Files
stack/packages/mosaic/framework/tools/git/wrapper-guard.sh
T
Hermes Agent 2a2a87251a
ci/woodpecker/pr/ci Pipeline failed
wrapper-guard: read the command the shell will run, and stop losing the client behind option values
Round-four review, three more absence-driven allows.

1. The guard read the command as TYPED. A backslash before a newline is removed
   before anything else happens, so an endpoint token split across the join
   (`.../iss\` + newline + `ues/1/comments`) executed the comments endpoint while
   the literal token never appeared in the text. Continuations are now joined
   before every check, because the joined form IS the command. This is the same
   defect as the split-across-variables case, minus the excuse: there the token
   genuinely does not exist until the shell expands it, here it was sitting in
   the input the whole time and the guard chose the wrong reading of it.

2. Transparent prefixes take option VALUES. `sudo -u root curl` hid a live write
   because `root` was a word the prefix list did not know. Enumerating option
   grammars per prefix is the wrong game, so what is skipped is an option and at
   most one value for it, plus a bare duration for `timeout` — never an
   arbitrary word. `xargs echo curl ...` therefore stays ALLOWED, because there
   the command is echo and the client is its argument.

3. `find -exec` runs the client. It opens command position the same way an
   operator does, and now reads that way.

All seven reviewer repros are fixtures, each with its counter-case in the
allowed direction: a continuation inside a heredoc document stays a document,
`xargs echo curl` stays allowed, `sudo apt-get install curl` stays allowed, a
prefixed READ stays allowed. 48/48, and the 18-command ordinary sweep still
blocks none.

Gates: sanitization, resident budget, test enumeration, tools-index (self-test
4/4, git suite 100%), prettier.
2026-08-12 17:46:25 -05:00

299 lines
15 KiB
Bash
Executable File

#!/usr/bin/env bash
# wrapper-guard.sh — PreToolUse hook on Bash.
#
# Blocks three specific, mechanically-detectable mistakes that prose has
# repeatedly failed to prevent:
#
# 1. A checkout (git clone / git worktree add) targeting $HOME.
# Root cause of a fleet host's /home filling to 100% — 255 GB, 842 dirs.
#
# 2. A raw provider API WRITE against an endpoint that already has a Mosaic
# wrapper. Constitution gate 7 requires the wrapper; the wrapper knows
# provider dialect, identity, and queue-guard ordering that raw curl does
# not. Reads are untouched — they are how you gather evidence.
#
# 3. The literal review event "APPROVE". Gitea's vocabulary is APPROVED;
# it accepts APPROVE with HTTP 200, silently files the review PENDING,
# and then 422s on submit. This one is unconditionally wrong on Gitea and
# is what a verdict silently failing to land looks like.
#
# Design constraint: this hook must not become something agents route around.
# It blocks WRITES to endpoints with a known wrapper, and nothing else. Raw
# curl for reads, for registry/manifest calls, and for endpoints with no
# wrapper (there are many) all pass untouched.
#
# Break-glass, for a genuine gap where no wrapper can express the call:
# MOSAIC_WRAPPER_OVERRIDE=1 <command>
# Using it means "no wrapper covers this" — if that is wrong, the fix is to
# extend the wrapper, not to keep typing the override.
#
# Exit codes (Claude Code PreToolUse): 0 = allow, 2 = block with message.
set -euo pipefail
INPUT="$(cat)"
CMD="$(printf '%s' "$INPUT" | jq -r '.tool_input.command // empty' 2>/dev/null || true)"
[ -z "$CMD" ] && exit 0
# Read the command the SHELL will run, not the text as typed. A backslash before
# a newline is removed before anything else happens, so
# curl -d@b https://host/api/v1/repos/a/b/iss\
# ues/1/comments
# executes the comments endpoint while the literal token `issues` never appears
# in the text. Every check below — position, URL, body, endpoint — reads the
# joined form, because that is the command.
CMD="$(printf '%s' "$CMD" | sed -e ':a' -e 'N' -e '$!ba' -e 's/\\\n//g')"
# Honour the override only when it is set in the command itself or the env.
case "$CMD" in *MOSAIC_WRAPPER_OVERRIDE=1*) exit 0 ;; esac
[ "${MOSAIC_WRAPPER_OVERRIDE:-0}" = "1" ] && exit 0
# The wrappers this guard points at are its own siblings. Resolving relative to
# this file — rather than to a hardcoded $HOME/.config/mosaic — means the guard
# names the wrappers from the same install it was launched from, and that it
# still works from a repo checkout with no installed mosaic home (which is how it
# is exercised in CI). $HOME remains the fallback for a guard invoked by an
# absolute path from somewhere unusual.
W="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
[ -x "$W/pr-review.sh" ] || W="$HOME/.config/mosaic/tools/git"
# ---- 1. checkout into $HOME ------------------------------------------------
if printf '%s' "$CMD" | grep -Eq 'git[^|;&]*(clone|worktree[[:space:]]+add)'; then
# Any argument that resolves under $HOME and is not under a work filesystem.
if printf '%s' "$CMD" | grep -Eq "(^|[[:space:]=\"'])(~|\\\$HOME|$HOME)/"; then
cat <<EOF
BLOCKED: this checks a repository out under \$HOME.
\$HOME holds configuration, credentials, state and caches. It does not hold
checkouts, worktrees, scratch files, or build output. One fleet host's /home hit
100% (394 G) with 255 GB of agent workspaces accumulated exactly this way.
Use the helper, which derives the path so you do not have to choose one:
~/.config/mosaic/tools/git/mosaic-worktree.sh new <branch> # /src/<repo>-worktrees/<slug>
~/.config/mosaic/tools/git/mosaic-worktree.sh path <branch> # show where it would go
~/.config/mosaic/tools/git/mosaic-worktree.sh rm <branch> # removal is part of the task
Worktrees, not clones: they share the object store, and \`git worktree list\`
makes every one of them enumerable — which is the only reason cleanup can
ever be safe.
EOF
exit 2
fi
fi
# ---- 2/3. provider API writes ---------------------------------------------
# A raw provider write is four things at once: an HTTP client, a URL, a mutating
# verb or a request body, and a path fragment naming an endpoint a wrapper
# already owns. All four are required, which is what keeps reads and unwrapped
# endpoints flowing.
#
# Deliberately NOT gated on the literal "/api/v1/repos/". An independent reviewer
# broke that version in one line: build the path in shell variables
# p=/api/v1/repo; q=s/a/b/pulls/1/reviews; curl -d@body "https://host${p}${q}"
# and the host-anchored literal never appears, so the check read clean while the
# write went through. The endpoint fragments below survive it, because the
# fragment has to appear somewhere for the URL to be constructible at all.
# The client must be at COMMAND POSITION, and that has to be judged against the
# CODE in the command, not against its text. Review caught the text version
# blocking ordinary work:
# grep -R "curl -d https://host/api/v1/repos/a/b/issues" docs/
# echo "curl -d https://host/api/v1/repos/a/b/pulls" > note.txt
# Talking about a call is not making one, and over-blocking is not the safe
# direction: a guard that blocks ordinary work gets switched off, and a guard
# that is off permits everything.
#
# A first fix required the client to follow a shell operator. That lasted until
# the author sent a message quoting one of these fixtures — the quoted text
# contained `... && GITEA_TOKEN=$T curl -d@b .../merge`, so an operator appeared
# INSIDE the quotes and the guard blocked the message. Same defect, one level
# in: an operator inside a string is not an operator.
#
# So the position test runs against a SKELETON — the command with its data spans
# (quoted strings, heredoc bodies) removed. Endpoint, URL and body detection all
# still run against the FULL text, because real calls quote their URLs and a
# skeleton would be blind to them.
#
# The exception is the reason quotes are data at all: if something is about to
# EXECUTE the quoted text, then the quotes hold code and the skeleton is the
# full text again. The first version of this list named only `bash -c`, `sh <<`
# and `eval`, and review immediately produced the spellings it did not know:
# printf '%s\n' 'curl -d@b .../comments' | sh
# cat <<EOF | sh
# sh -s <<EOF
# Each one executes; each one had its client filed away as data. So the test is
# not "which invocation form did I think of" but "does a shell stand between
# this data and execution" — including the pipe, which is the common form.
SHELL_EXECUTES_DATA='(^|[[:space:]|;&(])((ba|z)?sh|dash)([[:space:]]+-[a-z]*[cs]([[:space:]]|$)|[[:space:]]*<<)'
SHELL_EXECUTES_DATA="$SHELL_EXECUTES_DATA"'|(^|[[:space:]|;&(])eval([[:space:]]|$)'
SHELL_EXECUTES_DATA="$SHELL_EXECUTES_DATA"'|\|[[:space:]]*((ba|z)?sh|dash)([[:space:]]|$)'
# That decision is made PER LINE, not for the whole command. The first version
# switched globally, and it blocked its author again within the hour: a script
# whose only shell invocation was an unrelated `docker run ... sh -c 'echo hi'`
# had every OTHER quoted span on every other line promoted to code with it.
# A shell on one line does not execute a string on another, and a guard that
# says otherwise is back to blocking ordinary work.
SKEL="$(printf '%s' "$CMD" | awk -v inv="$SHELL_EXECUTES_DATA" '
function ascode(s) { gsub(/["\047]/, ";", s); return s } # quotes separate
function asdata(s) { gsub(/\047[^\047]*\047/, "", s); gsub(/"[^"]*"/, "", s); return s }
{
# Inside a heredoc: the body is code only if the line that OPENED it fed a
# shell (`cat <<EOF | sh`, `sh -s <<EOF`). Otherwise it is a document.
if (hd != "") { if ($0 == hd) { hd=""; next }
if (hdcode) print ascode($0); next }
if (match($0, /<<-?[[:space:]]*\047?"?[A-Za-z_][A-Za-z0-9_]*/)) {
t = substr($0, RSTART, RLENGTH); sub(/^<<-?[[:space:]]*[\047"]?/, "", t)
hd = t; hdcode = ($0 ~ inv)
}
print ($0 ~ inv) ? ascode($0) : asdata($0)
}')"
# Command position is not "the first word is literally `curl`". A word can sit
# in front of a command without displacing it, and review found four ordinary
# ones hiding a real write: `env GITEA_TOKEN=$T curl`, `command curl`,
# `timeout 10 curl`, `/usr/bin/curl`. The `env` form matters most, because it is
# precisely what an agent reaches for to keep a credential out of the global
# environment — the careful spelling was the invisible one.
CMD_PREFIX='([A-Za-z_][A-Za-z0-9_]*=[^[:space:]]*|env|command|builtin|exec|nohup|setsid|stdbuf|nice|ionice|sudo|doas|xargs|time|timeout|watch)[[:space:]]+'
# Their options take values: `sudo -u root curl` hid a live write because `root`
# was a word the list did not know. What is skipped is an OPTION and at most one
# value for it — not any word — so `xargs echo curl ...` stays allowed, because
# there `echo` is the command and curl is its argument. Plus a bare duration,
# which is `timeout`'s operand.
PREFIX_OPT='-[^[:space:]]+[[:space:]]+([^-][^[:space:]|;&(){}<>]*[[:space:]]+)?'
PREFIX_ARG='('"$PREFIX_OPT"'|[0-9]+[smhd]?[[:space:]]+)'
# A path in front of the client is still the client.
CLIENT='([^[:space:]]*/)?(curl|wget|httpie|http)'
# `find -exec` opens command position the same way an operator does.
CLIENT_AT_CMD_POS='(^|[;&|(){}]|`|\$\(|-execdir|-exec)[[:space:]]*('"$CMD_PREFIX"'('"$CMD_PREFIX"'|'"$PREFIX_ARG"')*)?'"$CLIENT"'([[:space:]]|$)'
if printf '%s' "$SKEL" | grep -Eq "$CLIENT_AT_CMD_POS" \
&& printf '%s' "$CMD" | grep -Eq 'https?://'; then
# Write detection. Every spelling curl accepts, because the guard is defeated
# by the one spelling it does not know: `-d@body` (no space) and
# `--request=POST` (equals form) both slipped past the first version.
is_write=0
printf '%s' "$CMD" | grep -Eq -- \
'-X[[:space:]]*(POST|PATCH|PUT|DELETE)|--request[[:space:]=]*(POST|PATCH|PUT|DELETE)' && is_write=1
# curl sends POST implicitly when handed a body, in any of these forms.
printf '%s' "$CMD" | grep -Eq -- \
'(^|[[:space:]])(-d|-F|-T)|--data([-a-z]*)?[[:space:]=]|--json[[:space:]=]|--form|--upload-file' && is_write=1
if [ "$is_write" -eq 1 ]; then
endpoint=""; wrapper=""
case "$CMD" in
*"/pulls/"*"/reviews"*|*"/pulls/"*"/requested_reviewers"*)
endpoint="pull-request review"; wrapper="pr-review.sh" ;;
*"/pulls/"*"/merge"*) endpoint="pull-request merge"; wrapper="pr-merge.sh" ;;
*"/issues/"*"/comments"*) endpoint="issue comment"; wrapper="issue-comment.sh" ;;
*"/pulls"*) endpoint="pull request"; wrapper="pr-create.sh" ;;
*"/issues"*) endpoint="issue"; wrapper="issue-create.sh" ;;
*"/milestones"*) endpoint="milestone"; wrapper="milestone-create.sh" ;;
esac
# A URL the guard cannot READ is a URL the guard must not CLEAR.
#
# Round one fixed one spelling of this and review immediately produced the
# general form: split the endpoint token itself across two variables —
# a=/api/v1/repos/o/r/iss; b=ues/1/comments
# curl -d@body "https://host${a}${b}"
# — and no fragment above ever appears contiguously. Chasing that with more
# fragments is unwinnable: the endpoint does not exist until the shell
# expands it, and this hook runs before that.
#
# So stop pretending to read it. If a write's URL contains an expansion,
# the guard has no endpoint to judge, and "no endpoint" must not mean
# "allowed" — that is the same absence-driven allow as the missing-wrapper
# case, wearing different clothes.
#
# Scoped to commands that are visibly forge-shaped, so an opaque webhook or
# artifact POST is untouched. A caller who splits `/api/` and the hostname
# as well does get through; that is no longer a mistake anyone makes by
# accident, and this hook stops mistakes. It is not a sandbox, and pretending
# otherwise is how you get a control nobody can trust the boundaries of.
if [ -z "$endpoint" ] \
&& printf '%s' "$CMD" | grep -Eq 'https?://[^[:space:]"'"'"'|;&)]*[$`]' \
&& printf '%s' "$CMD" | grep -Eq '/api/v[0-9]|/repos/|git\.|gitea|github\.com|gitlab|forgejo'; then
cat <<EOF
BLOCKED: raw provider API write whose URL this guard cannot read.
The URL is assembled from shell expansions, so the endpoint it names does not
exist until the shell builds it — after this check runs. The guard cannot tell
whether it is a wrapped endpoint, and an unreadable URL is not a cleared one.
$W/ <- the wrappers; use the one for the endpoint you are calling
If you are calling a wrapped endpoint (reviews, merges, comments, pulls,
issues, milestones), use the wrapper — it also resolves identity explicitly,
which matters on a host whose default provider login is an admin account.
If this is genuinely not a provider endpoint, either write the URL literally so
the guard can see what it is, or prefix MOSAIC_WRAPPER_OVERRIDE=1.
EOF
exit 2
fi
# Block on the ENDPOINT, never on whether the wrapper file happens to exist.
# The previous version required `[ -x "$W/$wrapper" ]`, which meant a host
# with a broken or absent install allowed exactly the raw writes the guard
# exists to stop — an absence-driven allow, and the second one found in this
# file. A missing wrapper is a broken install; it is not a licence to bypass
# gate 7. Say so, and say which is which.
if [ -n "$endpoint" ]; then
if [ -x "$W/$wrapper" ]; then
remedy="Use the wrapper the Constitution (gate 7) requires:
$W/$wrapper
Run \`$wrapper --help\` for the flags."
else
remedy="The wrapper that covers this endpoint is \`$wrapper\`, and it is NOT
present or not executable at:
$W/$wrapper
That is a broken or incomplete install, not permission to send the call raw.
Repair the install (\`mosaic doctor\`) and use the wrapper."
fi
cat <<EOF
BLOCKED: raw provider API write to the $endpoint endpoint.
$remedy
The wrappers are not a formality. They carry provider-dialect differences that
raw curl silently gets wrong — Gitea's review event is APPROVED, GitHub's is
APPROVE, and Gitea accepts the wrong one with HTTP 200 while filing the review
as PENDING. They also resolve identity explicitly, which matters on a host
where the default login is an admin account.
If no wrapper flag can express this call, that is a wrapper gap: extend the
wrapper. To proceed anyway for a genuine gap, prefix MOSAIC_WRAPPER_OVERRIDE=1.
EOF
exit 2
fi
fi
fi
# ---- 3. the APPROVE/APPROVED trap, wherever it appears ---------------------
if printf '%s' "$CMD" | grep -Eq '"event"[[:space:]]*:[[:space:]]*"APPROVE"'; then
cat <<EOF
BLOCKED: review event "APPROVE" is not valid on Gitea.
Gitea's vocabulary is "APPROVED". It accepts "APPROVE" with HTTP 200, silently
files the review as PENDING, and then fails the submit endpoint with
422 "review stay pending" — so the verdict looks placed and is not.
("REQUEST_CHANGES" is spelled identically on both providers; only the approve
path carries this trap.)
Use $W/pr-review.sh, which sends the correct token for the detected provider.
Whatever you use, re-read GET /pulls/{n}/reviews and assert state==APPROVED
before reporting a verdict placed.
EOF
exit 2
fi
exit 0