Compare commits

...
Author SHA1 Message Date
hugocasaandClaude Opus 5 7d88a37dcf test(agents): make the tree-root rows follow the checkout kind
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-18 16:41:09 +02:00
hugocasaandClaude Opus 5 5ecbd15fca refactor(agents): state the glob and cd rationale once
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-18 10:41:09 +02:00
hugocasaandClaude Opus 5 5a58edb782 fix(agents): stop a failed cd from hiding the directory form
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-18 10:23:47 +02:00
hugocasaandClaude Opus 5 8f97f247c6 docs: prefer Edit/Write over shell edits in agent guidance
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-18 10:01:06 +02:00
hugocasaandClaude Opus 5 32e2f936f5 docs: state the one-write-per-line rule in the scratch guidance
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-18 09:33:33 +02:00
hugocasaandClaude Opus 5 fbb348e3b3 fix(agents): leave directory-form cp and mv unproved
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-18 09:16:18 +02:00
hugocasaandClaude Opus 5 e006924df4 fix(agents): prove where a directory-form cp or mv actually lands
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-17 19:58:52 +02:00
hugocasaandClaude Opus 5 9a8b01a864 fix(agents): treat sibling checkouts as separate roots
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-17 19:50:58 +02:00
hugocasaandClaude Opus 5 d51de5b7be fix(agents): never prove a command carrying a substitution or relative cd
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-17 19:42:11 +02:00
hugocasaandClaude Opus 5 a56f5c0dc2 fix(agents): keep redirects and chained writes off the allow path
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-17 19:29:52 +02:00
hugocasaandClaude Opus 5 6e67406303 fix(agents): close two auto-allow holes in the scratch guards
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-17 18:55:35 +02:00
hugocasaandClaude Opus 5 502d519481 docs: describe the checkout root in the scratch guidance
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-17 18:45:30 +02:00
hugocasaandClaude Opus 5 5639bf470b fix(agents): prove scratch file ops per command segment
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-17 18:45:05 +02:00
5 changed files with 579 additions and 220 deletions
+220 -88
View File
@@ -1,17 +1,31 @@
#!/usr/bin/env bash
# PreToolUse allowance for scratch file ops: auto-allow a single, plain, single-line
# `mkdir` / `cp` / `mv` / `touch` / `chmod` / `tar` / `unzip` whose every path operand
# resolves under /tmp. Anything else makes no decision (exit 0) and falls back to the normal
# permission flow, except for `mv` and `chmod`: those get an explicit `ask`, the only prompt
# they get (see lib-guarded-verb.sh).
# PreToolUse allowance for scratch file ops: auto-allow `mkdir` / `cp` / `mv` / `touch` /
# `chmod` whose every path operand resolves inside one of the roots `path_class` recognizes —
# under /tmp, or inside a git working tree under $HOME — and `tar` / `unzip` confined to /tmp.
# Anything else makes no decision (exit 0) and falls back to the normal permission flow, except
# for `mv` and `chmod`: those get an explicit `ask`, the only prompt they get (see
# lib-guarded-verb.sh).
#
# The command is read one segment at a time, so chaining and line breaks carry no weight of
# their own: `cd /tmp/scratch && mv /tmp/a /tmp/b` is proved on the operands of the `mv`. A
# decision covers the whole command line, so `allow` is emitted only when every segment is one
# of these verbs proved here or a `cd` that resolved, AND exactly one of them writes (see the
# gate at the foot of this file — an earlier write can change what a later operand means). A
# line that mixes a proven op with some other command makes no decision instead and leaves that
# line to the normal permission flow, rather than waving an unexamined command through with it.
#
# This is a hook rather than an allow rule because permission rules match a command prefix, so
# they can only constrain the FIRST operand. `cp /tmp/x ~/.zshrc` matches a `cp /tmp/` prefix,
# and requiring every operand is the point.
#
# Requiring the sources under /tmp too (not just the destination) keeps this from becoming a
# read-exfiltration path around the `Read(**/.env)` / `Read(**/secrets/**)` deny rules: a copy
# out of the project into /tmp would land the content somewhere `Read(/tmp/**)` allows.
# One operation may not straddle two roots, sources included, and a sibling checkout is a
# different root — `path_class` names the git tree, not just its kind. A copy out of a checkout
# into /tmp would be a read-exfiltration path around the `Read(**/secrets/**)` / `Read(**/*.pem)`
# deny rules, since the content lands where `Read(/tmp/**)` allows it to be read back, and one
# out of a repo the Read tool is not confined to would do the same for that repo. Keeping every
# operand of one operation inside a single root closes both without restating those rules here.
# The checkout root itself is what makes an in-repo `mv` or `chmod` auto-allowable: deleting a
# file there has never prompted, and moving or chmod-ing one is not the graver act.
#
# Deny-by-default tokenizing, in the same spirit as guard-rm-outside-tmp.sh: every path token
# must consist only of alphanumerics and `. _ / -`. That set contains none of the characters
@@ -19,12 +33,19 @@
# any glob character, so all of those forms fail by construction. `realpath -m` then resolves
# `..` and existing symlinks, so `/tmp/link` pointing at /etc/passwd is caught.
#
# `tar` and `unzip` keep the stricter rule — /tmp only, and absolute operands only — because
# their positional grammar makes a bare word ambiguous: `tar P -xf ...` is --absolute-names,
# not a file named P, and resolving it as a path would put an option in a root and allow it.
# The other five take relative operands, resolved against the working directory that `cd`
# tracking maintains, since for those a bare word really is a path (a GNU option starts with
# `-`, and the option allowlist below rejects the ones that would change symlink handling).
#
# `tar` and `unzip` get their own parser: their write destination arrives as a flag VALUE
# (`-C`, `-d`) rather than a positional, and a bundle like `-xzf` consumes the token after it.
# Flags are an allowlist, not a denylist, so `-P` / `--absolute-names` — which turn off tar's
# refusal to extract `..` and absolute member paths — defer rather than needing enumeration.
# Extraction additionally requires an explicit destination under /tmp, or a cwd already under
# /tmp, since otherwise members land in the project checkout.
# Extraction additionally requires an explicit destination under /tmp, or a working directory
# already under /tmp, since otherwise members land in the project checkout.
#
# Residual risk accepted: an archive whose members include a symlink pointing out of /tmp
# followed by a write through it can still escape, because tar applies member symlinks as it
@@ -52,25 +73,51 @@ defer() {
exit 0
}
# A newline separates commands, and the tokenizer below only reads the first line — defer.
case "$cmd" in *$'\n'*) defer "multi-line command" ;; esac
has_substitution "$cmd" && defer "command substitution in the command line"
read -r -a toks <<< "$cmd"
# 0 iff the token is a literal path this hook may reason about. A glob never auto-allows: bash
# expands it only after the hook has decided, so realpath sees the unexpanded pattern —
# `/tmp/link*` canonicalizes to itself and passes, then expands onto a symlink whose target is
# outside, and `cp` and `chmod` follow a command-line symlink, so that is a write to the target.
# (guard-rm-outside-tmp.sh can allow globs because `rm` unlinks the symlink rather than following
# it.) The charset holds none of the characters bash uses for quoting, expansion or separation.
literal_path() {
case "$1" in *[*?[]*) return 1 ;; esac
[ -z "$(printf '%s' "$1" | tr -d 'A-Za-z0-9._/-')" ]
}
# 0 iff the token is charset-safe and resolves to a path strictly inside /tmp.
# Prints the root class of a path token, then the path it resolved to on a second line,
# resolving a relative one against the tracked working directory. Fails, printing nothing,
# when the token is unsafe to reason about or lands outside every root.
operand_class() {
local t="$1" canon alt cls alt_cls=""
literal_path "$t" || return 1
case "$t" in
/*) canon=$(realpath -m -- "$t" 2>/dev/null) ;;
*) # A `cd` may fail at runtime and leave the command where it started, so a relative
# operand has to land in the same root either way.
[ -n "$seg_cwd" ] || return 1
canon=$(realpath -m -- "$seg_cwd/$t" 2>/dev/null)
if [ -n "$alt_cwd" ]; then
alt=$(realpath -m -- "$alt_cwd/$t" 2>/dev/null)
[ -n "$alt" ] || return 1
alt_cls=$(path_class "$alt") || return 1
fi
;;
esac
[ -n "$canon" ] || return 1
cls=$(path_class "$canon") || return 1
[ -n "$alt_cls" ] && [ "$alt_cls" != "$cls" ] && return 1
# Class and resolved path together: a caller runs this in a command substitution, so a global
# set here would be set in that subshell and lost.
printf '%s\n%s' "$cls" "$canon"
}
# 0 iff the token is charset-safe and resolves to a path strictly inside /tmp. The archive
# parser's stricter check; everything else goes through operand_class.
under_tmp() {
local t="$1" canon
# Globs never auto-allow. Bash expands them only after this hook has decided, so realpath
# sees the unexpanded pattern: `/tmp/link*` canonicalizes to itself and passes, then
# expands onto a symlink whose target is outside /tmp. chmod and cp follow command-line
# symlinks, so that is a write to the target. guard-rm-outside-tmp.sh can allow globs
# because `rm` unlinks the symlink itself rather than following it.
case "$t" in *[*?[]*) return 1 ;; esac
[ -n "$(printf '%s' "$t" | tr -d 'A-Za-z0-9._/-')" ] && return 1
# Absolute only. Resolving a relative operand against the cwd makes any bare word look like
# a safe path whenever the cwd is under /tmp, while the tool itself reads it as an option:
# `tar P -xf ...` is --absolute-names, not ./P, and `cp /tmp/t -RL /tmp/o` is a
# dereferencing recursive copy, not a file named -RL.
literal_path "$t" || return 1
case "$t" in /*) ;; *) return 1 ;; esac
canon=$(realpath -m -- "$t" 2>/dev/null)
[ -n "$canon" ] || return 1
@@ -79,29 +126,16 @@ under_tmp() {
return 1
}
# Bare command word only; wrappers (`timeout cp`), env prefixes, and `/bin/cp` defer.
# Options are an allowlist per command, so anything that changes how symlinks are followed
# defers instead of needing enumeration. `cp -L` / `-H` matter most: they dereference while
# recursing, which copies the CONTENT of a symlink target from outside /tmp into a scratch
# dir that `Read(/tmp/**)` then exposes. Plain `-r` and `-a` (which implies `-d`) recreate
# such a symlink as a symlink instead, so no outside content is materialized.
case "${toks[0]:-}" in
mkdir) takes_mode=0; ok_opts='pv' ;;
cp) takes_mode=0; ok_opts='rRvfnpa' ;;
mv) takes_mode=0; ok_opts='vfn' ;;
touch) takes_mode=0; ok_opts='acmv' ;;
chmod) takes_mode=1; ok_opts='Rvfc' ;; # chmod's first operand is a mode, not a path
tar) ok_flags='xctzjJavfC'; val_flags='fC' ;;
unzip) ok_flags='oqnljvd'; val_flags='d' ;;
*) defer "not the leading command word" ;;
esac
# ---------------------------------------------------------------- tar / unzip
if [ -n "${ok_flags:-}" ]; then
saw_archive=0 saw_dest=0 extracting=0 listing=0 end_opts=0
i=1
while [ "$i" -lt "${#toks[@]}" ]; do
t="${toks[$i]}"
# Proves one `tar` / `unzip` segment ($1 = the verb), whose tokens are in SEG_TOKS.
check_archive_segment() {
local verb="$1" ok_flags val_flags t flags val
local saw_archive=0 saw_dest=0 extracting=0 listing=0 end_opts=0 i=1
case "$verb" in
tar) ok_flags='xctzjJavfC'; val_flags='fC' ;;
unzip) ok_flags='oqnljvd'; val_flags='d' ;;
esac
while [ "$i" -lt "${#SEG_TOKS[@]}" ]; do
t="${SEG_TOKS[$i]}"
i=$((i + 1))
if [ "$end_opts" = 0 ]; then
[ "$t" = "--" ] && { end_opts=1; continue; }
@@ -112,13 +146,13 @@ if [ -n "${ok_flags:-}" ]; then
# leave a residue here and defer rather than being enumerated as denials.
[ -n "$(printf '%s' "$flags" | tr -d "$ok_flags")" ] && defer "unrecognized option \`$t\`"
case "$flags" in *x*) extracting=1 ;; esac
case "${toks[0]}$flags" in unzip*[lv]*) listing=1 ;; esac
case "$verb$flags" in unzip*[lv]*) listing=1 ;; esac
# A flag consuming the next token must be alone in its bundle's final position
# (`-xzf a.tar`), else the token it eats is ambiguous.
case "${flags%?}" in *[$val_flags]*) defer "ambiguous option bundle \`$t\`" ;; esac
case "${flags: -1}" in
[$val_flags])
val="${toks[$i]:-}"
val="${SEG_TOKS[$i]:-}"
i=$((i + 1))
[ -n "$val" ] || defer "option \`$t\` has no value"
under_tmp "$val" || defer "\`$val\` is outside /tmp"
@@ -136,54 +170,152 @@ if [ -n "${ok_flags:-}" ]; then
# first is the archive. Requiring every one under /tmp is conservative for member names,
# which are not filesystem paths — those defer rather than being wrongly allowed.
under_tmp "$t" || defer "\`$t\` is outside /tmp"
[ "${toks[0]}" = "unzip" ] && saw_archive=1
[ "$verb" = "unzip" ] && saw_archive=1
done
# tar without -f reads a tape/stdin; unzip needs an archive
[ "$saw_archive" = 1 ] || defer "no archive operand"
# Writes land relative to the working directory unless a destination was given. `unzip -l`
# and `-v` only list, so they need no destination.
if [ "$extracting" = 1 ] || { [ "${toks[0]}" = "unzip" ] && [ "$listing" = 0 ]; }; then
[ "$saw_dest" = 1 ] || under_tmp "${cwd:-$PWD}" || defer "extraction target is outside /tmp"
if [ "$extracting" = 1 ] || { [ "$verb" = "unzip" ] && [ "$listing" = 0 ]; }; then
# An extraction with no destination lands in the working directory. Word splitting cannot
# tell a `cd` inside a quoted string from one the shell runs, and believing a false one
# would put an archive's members in the checkout, so once any `cd` is in the line only an
# explicit destination will do.
[ "$saw_dest" = 1 ] \
|| { [ "$saw_cd" = 0 ] && [ -n "$seg_cwd" ] && under_tmp "$seg_cwd"; } \
|| defer "extraction target is outside /tmp"
fi
decide allow "archive paths and extraction target are under /tmp"
fi
}
# ------------------------------------------- mkdir / cp / mv / touch / chmod
path_operand=0
seen_mode=0
end_opts=0
i=1
while [ "$i" -lt "${#toks[@]}" ]; do
t="${toks[$i]}"
i=$((i + 1))
# Proves one `mkdir` / `cp` / `mv` / `touch` / `chmod` segment ($1 = the verb), whose tokens
# are in SEG_TOKS.
check_fileops_segment() {
local verb="$1" takes_mode ok_opts t cls resolved seen_class=""
local path_operand=0 seen_mode=0 end_opts=0 i=1 rel_operand=0
local -a ops=()
# Options are an allowlist per command, so anything that changes how symlinks are followed
# defers instead of needing enumeration. `cp -L` / `-H` matter most: they dereference while
# recursing, which copies the CONTENT of a symlink target from outside /tmp into a scratch
# dir that `Read(/tmp/**)` then exposes. Plain `-r` and `-a` (which implies `-d`) recreate
# such a symlink as a symlink instead, so no outside content is materialized.
case "$verb" in
mkdir) takes_mode=0; ok_opts='pv' ;;
cp) takes_mode=0; ok_opts='rRvfnpa' ;;
mv) takes_mode=0; ok_opts='vfn' ;;
touch) takes_mode=0; ok_opts='acmv' ;;
chmod) takes_mode=1; ok_opts='Rvfc' ;; # chmod's first operand is a mode, not a path
esac
while [ "$i" -lt "${#SEG_TOKS[@]}" ]; do
t="${SEG_TOKS[$i]}"
i=$((i + 1))
if [ "$end_opts" = 0 ]; then
[ "$t" = "--" ] && { end_opts=1; continue; }
# Checked at any position, not just before the first operand: GNU utils permute, so
# `cp /tmp/tree -RL /tmp/out` still enables dereferencing recursion.
case "$t" in
-?*)
# Allowlist: long options and the dereferencing flags leave a residue and defer.
[ -n "$(printf '%s' "${t#-}" | tr -d "$ok_opts")" ] && defer "unrecognized option \`$t\`"
continue
;;
esac
fi
if [ "$end_opts" = 0 ]; then
[ "$t" = "--" ] && { end_opts=1; continue; }
# Checked at any position, not just before the first operand: GNU utils permute, so
# `cp /tmp/tree -RL /tmp/out` still enables dereferencing recursion.
case "$t" in
-?*)
# Allowlist: long options and the dereferencing flags leave a residue and defer.
[ -n "$(printf '%s' "${t#-}" | tr -d "$ok_opts")" ] && defer "unrecognized option \`$t\`"
continue
;;
esac
fi
# chmod: consume the mode operand without a path check. Octal, or symbolic clauses.
if [ "$takes_mode" = 1 ] && [ "$seen_mode" = 0 ]; then
case "$t" in
[0-7] | [0-7][0-7] | [0-7][0-7][0-7] | [0-7][0-7][0-7][0-7]) ;;
*) printf '%s' "$t" | grep -Eq '^[ugoa]*[+=-][rwxXst]*(,[ugoa]*[+=-][rwxXst]*)*$' || defer "unrecognized mode \`$t\`" ;;
esac
seen_mode=1
continue
fi
# chmod: consume the mode operand without a path check. Octal, or symbolic clauses.
if [ "$takes_mode" = 1 ] && [ "$seen_mode" = 0 ]; then
case "$t" in
[0-7] | [0-7][0-7] | [0-7][0-7][0-7] | [0-7][0-7][0-7][0-7]) ;;
*) printf '%s' "$t" | grep -Eq '^[ugoa]*[+=-][rwxXst]*(,[ugoa]*[+=-][rwxXst]*)*$' || defer "unrecognized mode \`$t\`" ;;
esac
seen_mode=1
continue
fi
under_tmp "$t" || defer "\`$t\` is outside /tmp"
path_operand=1
resolved=$(operand_class "$t") || defer "\`$t\` is outside /tmp and not inside a git checkout in \$HOME"
cls="${resolved%%$'\n'*}"
# Every operand of one operation stays in one root: see the exfiltration note above.
[ -n "$seen_class" ] && [ "$cls" != "$seen_class" ] && defer "\`$t\` puts this $verb across two roots"
seen_class="$cls"
ops+=("${resolved#*$'\n'}")
case "$t" in /*) ;; *) rel_operand=1 ;; esac
path_operand=1
done
[ "$path_operand" = 1 ] || defer "no path operand"
# In directory form the command writes a path it does not name: `cp x dir` writes `dir/x`,
# and `cp` follows that child when it is a symlink — this checkout is full of them, every
# `*_ee.rs` pointing into the sibling EE repo. Deriving that child would mean reproducing
# which name the tool picks (the operand as written, not as resolved — a symlinked source
# keeps its own name) and how deep `-r` recurses. The form is left unproved instead.
case "$verb" in
cp | mv)
[ "${#ops[@]}" -ge 2 ] || return 0
# Whether the destination is an existing directory is itself a question about which of
# the two candidate working directories the command ran in, and only one of them is in
# `ops`. A `cd` that fails at runtime would otherwise let the form through: the
# destination resolved against the directory the command never reached is some path that
# does not exist, while the one it actually ran in is a directory full of symlinks.
[ -n "$alt_cwd" ] && [ "$rel_operand" = 1 ] \
&& defer "a relative operand after a \`cd\` lands in one of two directories"
[ -d "${ops[-1]}" ] \
&& defer "\`${ops[-1]}\` already exists as a directory, so this $verb writes a path it does not name"
;;
esac
}
split_segments "$cmd"
seg_cwd="${cwd:-$PWD}"
alt_cwd="" # where a `cd` that failed would have left the command
saw_cd=0 # a `cd` moved the working directory somewhere
proved=0 # how many ops came out inside a single root
only_ours=1 # ... and nothing else shares the command line
for seg in "${SEGMENTS[@]}"; do
segment_tokens "$seg"
case "${SEG_TOKS[0]:-}" in
"") continue ;;
mkdir | cp | mv | touch | chmod)
check_fileops_segment "${SEG_TOKS[0]}"
proved=$((proved + 1))
continue
;;
tar | unzip)
check_archive_segment "${SEG_TOKS[0]}"
proved=$((proved + 1))
continue
;;
cd)
# A `cd` writes nothing, so it never blocks an allow; it only moves where a later relative
# operand points, to one of the two candidates `apply_cd` describes.
if [ "$saw_cd" = 0 ] && new_cwd=$(apply_cd "$seg_cwd" "${SEG_TOKS[@]:1}"); then
alt_cwd="$seg_cwd"
seg_cwd="$new_cwd"
else
# Not the harmless segment an allow assumes: whatever this guard could not account for
# may be a redirect, and a redirect writes. Leave the line to the normal flow.
seg_cwd="" alt_cwd=""
only_ours=0
fi
saw_cd=1
continue
;;
esac
# Some other command shares the line. If an `mv` or `chmod` runs inside it after all — behind
# a wrapper, an env prefix or a path — this hook cannot say what it writes to.
for verb in mv chmod; do
segment_runs_verb "$verb" "$seg" && defer "$verb is not the leading command word in \`$seg\`"
done
only_ours=0
done
[ "$path_operand" = 1 ] || defer "no path operand"
decide allow "every path operand is under /tmp"
# Exactly one write per line. Each segment is proved against the filesystem as it stands now,
# and an earlier write can change what a later operand means: `cp -r /tmp/tree /tmp/live` that
# recreates a symlink out of /tmp turns `/tmp/live/link` — a path under /tmp when this ran —
# into a write through that symlink. Deletes compose safely and guard-rm-outside-tmp.sh allows
# several, because `rm` unlinks a symlink rather than following it.
[ "$proved" -ge 1 ] || exit 0
[ "$only_ours" = 1 ] && [ "$proved" = 1 ] && decide allow "every path operand is inside a single root"
exit 0
+104 -85
View File
@@ -1,14 +1,17 @@
#!/usr/bin/env bash
# PreToolUse guard for `rm`: auto-allow ONLY a single, plain, single-line `rm` whose every
# operand is a whitelisted target — under /tmp, or inside a git working tree located in $HOME
# (a version-controlled project dir). Any other command that runs `rm` gets an explicit `ask`,
# which is the ordinary permission prompt and the only one `rm` gets (see lib-guarded-verb.sh);
# a command that runs no `rm` at all makes no decision (exit 0).
# PreToolUse guard for `rm`: auto-allow deletes whose every operand is a whitelisted target —
# under /tmp, or inside a git working tree located in $HOME (a version-controlled project dir).
# Any other command that runs `rm` gets an explicit `ask`, which is the ordinary permission
# prompt and the only one `rm` gets (see lib-guarded-verb.sh); a command that runs no `rm` at
# all makes no decision (exit 0).
#
# The git-tree allowance trades on "this is a project under version control" being lower-stakes
# than a delete elsewhere — NOT on full recoverability: committed content is restorable via git,
# but untracked / .gitignore'd / uncommitted content, and an independent nested repo's history
# under a recursively-deleted parent, are NOT. Accepted as a deliberate convenience tradeoff.
# The command is read one segment at a time, so chaining and line breaks carry no weight of
# their own: `rm -f /tmp/a && rm -rf /tmp/b` is two deletes, each proved on its own operands.
# A decision covers the whole command line, so `allow` is emitted only when every segment is
# an `rm` this guard proved or a `cd` it could resolve. A line that mixes a proven `rm` with
# some other command makes no decision instead and leaves that line to the normal permission
# flow: the delete is not what needed a prompt, and waving the rest of the line through with
# it would turn a trailing `rm -f /tmp/x` into a way to auto-approve anything.
#
# Deny-by-default: every token must consist only of a safe character set (alphanumerics,
# `. _ / -` and glob chars `* ? [ ]`). That set contains none of the characters bash uses for
@@ -17,12 +20,12 @@
# and existing symlinks (so a symlink out of the allowed roots is caught), and a wildcard in a
# non-final path segment is refused because it can expand through a symlink realpath can't see.
#
# The git-repo allowance covers targets inside a git working tree under $HOME, and the tree's
# own root folder only when it is a linked worktree (`.git` is a pointer file, so history in
# the main repo survives); a primary checkout's root (`.git` is a history dir) and any `.git`,
# `.claude` or `.env` path are never auto-allowed. Globs auto-allow only under /tmp — elsewhere their expansion
# Which targets those two roots cover, and the tradeoff they rest on, is `path_class` in
# lib-guarded-verb.sh. Globs auto-allow only under /tmp — elsewhere their expansion
# could reach `.git` or a dotfile the literal checks never see. Relative operands resolve
# against the command's cwd (from the hook input).
# against the working directory the command runs from, which a `cd` in an earlier segment
# moves; once a `cd` is one this guard cannot resolve, that directory is unknown and a
# relative operand can no longer be proved.
#
# Assumes GNU `realpath` (-m) and `jq`, both present in this repo's Linux dev env.
set -uo pipefail
@@ -35,87 +38,103 @@ cmd=$(printf '%s' "$input" | jq -r '.tool_input.command // empty' 2>/dev/null)
cwd=$(printf '%s' "$input" | jq -r '.cwd // empty' 2>/dev/null)
# Every bail-out below goes through `defer`, so the forms this guard refuses to reason about —
# compound, quoted, wrapped — still reach the user as a prompt whenever an `rm` runs among them.
# wrapped, quoted, expanded — still reach the user as a prompt whenever an `rm` runs among them.
runs_verb rm "$cmd" && guarded=1 || guarded=0
defer() {
[ "$guarded" = 1 ] && decide ask "$1"
exit 0
}
# A newline separates commands, and the tokenizer below only reads the first line — defer.
case "$cmd" in *$'\n'*) defer "multi-line command" ;; esac
has_substitution "$cmd" && defer "command substitution in the command line"
read -r -a toks <<< "$cmd"
# Bare leading `rm` only; wrappers (`timeout rm`), env prefixes, and `/bin/rm` defer.
[ "${toks[0]:-}" = "rm" ] || defer "rm is not the leading command word"
# 0 (allow) iff the canonical path is an auto-allowable rm target: under /tmp, or strictly
# inside a git working tree located under $HOME. The walk stops at $HOME, so a dotfiles repo at
# ~ can't make all of $HOME deletable, and top-level ~ files stay protected.
allowed_target() {
local canon="$1" d root=""
case "$canon" in /tmp/?*) return 0 ;; esac
[ -n "${HOME:-}" ] || return 1
case "$canon" in "$HOME"/?*) ;; *) return 1 ;; esac
# Never auto-allow: history, and the two kinds of path the "it's under version control"
# premise doesn't hold for — the agent's own guards and settings (deleting them is what
# removes the prompt on everything else), and gitignored `.env` files.
case "$canon" in
*"/.git" | *"/.git/"* | *"/.claude" | *"/.claude/"*) return 1 ;;
*"/.env" | *"/.env."*) return 1 ;;
esac
d="$canon"
while [ "$d" != "/" ] && [ "$d" != "$HOME" ]; do
[ -e "$d/.git" ] && { root="$d"; break; }
d=$(dirname "$d")
# Proves one `rm` segment, whose tokens are in SEG_TOKS with `rm` at index 0, resolving relative
# operands against $seg_cwd. Returns only once every operand is an auto-allowable target;
# anything it cannot prove defers instead.
check_rm_segment() {
local i=1 t canon candidates had_operand=0 end_opts=0
while [ "$i" -lt "${#SEG_TOKS[@]}" ]; do
t="${SEG_TOKS[$i]}"
i=$((i + 1))
# Whitelist every token (flags included, so an operator hidden in a flag like `-rf;rm`
# can't slip past): any character outside the safe set makes it unsafe to reason about.
[ -n "$(printf '%s' "$t" | tr -d 'A-Za-z0-9._/*?[]-')" ] && defer "unsafe characters in \`$t\`"
# A glob in an option-looking token (`-[-]`) can expand to `--` and turn a later `-name`
# into an operand — never a real option, so defer.
case "$t" in -*[*?[]*) defer "glob inside the option \`$t\`" ;; esac
if [ "$end_opts" = 0 ]; then
[ "$t" = "--" ] && { end_opts=1; continue; }
# Skip real options only before the first operand. A bare `-` is a filename, and under
# POSIXLY_CORRECT GNU rm stops option parsing at the first operand, so a later `-name`
# is a filename too — validate it rather than skipping it.
if [ "$had_operand" = 0 ]; then
case "$t" in -?*) continue ;; esac
fi
fi
had_operand=1
# No wildcard in a non-final path segment (`a/*/b`): it can expand through a symlink
# realpath can't see. A slashless glob (`*.rs`) is a final-segment match — fine.
case "$t" in */*) case "${t%/*}" in *[*?[]*) defer "glob in a non-final segment of \`$t\`" ;; esac ;; esac
# A relative operand has as many candidate paths as the command has candidate working
# directories, and every one of them has to be auto-allowable: a `cd` that fails at runtime
# leaves the delete running in the directory it started in.
case "$t" in
/*) candidates=$(realpath -m -- "$t" 2>/dev/null) ;;
*) [ -n "$seg_cwd" ] || defer "\`$t\` is relative to a working directory this guard cannot pin down"
candidates=$(realpath -m -- "$seg_cwd/$t" 2>/dev/null)
[ -n "$alt_cwd" ] && candidates="$candidates
$(realpath -m -- "$alt_cwd/$t" 2>/dev/null)"
;;
esac
while IFS= read -r canon; do
[ -n "$canon" ] || defer "cannot resolve \`$t\`"
# A glob may auto-allow only under /tmp, where everything is deletable. Elsewhere its
# expansion could match `.git`, a dotfile like `.*`, or a nested checkout root that the
# literal-path checks never see — so require literal operands in git repos.
case "$t" in *[*?[]*) case "$canon" in /tmp/?*) ;; *) defer "glob \`$t\` is outside /tmp" ;; esac ;; esac
path_class "$canon" >/dev/null || defer "\`$canon\` is outside /tmp and not inside a git checkout in \$HOME"
done <<< "$candidates"
done
[ -n "$root" ] || return 1 # not inside a git working tree under $HOME
if [ "$canon" = "$root" ]; then
# Deleting the repo root folder itself: allow only for a linked worktree, whose `.git` is
# a file/pointer so the history lives in the main repo and survives. A primary checkout's
# `.git` is a directory holding the history, so deleting it is unrecoverable — defer.
[ -f "$root/.git" ] && return 0
return 1
fi
return 0
[ "$had_operand" = 1 ] || defer "no operand"
}
had_operand=0
end_opts=0
i=1
while [ "$i" -lt "${#toks[@]}" ]; do
t="${toks[$i]}"
i=$((i + 1))
# Whitelist every token (flags included, so an operator hidden in a flag like `-rf;rm`
# can't slip past): any character outside the safe set makes it unsafe to reason about.
[ -n "$(printf '%s' "$t" | tr -d 'A-Za-z0-9._/*?[]-')" ] && defer "unsafe characters in \`$t\`"
# A glob in an option-looking token (`-[-]`) can expand to `--` and turn a later `-name`
# into an operand — never a real option, so defer.
case "$t" in -*[*?[]*) defer "glob inside the option \`$t\`" ;; esac
if [ "$end_opts" = 0 ]; then
[ "$t" = "--" ] && { end_opts=1; continue; }
# Skip real options only before the first operand. A bare `-` is a filename, and under
# POSIXLY_CORRECT GNU rm stops option parsing at the first operand, so a later `-name`
# is a filename too — validate it rather than skipping it.
if [ "$had_operand" = 0 ]; then
case "$t" in -?*) continue ;; esac
fi
fi
had_operand=1
# No wildcard in a non-final path segment (`a/*/b`): it can expand through a symlink
# realpath can't see. A slashless glob (`*.rs`) is a final-segment match — fine.
case "$t" in */*) case "${t%/*}" in *[*?[]*) defer "glob in a non-final segment of \`$t\`" ;; esac ;; esac
case "$t" in
/*) canon=$(realpath -m -- "$t" 2>/dev/null) ;;
*) canon=$(realpath -m -- "${cwd:-$PWD}/$t" 2>/dev/null) ;;
split_segments "$cmd"
seg_cwd="${cwd:-$PWD}"
alt_cwd="" # where a `cd` that failed would have left the command
saw_cd=0
proved=0 # at least one `rm` segment came out auto-allowable
only_ours=1 # ... and nothing else shares the command line
for seg in "${SEGMENTS[@]}"; do
segment_tokens "$seg"
case "${SEG_TOKS[0]:-}" in
"") continue ;;
rm)
check_rm_segment
proved=1
continue
;;
cd)
# A `cd` writes nothing, so it never blocks an allow; it only moves where a later relative
# operand points, to one of the two candidates `apply_cd` describes.
if [ "$saw_cd" = 0 ] && new_cwd=$(apply_cd "$seg_cwd" "${SEG_TOKS[@]:1}"); then
alt_cwd="$seg_cwd"
seg_cwd="$new_cwd"
else
# Not the harmless segment an allow assumes: whatever this guard could not account for
# may be a redirect, and a redirect writes. Leave the line to the normal flow.
seg_cwd="" alt_cwd=""
only_ours=0
fi
saw_cd=1
continue
;;
esac
[ -n "$canon" ] || defer "cannot resolve \`$t\`"
# A glob may auto-allow only under /tmp, where everything is deletable. Elsewhere its
# expansion could match `.git`, a dotfile like `.*`, or a nested checkout root that the
# literal-path checks never see — so require literal operands in git repos.
case "$t" in *[*?[]*) case "$canon" in /tmp/?*) ;; *) defer "glob \`$t\` is outside /tmp" ;; esac ;; esac
allowed_target "$canon" || defer "\`$canon\` is outside /tmp and not inside a git checkout in \$HOME"
# Some other command shares the line. If an `rm` runs inside it after all — behind a wrapper,
# an env prefix or a path — this guard cannot say what it deletes.
segment_runs_verb rm "$seg" && defer "rm is not the leading command word in \`$seg\`"
only_ours=0
done
[ "$had_operand" = 1 ] || defer "no operand"
decide allow 'rm operands are under /tmp or inside a git checkout in $HOME'
[ "$proved" = 1 ] || exit 0
[ "$only_ours" = 1 ] && decide allow 'rm operands are under /tmp or inside a git checkout in $HOME'
exit 0
+152 -39
View File
@@ -10,17 +10,6 @@
# expand a glob operand against the filesystem. Neither guard relies on pathname expansion.
set -f
# 0 iff <verb> ($1) runs as a command word anywhere in <command> ($2). Mirrors how a Bash
# permission rule matches, so that owning the prompt here doesn't narrow what used to prompt:
# the command splits on `; & |` and newlines, and a leading env assignment or process wrapper
# (`timeout 5 rm`, `xargs rm`) is skipped before the command word is read.
#
# The split set also carries the characters that open a nested command — `$(`, backticks and
# `( )` — because a rule matches the verb inside one (`echo $(rm -rf ~)` prompts), and a
# separator that only ends statements would read that as an `echo`. Braces are handled as
# words rather than separators, since splitting on them cuts `xargs -I {} … rm` in half and
# strands the `rm` in a segment that no longer knows a wrapper preceded it.
# 0 iff <text> ($1) starts with a command that only reads its input. An allowlist, because the
# opposite — naming the shells to avoid — would have to be complete: an unlisted one (`ash`,
# `rbash`, `busybox sh`) executes the body while the guard calls it data. Unrecognized here only
@@ -115,35 +104,159 @@ strip_heredoc_bodies() {
done
}
# 0 iff <verb> ($1) runs as a command word in <segment> ($2), which must already be one
# segment (no separator left in it). Wrapper, env-prefix and `/bin/<verb>` forms all count.
segment_runs_verb() {
local verb="$1" w wrapped=0
for w in $2; do
# The shell strips quotes and backslashes before it looks up the command, so `'rm'` and
# `r\m` run rm and have to compare equal to it.
w="${w//[\"\'\\]/}"
case "$w" in
"$verb" | */"$verb") return 0 ;;
*=*) ;; # leading env assignment
-* | *'>'* | *'<'*) ;; # a flag, or a leading redirect
[0-9]*) [ "$wrapped" = 1 ] || break ;; # a wrapper's duration, not `1:` in prose
'!' | '{' | '}' | if | then | elif | else | while | until | do) ;; # never the command
timeout | time | nice | nohup | stdbuf | command | builtin | noglob | xargs | sudo | env)
wrapped=1 ;;
# A wrapper's option value is indistinguishable from a command name (`stdbuf -o L rm`),
# so past a wrapper the scan runs to the end of the segment instead of stopping at the
# first ordinary word. Before one, that word is the command and the verb cannot follow
# it. Nothing bounds the scan: a wrapper takes unboundedly many operands
# (`env -u A -u B ...`), and any cutoff — a word count, or stopping at the first quoted
# word — drops the prompt for a real `sudo -u 'root' rm`. Prose after a wrapper is the
# price, and it only over-prompts.
*) [ "$wrapped" = 1 ] || break ;;
esac
done
return 1
}
# Splits <command> ($1) into its command segments, into the global array SEGMENTS. Every guard
# reasons one segment at a time, so `a && b` is two commands here rather than one unparsable
# blob, and a newline is a separator like any other.
#
# The split set carries more than `; & |` and newlines: `$(`, backticks and `( )` open a nested
# command, and a separator that only ended statements would read `echo $(rm -rf ~)` as an
# `echo`. Braces are handled as words rather than separators, since splitting on them cuts
# `xargs -I {} … rm` in half and strands the `rm` in a segment that no longer knows a wrapper
# preceded it.
#
# `tr` and not `${1//[...]}`: a `}` inside the bracket expression closes the expansion itself,
# which silently leaves the command unsplit and every separator unseen.
split_segments() {
local seg
SEGMENTS=()
while IFS= read -r seg; do SEGMENTS+=("$seg"); done <<< "$(strip_heredoc_bodies "$1" | tr ';&|()`' '\n')"
}
# 0 iff <command> ($1) carries a command substitution outside a heredoc body. A substitution is
# concatenated into the word it sits in, and splitting on its opener cuts that word in half:
# `/tmp/a/`printf ../../etc`` would be proved as `/tmp/a/`, with the traversal validated as an
# unrelated segment. Nothing here can evaluate it, so a guard proves nothing about such a
# command. Heredoc bodies are excepted — those are data the split has already dropped.
has_substitution() {
case "$(strip_heredoc_bodies "$1")" in
*'$('* | *'`'*) return 0 ;;
esac
return 1
}
# Reads <segment> ($1) into the global array SEG_TOKS, dropping the shell keywords that can
# precede a command word so that `then rm -rf x` is analyzed as the `rm` it runs. Word
# splitting only: quotes are left in the token and fail the guards' charset check downstream,
# which is what keeps `rm -rf "$HOME/x"` unprovable.
segment_tokens() {
SEG_TOKS=()
read -r -a SEG_TOKS <<< "$1"
while [ "${#SEG_TOKS[@]}" -gt 0 ]; do
case "${SEG_TOKS[0]}" in
'!' | '{' | '}' | if | then | elif | else | while | until | do) SEG_TOKS=("${SEG_TOKS[@]:1}") ;;
*) break ;;
esac
done
}
# Prints the directory a `cd` lands in, given the current one ($1) and the tokens after the
# `cd` ($2...). Fails, printing nothing, when the destination cannot be resolved — a variable,
# `-`, an option, a relative path, no operand at all (`cd` alone is $HOME), or more than one.
#
# Resolving says nothing about whether the `cd` will SUCCEED: the destination may not exist, and
# `;` runs the next command anyway, leaving it in the directory it started in. So a caller may
# never treat this as the working directory outright — it is one of two candidates, and a
# relative operand has to be provable against the one the command started in as well. That also
# makes a `cd` word splitting invented out of quoted text harmless: it can only add a candidate,
# never drop one. Past the first `cd` the branching outruns two candidates, so a caller that
# sees a second gives up on relative operands entirely.
apply_cd() {
local cwd="$1" t
shift
[ "$#" -eq 1 ] || return 1
t="$1"
[ -n "$(printf '%s' "$t" | tr -d 'A-Za-z0-9._/-')" ] && return 1
# Absolute only. A relative destination is not `$cwd/$t`: the shell searches $CDPATH first,
# so `cd ssh` may land in /etc/ssh, and this cannot see the caller's $CDPATH to rule it out.
case "$t" in /*) ;; *) return 1 ;; esac
realpath -m -- "$t" 2>/dev/null
}
# Prints the class of a canonical path and returns 0: `tmp` for one strictly under /tmp, or
# `repo:<root>` for one strictly inside the git working tree at <root>, itself under $HOME.
# Fails, printing nothing, for anything else — those are the only roots the guards are willing
# to touch unprompted. The root is part of the class so that a caller pairing two operands can
# tell one checkout from another: sibling repos are separate permission boundaries, not one.
#
# The `repo` class trades on "this is a project under version control" being lower-stakes than
# the same act elsewhere — NOT on full recoverability: committed content is restorable via git,
# but untracked / .gitignore'd / uncommitted content, and an independent nested repo's history
# under a recursively-deleted parent, are NOT. Accepted as a deliberate convenience tradeoff.
#
# The walk stops at $HOME, so a dotfiles repo at ~ can't put all of $HOME in a class, and
# top-level ~ files stay out of one. A working tree's own root folder counts only when it is a
# linked worktree, whose `.git` is a pointer file so the history lives in the main repo and
# survives; a primary checkout's `.git` is a directory holding the history itself, so losing it
# is unrecoverable.
#
# Some paths are in no class in any root, /tmp included. Git history, and the agent's own guards
# and settings, because removing those is what removes the prompt on everything else. And every
# path `.claude/settings.json` refuses to read — `.env`, `secrets/`, `*.pem`, `*.key`,
# `credentials.json`, `.secret*` — because a `cp` or `mv` that is auto-allowed on both ends
# would rename one out of those globs and hand back through `Read` exactly what they deny.
path_class() {
local canon="$1" d root=""
case "$canon" in
*"/.git" | *"/.git/"* | *"/.claude" | *"/.claude/"*) return 1 ;;
*"/.env" | *"/.env."*) return 1 ;;
*"/secrets" | *"/secrets/"*) return 1 ;;
*.pem | *.key | *"/credentials.json") return 1 ;;
*"/.secret"* | *.secret | *.secrets) return 1 ;;
esac
case "$canon" in /tmp/?*) printf 'tmp'; return 0 ;; esac
[ -n "${HOME:-}" ] || return 1
case "$canon" in "$HOME"/?*) ;; *) return 1 ;; esac
d="$canon"
while [ "$d" != "/" ] && [ "$d" != "$HOME" ]; do
[ -e "$d/.git" ] && { root="$d"; break; }
d=$(dirname "$d")
done
[ -n "$root" ] || return 1 # not inside a git working tree under $HOME
if [ "$canon" = "$root" ]; then
[ -f "$root/.git" ] || return 1
fi
printf 'repo:%s' "$root"
}
# 0 iff <verb> ($1) runs as a command word anywhere in <command> ($2). Mirrors how a Bash
# permission rule matches, so that owning the prompt here doesn't narrow what used to prompt:
# a guard consults this before it starts proving segments, and every bail-out it then takes
# is a prompt for exactly the commands a rule would have caught.
runs_verb() {
local verb="$1" seg w wrapped
while IFS= read -r seg; do
wrapped=0
for w in $seg; do
# The shell strips quotes and backslashes before it looks up the command, so `'rm'` and
# `r\m` run rm and have to compare equal to it.
w="${w//[\"\'\\]/}"
case "$w" in
"$verb" | */"$verb") return 0 ;;
*=*) ;; # leading env assignment
-* | *'>'* | *'<'*) ;; # a flag, or a leading redirect
[0-9]*) [ "$wrapped" = 1 ] || break ;; # a wrapper's duration, not `1:` in prose
'!' | '{' | '}' | if | then | elif | else | while | until | do) ;; # never the command
timeout | time | nice | nohup | stdbuf | command | builtin | noglob | xargs | sudo | env)
wrapped=1 ;;
# A wrapper's option value is indistinguishable from a command name (`stdbuf -o L rm`),
# so past a wrapper the scan runs to the end of the segment instead of stopping at the
# first ordinary word. Before one, that word is the command and the verb cannot follow
# it. Nothing bounds the scan: a wrapper takes unboundedly many operands
# (`env -u A -u B …`), and any cutoff — a word count, or stopping at the first quoted
# word — drops the prompt for a real `sudo -u 'root' rm`. Prose after a wrapper is the
# price, and it only over-prompts.
*) [ "$wrapped" = 1 ] || break ;;
esac
done
# `tr` and not `${2//[...]}`: a `}` inside the bracket expression closes the expansion
# itself, which silently leaves the command unsplit and every separator unseen.
done <<< "$(strip_heredoc_bodies "$2" | tr ';&|()`' '\n')"
local verb="$1" seg
split_segments "$2"
for seg in "${SEGMENTS[@]}"; do
segment_runs_verb "$verb" "$seg" && return 0
done
return 1
}
+90 -4
View File
@@ -4,12 +4,26 @@
# What this pins is the `ask` column: a matcher change that turns one into a no-decision drops
# that command's only prompt (see lib-guarded-verb.sh). The wrapper, nested-command and quoted
# rows are the ones that catch it.
#
# The `allow` column carries its own weight, because a decision covers the whole command line:
# `allow` may only appear where every segment was proved here, and a line that also runs
# something unexamined has to come out `none` so the normal permission flow still sees it.
set -uo pipefail
H="$(cd "${BASH_SOURCE[0]%/*}" && pwd)"
CWD="$(git -C "$H" rev-parse --show-toplevel)"
OUT="$HOME/not-a-git-tree" # never written to; only the guards' path checks look at it
fails=0
# A tree's own root is auto-allowable only when it is a LINKED worktree, whose `.git` is a
# pointer file so the history lives in the main repo and survives; a primary checkout's `.git`
# is the history itself. The suite runs from either kind, so the rows that name the root follow
# the one it is run in — which is also what pins both halves of that rule.
if [ -f "$CWD/.git" ]; then
ROOT_SOLO=allow ROOT_CHAINED=none # linked worktree
else
ROOT_SOLO=ask ROOT_CHAINED=ask # primary checkout
fi
run() { # run <hook> <allow|ask|none> <command>
local hook="$1" want="$2" cmd="$3" out got
out=$(jq -nc --arg c "$cmd" --arg w "$CWD" \
@@ -41,15 +55,16 @@ run $G ask "rm $CWD/.claude/settings.json"
run $G ask "rm $CWD/.claude/settings.local.json"
run $G ask "rm -rf $CWD/backend/.env"
run $G ask "rm -rf $CWD/.env.local"
run $G ask "rm -rf $CWD"
run $G $ROOT_SOLO "rm -rf $CWD"
run $G ask "rm -rf $CWD/*"
run $G ask "rm -rf /etc/passwd"
run $G ask 'rm -rf "$HOME/x"'
run $G ask "rm -rf /tmp/../$OUT"
run $G ask "ls /tmp && rm -rf /tmp/x"
run $G none "ls /tmp && rm -rf /tmp/x" # proved delete, unexamined neighbour
run $G ask 'echo $(rm -rf /etc)'
run $G ask 'echo `rm -rf /etc`'
run $G ask "{ rm -rf /etc; }"
run $G allow "{ rm -rf /tmp/scratch/x; }" # the keyword drops, the delete still proves
run $G ask "find . -name x | xargs rm"
run $G ask "timeout 5 rm -rf /tmp/x"
run $G ask "stdbuf -o L rm -rf /etc"
@@ -107,6 +122,33 @@ run $G none 'echo $(ls /tmp)'
run $G none 'grep -rn "rm" backend/'
run $G none "cargo build --release"
# Chaining and line breaks are not themselves a reason to prompt: each segment is proved on its
# own operands, and a `cd` moves where a relative one points.
run $G allow "rm -f /tmp/a; rm -rf /tmp/b"
run $G allow "$(printf 'rm -f /tmp/a\nrm -rf %s/frontend/scratch' "$CWD")"
run $G allow "cd /tmp/scratch && rm -rf sub"
run $G none "mkdir -p /tmp/x && rm -rf /tmp/x"
run $G ask "$(printf 'ls /tmp\nrm -rf /etc')"
# A `cd` this guard can resolve is where the relative operand lands; one it cannot leaves the
# working directory unknown, and an unknown one proves nothing.
run $G ask "cd /etc && rm -rf foo"
run $G ask 'cd "$D" && rm -rf foo'
run $G ask "cd $CWD && rm -rf .git"
run $G ask "cd /etc && cd /tmp/scratch && rm -rf sub" # a cd out is not walked back
# A `cd` can fail at runtime, and `;` runs the delete from where the command started, so a
# relative operand is proved from both directories.
run $G ask "cd /tmp/does-not-exist; rm -rf .git"
run $G ask "cd /tmp/does-not-exist; rm -rf backend/.env"
run $G ask "cd /tmp/a && cd /tmp/b && rm -rf sub"
run $G ask "rm -rf /tmp/clone/.git" # history is never in a class
run $G ask "rm -rf /tmp/scratch/id_rsa.key"
run $G none "cd /tmp >$OUT; rm -f /tmp/a"
# A substitution is concatenated into its word, so splitting on it would prove only the literal
# half; a relative `cd` is not $cwd/$t either, since the shell searches $CDPATH first.
run $G ask 'rm -rf /tmp/a/`printf ../../etc`'
run $G ask 'rm -rf /tmp/a/$(printf ../../etc)'
run $G ask "cd ssh && rm -rf moduli"
echo
echo "== allow-fileops-in-tmp.sh =="
A=allow-fileops-in-tmp.sh
@@ -116,8 +158,8 @@ run $A allow "cp -r /tmp/a /tmp/b"
run $A allow "tar -xzf /tmp/a.tar.gz -C /tmp/out"
run $A ask "mv /tmp/a $OUT"
run $A ask "mv $CWD/AGENTS.md /tmp/a"
run $A ask "chmod -R 777 $CWD"
run $A ask "ls && mv /tmp/a /tmp/b"
run $A $ROOT_SOLO "chmod -R 777 $CWD"
run $A none "ls && mv /tmp/a /tmp/b" # proved move, unexamined neighbour
run $A ask 'echo $(mv /tmp/a /etc)'
run $A ask "timeout --signal KILL 5 mv /tmp/a /etc"
run $A ask "time -f FORMAT chmod 777 $OUT"
@@ -129,5 +171,49 @@ run $A none "cp $CWD/AGENTS.md /tmp/a"
run $A none "tar -xzf /tmp/a.tar.gz -C $OUT"
run $A none "cargo build"
run $A none "mkdir -p /tmp/x; mv /tmp/a /tmp/x; chmod 755 /tmp/x" # one write per line
run $A none "$(printf 'mv /tmp/a /tmp/b\nchmod 755 /tmp/b')"
run $A ask "ls && mv /tmp/a /etc"
run $A $ROOT_CHAINED "$(printf 'mkdir -p /tmp/x\nchmod -R 777 %s' "$CWD")"
run $A allow "cd /tmp/x && tar -xzf /tmp/a.tar.gz -C /tmp/out"
# The checkout is a root of its own, so an in-repo move or chmod is as auto-allowable as the
# in-repo delete already was — but one operation may not straddle it and /tmp.
run $A allow "chmod +x scripts/worktree-env"
run $A allow "mv backend/.sqlx backend/.sqlx.bad"
run $A allow "mv $CWD/frontend/a.ts $CWD/frontend/b.ts"
run $A ask "mv /tmp/a $CWD/frontend/a.ts"
run $A ask "chmod -R 777 $CWD/.git"
run $A ask "mv $CWD/backend/.env $CWD/backend/.env.bak"
run $A ask "mv $CWD/AGENTS.md $OUT"
run $A ask "cd /etc && mv a b"
# An auto-allowed rename may not carry a path out of the `Read` deny globs.
run $A ask "mv backend/server.pem backend/server.txt"
run $A none "cp backend/secrets/token frontend/token.txt" # cp has no prompt of its own,
# so what matters is it is not allowed
run $A ask "mv $CWD/backend/credentials.json /tmp/x"
run $A ask "cd /tmp/does-not-exist; mv .claude/settings.json settings.bak"
# A segment this hook cannot read whole may carry a redirect, and an earlier write can change
# what a later operand resolves to — neither may ride along on an allow.
run $A none "cd /tmp >$OUT; mv /tmp/a /tmp/b"
run $A none "cp -r /tmp/tree /tmp/live; cp /tmp/payload /tmp/live/link"
run $A ask 'mv /tmp/a/`printf ../../etc/x` /tmp/b'
# A sibling checkout is a different root: its files are outside what the Read tool is confined
# to, and copying them in would hand back what that confinement withholds.
EE="$(dirname "$CWD")/windmill-ee-private" # a sibling checkout; absent elsewhere, still not a root
run $A ask "mv $EE/backend/x.rs $CWD/backend/x.rs"
run $A none "cp $EE/README.md $CWD/README.copy"
# Directory form writes a path the command does not name — DEST/basename(SRC) — and `cp`
# follows that child when it is a symlink, as every `*_ee.rs` in this checkout is.
run $A ask "mv frontend/apps_ee.rs backend/windmill-api/src"
run $A none "cp frontend/apps_ee.rs backend/windmill-api/src"
run $A none "cp frontend/a.ts backend"
run $A ask "mv /tmp/a $CWD/backend"
# ... and a `cd` that fails at runtime may not hide that form: the destination is a directory
# in the directory the command actually ran in, whichever of the two that turns out to be.
run $A none "cd $CWD/AGENTS.md; cp frontend/apps_ee.rs backend/windmill-api/src"
run $A ask "cd $CWD/AGENTS.md; mv frontend/apps_ee.rs backend/windmill-api/src"
run $A none "cd /tmp/x && tar -xzf /tmp/a.tar.gz" # no -C, and the cwd is now two candidates
run $A allow "cp frontend/a.ts backend/a.ts" # ... naming the destination proves fine
echo
[ "$fails" = 0 ] && echo "ALL PASS" || { echo "$fails FAILURES"; exit 1; }
+13 -4
View File
@@ -146,10 +146,19 @@ $NAV --root backend callees "X" # what does X call?
- **MUST `outline` before `Read`** on unfamiliar files — then `body` or `Read` with offset/limit for specifics
- **Scratch stays outside the checkout.** Temp scripts, data dumps, cache backups and
screenshots go in the session scratch directory or `/tmp`, so nothing temporary can end up
committed. Write `rm`/`mv`/`cp` as one plain unchained command: a PreToolUse hook
auto-allows those when every operand is under `/tmp` or inside this checkout, but it defers
on `&&`, `;`, redirects, quotes and `$VAR` — that deferral, not the delete itself, is what
turns a routine cleanup into a permission prompt.
committed. Write the paths in `rm`/`mv`/`cp` out literally: a PreToolUse hook proves each
operand, and auto-allows deletes, moves, copies and mode changes under `/tmp` or inside a git
checkout under `$HOME`, as long as one operation stays within a single root — a sibling
checkout is a root of its own (`tar` and `unzip` stay `/tmp`-only). Chain deletes freely, each
proved on its own operands, but keep writes to one per line, name the destination rather than
a directory to drop it in, and put anything else on its own line: a command the hook does not
prove drops the whole line back to the normal permission flow. A
quoted or `$VAR` operand, a `~`, a redirect, a `$(…)`, a relative `cd`, or a wrapper like
`xargs rm` cannot be proved, and that deferral is what turns a cleanup into a prompt.
- **Change files with Edit/Write, not the shell.** `sed -i`, `cat > file <<'EOF'` and inline
`python3 - <<'PY'` scripts put an edit through the PreToolUse guards and the permission
classifier, which match `Bash` and nothing else, so a routine edit arrives as a prompt. Bash
stays right for running things — tests, builds, git, one-off queries.
- Search for existing code to reuse before writing new code
- Follow established patterns in the codebase
- Keep changes focused — don't refactor beyond what's asked