From 5639bf470b968bf88958ff7f79dc6bee6dd85f53 Mon Sep 17 00:00:00 2001
From: hugocasa <hugo@casademont.ch>
Date: Mon, 17 Aug 2026 18:26:03 +0200
Subject: [PATCH] fix(agents): prove scratch file ops per command segment

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 .claude/hooks/allow-fileops-in-tmp.sh | 248 +++++++++++++++++---------
 .claude/hooks/guard-rm-outside-tmp.sh | 178 +++++++++---------
 .claude/hooks/lib-guarded-verb.sh     | 166 +++++++++++++----
 .claude/hooks/test-hooks.sh           |  38 +++-
 AGENTS.md                             |   9 +-
 5 files changed, 426 insertions(+), 213 deletions(-)

diff --git a/.claude/hooks/allow-fileops-in-tmp.sh b/.claude/hooks/allow-fileops-in-tmp.sh
index 665c23cf0a..00d474940a 100755
--- a/.claude/hooks/allow-fileops-in-tmp.sh
+++ b/.claude/hooks/allow-fileops-in-tmp.sh
@@ -1,17 +1,28 @@
 #!/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: `mkdir -p /tmp/a && mv /tmp/b /tmp/a` is two operations, each proved on its own
+# operands. 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. 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 the two roots, sources included. 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; keeping
+# every operand of one operation inside a single root closes that without restating the deny
+# list 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 +30,19 @@
 # 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
@@ -52,12 +70,27 @@ 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
+# Prints the root class of a charset-safe path token, 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="$1" canon
+  # Globs never auto-allow: bash expands them only after this hook has decided, so realpath
+  # sees the unexpanded pattern and `/tmp/link*` passes before expanding onto a symlink whose
+  # target is outside. chmod and cp follow command-line symlinks, so that is a write to it.
+  case "$t" in *[*?[]*) return 1 ;; esac
+  [ -n "$(printf '%s' "$t" | tr -d 'A-Za-z0-9._/-')" ] && return 1
+  case "$t" in
+    /*) canon=$(realpath -m -- "$t" 2>/dev/null) ;;
+    *)  [ -n "$seg_cwd" ] || return 1
+        canon=$(realpath -m -- "$seg_cwd/$t" 2>/dev/null) ;;
+  esac
+  [ -n "$canon" ] || return 1
+  path_class "$canon"
+}
 
-read -r -a toks <<< "$cmd"
-
-# 0 iff the token is charset-safe and resolves to a path strictly inside /tmp.
+# 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="$1" canon
   # Globs never auto-allow. Bash expands them only after this hook has decided, so realpath
@@ -67,10 +100,6 @@ under_tmp() {
   # because `rm` unlinks the symlink itself rather than following it.
   case "$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.
   case "$t" in /*) ;; *) return 1 ;; esac
   canon=$(realpath -m -- "$t" 2>/dev/null)
   [ -n "$canon" ] || return 1
@@ -79,29 +108,16 @@ 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
       [ "$t" = "--" ] && { end_opts=1; continue; }
@@ -112,13 +128,13 @@ if [ -n "${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 "\`$val\` is outside /tmp"
@@ -136,54 +152,122 @@ if [ -n "${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="$1" takes_mode ok_opts t cls seen_class=""
+  local path_operand=0 seen_mode=0 end_opts=0 i=1
+  # Options are an allowlist per command, so anything that changes how symlinks are followed
+  # defers instead of needing enumeration. `cp -L` / `-H` matter most: they dereference while
+  # recursing, which copies the CONTENT of a symlink target from outside /tmp into a scratch
+  # dir that `Read(/tmp/**)` then exposes. Plain `-r` and `-a` (which implies `-d`) recreate
+  # such a symlink as a symlink instead, so no outside content is materialized.
+  case "$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
+    cls=$(operand_class "$t") || defer "\`$t\` is outside /tmp and not inside a git checkout in \$HOME"
+    # Every operand of one operation stays in one root: see the exfiltration note above.
+    [ -n "$seen_class" ] && [ "$cls" != "$seen_class" ] && defer "\`$t\` puts this $verb across both /tmp and a checkout"
+    seen_class="$cls"
+    path_operand=1
+  done
+
+  [ "$path_operand" = 1 ] || defer "no path operand"
+}
+
+split_segments "$cmd"
+seg_cwd="${cwd:-$PWD}"
+saw_cd=0                                  # a `cd` moved the working directory somewhere
+proved=0                                  # at least one op 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=1
+      continue
+      ;;
+    tar | unzip)
+      check_archive_segment "${SEG_TOKS[0]}"
+      proved=1
+      continue
+      ;;
+    cd)
+      # A `cd` writes nothing, so it never blocks an allow; it only moves where a relative
+      # operand points. Tracking it is only sound while every step stays known and inside the
+      # roots, so an unresolvable destination, or one outside them, makes the working directory
+      # unknown — and a `cd` is only followed from a known one, so a later `cd` cannot walk it
+      # back into a root it has already left.
+      saw_cd=1
+      if [ -n "$seg_cwd" ] && new_cwd=$(apply_cd "$seg_cwd" "${SEG_TOKS[@]:1}") \
+         && path_class "$new_cwd" >/dev/null; then
+        seg_cwd="$new_cwd"
+      else
+        seg_cwd=""
+      fi
+      continue
+      ;;
+  esac
+  # Some other command shares the line. If an `mv` or `chmod` runs inside it after all — behind
+  # a wrapper, an env prefix or a path — this hook cannot say what it writes to.
+  for verb in mv chmod; do
+    segment_runs_verb "$verb" "$seg" && defer "$verb is not the leading command word in \`$seg\`"
+  done
+  only_ours=0
 done
 
-[ "$path_operand" = 1 ] || defer "no path operand"
-decide allow "every path operand is under /tmp"
+[ "$proved" = 1 ] || exit 0
+[ "$only_ours" = 1 ] && decide allow "every path operand is under /tmp"
+exit 0
diff --git a/.claude/hooks/guard-rm-outside-tmp.sh b/.claude/hooks/guard-rm-outside-tmp.sh
index 5aff497d89..703dedc885 100755
--- a/.claude/hooks/guard-rm-outside-tmp.sh
+++ b/.claude/hooks/guard-rm-outside-tmp.sh
@@ -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 $HOME, and the tree's
-# own root folder only when it is a linked worktree (`.git` is a pointer file, so history in
-# the main repo survives); a primary checkout's root (`.git` is a history dir) and any `.git`,
-# `.claude` or `.env` path are never auto-allowed. Globs auto-allow only under /tmp — elsewhere their expansion
+# Which targets those two roots cover, and the tradeoff they rest on, is `path_class` in
+# lib-guarded-verb.sh. Globs auto-allow only under /tmp — elsewhere their expansion
 # could reach `.git` or a dotfile the literal checks never see. Relative operands resolve
-# against the command's cwd (from the hook input).
+# against the working directory the command runs from, which a `cd` in an earlier segment
+# moves; once a `cd` is one this guard cannot resolve, that directory is unknown and a
+# relative operand can no longer be proved.
 #
 # Assumes GNU `realpath` (-m) and `jq`, both present in this repo's Linux dev env.
 set -uo pipefail
@@ -35,87 +38,90 @@ cmd=$(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
-
-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 $seg_cwd. Returns only once every operand is an auto-allowable target;
+# anything it cannot prove defers instead.
+check_rm_segment() {
+  local i=1 t canon had_operand=0 end_opts=0
+  while [ "$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
+    case "$t" in
+      /*) canon=$(realpath -m -- "$t" 2>/dev/null) ;;
+      *)  [ -n "$seg_cwd" ] || defer "\`$t\` is relative to a working directory left by a \`cd\` this guard cannot resolve"
+          canon=$(realpath -m -- "$seg_cwd/$t" 2>/dev/null) ;;
+    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 \`$t\` is outside /tmp" ;; esac ;; esac
+    path_class "$canon" >/dev/null || defer "\`$canon\` is outside /tmp and not inside a git checkout in \$HOME"
   done
-  [ -n "$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:-$PWD}"
+proved=0                                  # at least one `rm` segment came out auto-allowable
+only_ours=1                               # ... and nothing else shares the command line
+
+for seg in "${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 the operands
+      # of a later segment point. Tracking it is only sound while every step stays known and
+      # inside the allowed roots, so an unresolvable destination, or one this guard would not
+      # delete in, makes the working directory unknown — and a `cd` is only followed from a
+      # known one, so a later `cd` cannot walk it back into a root it has already left.
+      if [ -n "$seg_cwd" ] && new_cwd=$(apply_cd "$seg_cwd" "${SEG_TOKS[@]:1}") \
+         && path_class "$new_cwd" >/dev/null; then
+        seg_cwd="$new_cwd"
+      else
+        seg_cwd=""
+      fi
+      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
diff --git a/.claude/hooks/lib-guarded-verb.sh b/.claude/hooks/lib-guarded-verb.sh
index c4d5f5ac9a..dc83a73220 100644
--- a/.claude/hooks/lib-guarded-verb.sh
+++ b/.claude/hooks/lib-guarded-verb.sh
@@ -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 +104,134 @@ 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')"
+}
+
+# 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, no operand at all (`cd` alone is $HOME), or more than one. The caller then
+# treats the working directory as unknown, so that a later relative operand is never resolved
+# against a directory the command has already left.
+apply_cd() {
+  local cwd="$1" t
+  shift
+  [ "$#" -eq 1 ] || return 1
+  t="$1"
+  [ -n "$(printf '%s' "$t" | tr -d 'A-Za-z0-9._/-')" ] && return 1
+  case "$t" in -*) return 1 ;; esac
+  case "$t" in
+    /*) realpath -m -- "$t" 2>/dev/null ;;
+    *) [ -n "$cwd" ] || return 1
+       realpath -m -- "$cwd/$t" 2>/dev/null ;;
+  esac
+}
+
+# Prints the class of a canonical path and returns 0: `tmp` for one strictly under /tmp, `repo`
+# for one strictly inside a git working tree located under $HOME. Fails, printing nothing, for
+# anything else — those are the only two roots the guards are willing to touch unprompted.
+#
+# 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. Never in a class at all: git history, the agent's own
+# guards and settings (removing those is what removes the prompt on everything else), and
+# gitignored `.env` files — the version-control premise holds for none of the three. 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.
+path_class() {
+  local canon="$1" d root=""
+  case "$canon" in /tmp/?*) printf 'tmp'; return 0 ;; esac
+  [ -n "${HOME:-}" ] || return 1
+  case "$canon" in "$HOME"/?*) ;; *) return 1 ;; esac
+  case "$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")
+  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'
+}
+
+# 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
 }
 
diff --git a/.claude/hooks/test-hooks.sh b/.claude/hooks/test-hooks.sh
index 65631b6a48..b225827f2c 100644
--- a/.claude/hooks/test-hooks.sh
+++ b/.claude/hooks/test-hooks.sh
@@ -4,6 +4,10 @@
 # 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)"
@@ -46,7 +50,7 @@ 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; }"
@@ -107,6 +111,20 @@ 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
+
 echo
 echo "== allow-fileops-in-tmp.sh =="
 A=allow-fileops-in-tmp.sh
@@ -117,7 +135,7 @@ 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 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 +147,21 @@ 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 allow "mkdir -p /tmp/x; mv /tmp/a /tmp/x; chmod 755 /tmp/x"
+run $A allow "$(printf 'mv /tmp/a /tmp/b\nchmod 755 /tmp/b')"
+run $A ask   "ls && mv /tmp/a /etc"
+run $A ask   "$(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 $A ask   "cd /etc && mv a b"
+
 echo
 [ "$fails" = 0 ] && echo "ALL PASS" || { echo "$fails FAILURES"; exit 1; }
diff --git a/AGENTS.md b/AGENTS.md
index aab920746b..4096e2fc9c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -146,10 +146,11 @@ $NAV --root backend callees "X"                           # what does X call?
 - **MUST `outline` before `Read`** on unfamiliar files — then `body` or `Read` with offset/limit for specifics
 - **Scratch stays outside the checkout.** Temp scripts, data dumps, cache backups and
   screenshots go in the session scratch directory or `/tmp`, so nothing temporary can end up
-  committed. Write `rm`/`mv`/`cp` as one plain unchained command: a PreToolUse hook
-  auto-allows those when every operand is under `/tmp` or inside this checkout, but it defers
-  on `&&`, `;`, redirects, quotes and `$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 under `/tmp` or inside this checkout, and `mv`/`cp`/`chmod`
+  under `/tmp`. Chaining and line breaks are fine, since every command on the line is proved on
+  its own operands, but a quoted or `$VAR` operand, a redirect, or a wrapper like `xargs rm`
+  cannot be proved, and that deferral is what turns a routine cleanup into a permission prompt.
 - 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