CHANGELOG.md

@@ -214,6 +214,27 @@ on the way to cutting its first release, and this file starts there.
   pty (util-linux `script`, skipped where it is absent) and asserting the
   MESSAGE rather than the exit code, which the bug also produced.
 
+- **`users apply` now tells "revoke everyone" apart from "I truncated the
+  file"** (#65) — a users file naming zero users is a valid instruction to
+  revoke every operator on the box, and it is indistinguishable from a file a
+  stray `>` produced. The per-user warnings apply already emitted arrive after
+  the decision and scale wrong: twenty operators is twenty lines of scrollback,
+  so the signal was loudest exactly where it read as noise. The `/etc/rig/users`
+  ledger draws the line apply needs — an empty file against an empty ledger is
+  an unambiguous no-op; against a populated one it closes every named door — so
+  only the second case now stops, states how many operators are about to be
+  revoked, and requires explicit consent: `--yes`, `RIG_YES=1` (the
+  installer-family variable `rig uninstall` already reads), or a `y` on a TTY.
+  Without a terminal and without consent it exits 2, in `uninstall_confirm`'s
+  words, rather than assume a yes it cannot ask for or hang on a prompt nothing
+  can answer. A **confirmation**, not `rig bootstrap`'s flat refusal of the same
+  file (#57/#59): bootstrap asserts who lives on a box, apply converges, and
+  converging to zero stays a legitimate de-provisioning. Ledger entries already
+  marked `revoked` don't count toward the number, so a second identical run
+  stays the silent no-op. Mass revocation below the empty-file bright line (a
+  file dropping 19 of 20) is deliberately still ungated — that needs a threshold
+  someone has to justify, and #65 stays open for it.
+
 ## 0.2.0 — 2026-07-19
 
 ### Added

README.md

@@ -876,6 +876,35 @@ refused as a username: this file names operators; root's fate is root-door polic
 `--file -` reads stdin. A bad file exits 2 with **every** error listed at
 once, before anything changes — one fix cycle, not one round-trip per line.
 
+**A file naming zero users is confirmed, not refused (#65).** "Revoke
+everyone" and "I truncated the file" are the same instruction in this format,
+and a stray `>` writes the second one. apply cannot read intent — but it can
+read the `/etc/rig/users` ledger, which draws the only line worth drawing:
+against an empty ledger an empty file is an unambiguous no-op, and against a
+populated one it closes every named door on the box. So that second case, and
+only it, stops and asks — naming how many operators are about to go:
+
+```sh
+rig users apply --file ./users            # empty file + managed operators -> asks
+rig users apply --file ./users --yes      # ...or say yes up front
+RIG_YES=1 rig users apply --file ./users  # same, for automation
+```
+
+**Without a terminal and without consent it exits 2** rather than assume a
+yes it cannot ask for — the contract `rig uninstall` already uses, and the
+same `RIG_YES` the installer family reads. Consent that cannot be obtained is
+not consent, and a prompt nothing can answer must not hang either.
+
+This is a *confirmation*, deliberately unlike `rig bootstrap`'s flat refusal
+of the same file (see *Bootstrap*): bootstrap **asserts** who lives on a box,
+so an empty answer contradicts itself, while apply **converges** — and
+converging to zero is a complete, legitimate de-provisioning that has to keep
+working. Already-revoked ledger entries don't count toward the number, so a
+second identical run of an emptied file stays the silent no-op convergence
+promises. Dropping *most* users — nineteen of twenty — is not yet gated;
+that needs a threshold, where "the file is empty" is a bright line that needs
+none.
+
 **`@root` — seed keys from the door you came in through (#17).** A key field
 of exactly `@root` means "this user's `authorized_keys` becomes root's
 CURRENT `/root/.ssh/authorized_keys`". The point is lockout-avoidance: you

commands/users-apply.sh

@@ -18,9 +18,15 @@ die()  { printf 'rig-users: ERROR: %s\n' "$1" >&2; exit "${2:-1}"; }
 
 usage() {
   cat <<'EOF'
-usage: rig users apply --file <path>
+usage: rig users apply --file <path> [--yes]
 
   --file <path>   users file (required; '-' reads it from stdin)
+  --yes           consent, up front, to the one prompt this command can ask:
+                  a file naming ZERO users against a box that still has
+                  managed operators revokes all of them. RIG_YES=1 says the
+                  same thing (the installer-family convention). Without
+                  either, that case asks on a TTY and REFUSES (exit 2)
+                  without one — it never assumes consent it cannot get.
 
 The file is line-based and bash-parseable on purpose — a rig box has no YAML
 parser and no jq, and gets neither for this. Whitespace-separated: user,
@@ -72,11 +78,17 @@ EOF
 
 # --- args (validated before the root check, so errors are testable) ---------
 FILE=""
+# RIG_YES is the installer-family consent contract (bin/rig's uninstall_confirm
+# reads the same variable): how automation says yes where there is no terminal
+# to say it on. Set here so --yes and the env var are one flag with two doors.
+ASSUME_YES=0
+[ -n "${RIG_YES:-}" ] && ASSUME_YES=1
 while [ $# -gt 0 ]; do
   case "$1" in
     --file)
       [ $# -ge 2 ] || die "--file needs a value" 2
       FILE="$2"; shift 2 ;;
+    --yes) ASSUME_YES=1; shift ;;
     -h|--help) usage; exit 0 ;;
     *) die "unknown flag: $1" 2 ;;
   esac
@@ -501,6 +513,62 @@ done
 # revoked, data kept, convergence never destroys.
 LEDGER=/etc/rig/users
 REVOKED=()
+
+# --- the empty-file gate (#65) -----------------------------------------------
+# "Revoke everyone" and "I truncated the file" are the same instruction in this
+# file format, and apply cannot read intent. What it CAN read is the ledger, and
+# that draws the only line worth drawing: a file naming zero users against an
+# empty ledger is an unambiguous no-op, while the same file against a populated
+# one closes every named door on the box. Only the second is dangerous.
+#
+# So this is a CONFIRMATION, not a refusal — deliberately unlike bootstrap's
+# flat die on the same file (#57/#59). bootstrap ASSERTS who lives on a box, so
+# an empty answer there is a self-contradiction; apply CONVERGES, and converging
+# to zero is a complete, legitimate de-provisioning that must keep working. The
+# difference between the two commands is the whole point, and it survives here.
+#
+# The per-user warnings below are not this gate and cannot replace it: they
+# arrive after the decision, one line per operator, so the signal is loudest
+# exactly where it reads as scrollback rather than as a question.
+#
+# Count FIRST, then speak: the message states a real number, and counting is
+# what makes the already-converged case silent. An entry the ledger already
+# marks `revoked`, or one whose account no longer exists, is not at risk — this
+# run would not change it — so a second identical run of an emptied file stays
+# the clean no-op convergence promises, with no prompt to answer twice.
+#
+# SCOPE: the bright line only. Whether a file that drops 19 of 20 operators
+# deserves the same gate is the issue's open question — it needs a threshold
+# someone has to justify, where "the file is empty" needs nothing. Left open.
+if [ "${#USERS[@]}" -eq 0 ] && [ -r "$LEDGER" ] && [ "$ASSUME_YES" -eq 0 ]; then
+  AT_RISK=0
+  while read -r prev pstate _; do
+    [ -n "$prev" ] || continue
+    [ "${pstate:-active}" != "revoked" ] || continue
+    id -u "$prev" >/dev/null 2>&1 || continue
+    AT_RISK=$((AT_RISK + 1))
+  done < "$LEDGER"
+  if [ "$AT_RISK" -gt 0 ]; then
+    warn "this users file names ZERO users, and this box still manages $AT_RISK operator(s): applying it revokes every one of them — accounts expired, authorized_keys renamed, rig groups stripped. If the file was meant to be empty this is de-provisioning; if it was truncated by accident, stop here."
+    # No terminal means no consent, and a question nobody can answer must not
+    # be assumed into a yes or left to hang. Same shape and same words as
+    # bin/rig's uninstall_confirm, on purpose — one refusal in this codebase.
+    if [ ! -t 0 ]; then
+      printf 'rig-users: refusing to revoke every managed operator without --yes (no terminal to confirm on; RIG_YES=1 also means yes)\n' >&2
+      exit 2
+    fi
+    # `|| reply=""` is load-bearing: under `set -e` a read that hits EOF is a
+    # non-zero command, and an unguarded read would abort the script rather
+    # than take the safe default (#68, the same bug class).
+    printf 'rig-users: revoke all %s managed operator(s) on this box? [y/N] ' "$AT_RISK"
+    read -r reply || reply=""
+    case "$reply" in
+      y|Y|yes|YES|Yes) ;;
+      *) die "aborted — no operator was revoked" ;;
+    esac
+  fi
+fi
+
 if [ -r "$LEDGER" ]; then
   while read -r prev pstate _; do
     [ -n "$prev" ] || continue

test/cli.sh

@@ -1070,11 +1070,110 @@ if [ "$(id -u)" -ne 0 ]; then
   check "users apply: '@root' fixture parses, refuses non-root" 1 "must run as root" \
     "$ROOT/commands/users-apply.sh" --file "$FIX_BAD"
   check "users status: refuses non-root" 1 "must run as root" "$ROOT/commands/users-status.sh"
+  # --- the empty-file gate's flag surface (#65) ------------------------------
+  # Arg parsing precedes the root check, so flag ACCEPTANCE is provable here:
+  # reaching "must run as root" (exit 1) means --yes was taken, and an exit 2
+  # "unknown flag" would mean it was not. The gate's behaviour itself is
+  # root-only (it reads /etc/rig/users and revokes) and is grep-pinned below.
+  check "users apply: --yes is accepted" 1 "must run as root" \
+    "$ROOT/commands/users-apply.sh" --file "$FIX_OK" --yes
+  # Order-independent: a consent flag that only worked before --file would be
+  # a trap for anyone appending it to an existing command line.
+  check "users apply: --yes is accepted before --file" 1 "must run as root" \
+    "$ROOT/commands/users-apply.sh" --yes --file "$FIX_OK"
+  # --yes takes no value: it must not swallow the next argument.
+  check "users apply: --yes does not eat the following flag" 2 "unknown flag" \
+    "$ROOT/commands/users-apply.sh" --file "$FIX_OK" --yes --nope
+  # The env door, same as --yes: RIG_YES is the installer-family contract
+  # (bin/rig's uninstall_confirm reads it), so it must not be an unknown-flag
+  # equivalent or a parse error either.
+  check "users apply: RIG_YES=1 parses" 1 "must run as root" \
+    env RIG_YES=1 "$ROOT/commands/users-apply.sh" --file "$FIX_OK"
 else
   echo "skip: users non-root refusals (running as root)"
 fi
 rm -f "$FIX_OK" "$FIX_BAD"
 
+# --- the empty-file gate itself (#65) ----------------------------------------
+# The gated path needs root and a populated /etc/rig/users, so the shipped
+# script is grep-pinned instead — the house precedent for root-only refusals
+# (the '@root' keyless-seed die above, the invoker gate below).
+#
+# Consent has three doors and no fourth: --yes, RIG_YES, or a y on a TTY.
+check "users apply: --yes sets consent" 0 "" \
+  grep -qE '^[[:space:]]*--yes\) ASSUME_YES=1' "$ROOT/commands/users-apply.sh"
+check "users apply: RIG_YES is the env door for consent" 0 "" \
+  grep -qF 'RIG_YES:-' "$ROOT/commands/users-apply.sh"
+# The gate is ledger-gated, not file-gated: zero users ALONE is not the
+# condition, or it would refuse the empty-ledger no-op the issue calls
+# unambiguous. Both halves of the test must be present on the one line.
+# The gate condition and the counter are grepped as LITERALS — single quotes
+# intended throughout this block, the expansions are the script's own.
+# shellcheck disable=SC2016
+check "users apply: the gate is zero-users AND a readable ledger" 0 "" \
+  grep -qF 'if [ "${#USERS[@]}" -eq 0 ] && [ -r "$LEDGER" ] && [ "$ASSUME_YES" -eq 0 ]; then' \
+  "$ROOT/commands/users-apply.sh"
+# Counting precedes speaking, so the warning states a real number rather than
+# "some users" — and an already-revoked entry is not at risk, which is what
+# keeps a second identical run the silent no-op convergence promises.
+# shellcheck disable=SC2016
+check "users apply: the gate counts before it warns" 0 "" \
+  grep -qF 'AT_RISK=$((AT_RISK + 1))' "$ROOT/commands/users-apply.sh"
+# shellcheck disable=SC2016
+check "users apply: already-revoked ledger entries are not at risk" 0 "" \
+  grep -qF '[ "${pstate:-active}" != "revoked" ] || continue' "$ROOT/commands/users-apply.sh"
+# shellcheck disable=SC2016
+count_at="$(grep -nF 'AT_RISK=$((AT_RISK + 1))' "$ROOT/commands/users-apply.sh" | head -n1 | cut -d: -f1)"
+warn_at="$(grep -nF 'this users file names ZERO users' "$ROOT/commands/users-apply.sh" | head -n1 | cut -d: -f1)"
+check "users apply: the count is taken before the message quotes it" \
+  0 "" test "${count_at:-999999}" -lt "${warn_at:-0}"
+# No terminal and no consent is a REFUSAL, not an assumed yes and not a hang.
+check "users apply: no TTY and no consent exits 2" 0 "" \
+  grep -qF 'refusing to revoke every managed operator without --yes' \
+  "$ROOT/commands/users-apply.sh"
+check "users apply: the no-TTY refusal names RIG_YES as the other yes" 0 "" \
+  grep -qF 'no terminal to confirm on; RIG_YES=1 also means yes' \
+  "$ROOT/commands/users-apply.sh"
+# EOF-safe read (#68's bug class): an unguarded `read -r reply` aborts under
+# `set -e` instead of taking the safe default. The || is the whole fix.
+check "users apply: the confirm read survives EOF" 0 "" \
+  grep -qF 'read -r reply || reply=""' "$ROOT/commands/users-apply.sh"
+check "users apply: no unguarded read in the gate" 1 "" \
+  grep -nE '^[[:space:]]*read -r reply$' "$ROOT/commands/users-apply.sh"
+# The gate must sit BEFORE the revocation loop — a confirmation asked after
+# the first account is expired is not a confirmation. Line numbers, defaults
+# fail closed, same idiom as the visudo ordering assert.
+gate_at="$(grep -nF 'this users file names ZERO users' "$ROOT/commands/users-apply.sh" | head -n1 | cut -d: -f1)"
+revoke_at="$(grep -nF 'usermod -L -e 1' "$ROOT/commands/users-apply.sh" | head -n1 | cut -d: -f1)"
+check "users apply: the gate precedes the revocation loop" \
+  0 "" test "${gate_at:-999999}" -lt "${revoke_at:-0}"
+# Scope guard, the mirror of the #57 one above: this is a CONFIRMATION, and
+# apply must not have grown bootstrap's flat refusal of an empty file. A grep
+# that finds nothing is the pass — the repo's negative-law idiom.
+check "users apply: an empty file is still a legal de-provisioning input (gated, not refused)" 1 "" \
+  grep -nE 'names no users' "$ROOT/commands/users-apply.sh"
+# The deferred half of #65: mass revocation below the empty-file bright line
+# is NOT gated. The gate's only trigger is a file naming zero users, so no
+# CODE line may compare a revocation count against a threshold — comments are
+# stripped first, since the scope note beside the gate says the word on
+# purpose. Pinned so that adding a threshold is a deliberate edit to a failing
+# test rather than a silent contract change.
+check "users apply: partial mass revocation stays ungated (#65 open question)" 1 "" \
+  grep -nEi '^[[:space:]]*[^#[:space:]].*(threshold|RIG_REVOKE_MAX|AT_RISK[^)]*(-gt|-ge)[[:space:]]*\$)' \
+  "$ROOT/commands/users-apply.sh"
+
+# The empty-file gate is reachable only from a caller that can answer it. The
+# ONE in-tree caller of apply is bootstrap's users phase, and it refuses a
+# zero-user file at pre-flight (#57) — before it ever invokes apply — so no
+# in-tree path reaches the gate without a TTY. Pin both halves: if a second
+# caller appears, or bootstrap's refusal goes away, this stops being true.
+callers="$(grep -rlF 'users-apply.sh' "$ROOT/commands" | grep -v 'users-apply.sh$' || true)"
+check "users apply: bootstrap is its only in-tree caller" 0 "" \
+  test "$callers" = "$ROOT/commands/bootstrap.sh"
+check "users apply: bootstrap refuses a zero-user file before invoking it" \
+  0 "" test "$(grep -nF 'names no users' "$ROOT/commands/bootstrap.sh" | head -n1 | cut -d: -f1)" \
+  -lt "${users_apply_at:-0}"
+
 # Validate-then-apply: `visudo -c` must pass before anything lands in
 # /etc/sudoers.d — a bad drop-in takes down ALL of sudo, locking every admin
 # out of the escalation path apply just granted. Assert the order in the file,