 run_check "plugin structure" bash .github/scripts/validate-plugins.sh
 run_check "frontmatter" bash .github/scripts/validate-frontmatter.sh
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 44b5c24..521664e 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -16,6 +16,8 @@ jobs:
         run: sudo apt-get update && sudo apt-get install -y jq
       - name: Validate plugins
         run: bash .github/scripts/validate-plugins.sh
+      - name: Check CI/validate-all parity
+        run: bash .github/scripts/check-ci-parity.sh
 
   test:
     name: Plugin tests
@@ -28,7 +30,7 @@ jobs:
           sudo sh -c 'curl -sSL https://github.com/mvdan/sh/releases/download/v3.10.0/shfmt_v3.10.0_linux_amd64 > /usr/local/bin/shfmt'
           sudo chmod +x /usr/local/bin/shfmt
       - name: Run tests
-        run: bash tests/test.sh
+        run: bash test.sh
 
   validate-frontmatter:
     name: Validate frontmatter
diff --git a/CLAUDE.md b/CLAUDE.md
index bdffb17..1d3ad56 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -229,9 +229,18 @@ bash .github/scripts/validate-plugins.sh       # plugin structure
 bash .github/scripts/validate-frontmatter.sh   # command/skill frontmatter
 rumdl check .                                  # markdown linting
 bash .github/scripts/lint-shell.sh             # shellcheck + mock-recursion guard
+bash .github/scripts/check-ci-parity.sh        # CI and validate-all agree
 ```
 
-Adding a CI job means adding it to `validate-all.sh` too; the script says so.
+Adding a CI check means adding it to `validate-all.sh` too — and
+`check-ci-parity.sh` fails the build if you forget, so the two lists cannot
+drift apart again.
+
+**Run `bash test.sh`, not `bash tests/test.sh`.** The top-level script is the
+aggregator: it wraps the plugin suite and already parses pytest and cargo
+output, so a second suite root added there is picked up automatically. CI used
+to call the inner script directly, which would have silently skipped any such
+addition.
 
 Run a single test suite directly: