test: run the code under test, not grep over its source

The suite is rewritten against the sealed harness: a fresh $HOME per case, a
recorded call log per box, and assertions about what dotup DID rather than
about which words appear in the file.

ptydrive.py drives a command on a real pty from a small expect/send script, so
the private tier's password prompt, its retry loop, the token file and the
clone can be exercised at all — `expect` is not installed everywhere and
python3 is.

New guard for DU-C1: `dotup pick` is actually run, on a pty, against an fzf
fake that dumps its argv one argument per line. It must exit 0, hand fzf every
--bind the source writes (counted against the source, not a hard-coded 7), bind
all seven keys, and leave the completed-pick marker behind. Every other test
here greps the source for bind strings, which is exactly why all of them passed
while the picker was dead. A static check alongside it rejects any comment
sitting between continued lines, anywhere in the file.
This commit is contained in:
bcherb2
2026-08-21 22:34:39 -04:00
parent 042320208e
commit 8448ff551a
2 changed files with 666 additions and 84 deletions
+543 -84
View File
@@ -60,6 +60,52 @@ missed(){ awk -F'\t' -v f="$1" '!/^#/ && !/^@/ && NF>=3 && $3==f {print $1"/"$2}
| while read -r k; do grep -qxF "$k" state/selected || echo "$k"; done; :; }
reset() { rm -rf state; sh $D preset defaults; }
# ---------------------------------------------------------------- the seal ---
# The old harness put .tests/fakebin at the FRONT of $PATH. That fails OPEN: a
# package manager with no fake fell through to the real one, and an audit of
# this suite ran the host's actual `brew`. Everything below runs under a SEALED
# PATH instead (see seal.sh) with `env -i`, so the run inherits nothing at all
# and a tool nobody thought to fake is command-not-found rather than the
# machine's own copy.
SEAL=${TMPDIR:-/tmp}/dotup-seal.$$
SEALSUDO=${TMPDIR:-/tmp}/dotup-seal-sudo.$$
sh ./seal.sh "$SEAL" >/dev/null
sh ./seal.sh "$SEALSUDO" --with sudo >/dev/null
BOXN=0; SB=; LOG=; BOXENV=; BOXPATH=$SEAL; PW=
cleanup_boxes() { rm -rf "$SEAL" "$SEALSUDO" ${TMPDIR:-/tmp}/dotup-box."$$".* \
${TMPDIR:-/tmp}/dotup-sealtest."$$" ; }
trap cleanup_boxes EXIT INT TERM
# A fresh $HOME per case, because find_tool probes ~/.local/bin, ~/bin and
# ~/.npm-global/bin: a run against the developer's real $HOME is a run against
# whatever the developer happens to have installed.
newbox() {
BOXN=$((BOXN+1)); SB=${TMPDIR:-/tmp}/dotup-box.$$.$BOXN
rm -rf "$SB"; mkdir -p "$SB/.config/dotfiles" "$SB/tmp" "$SB/.local/bin"
: > "$SB/.config/dotfiles/expanded"; : > "$SB/.config/dotfiles/selected"
LOG=$SB/calls.log; : > "$LOG"; BOXENV=; BOXPATH=$SEAL
}
# $BOXENV is deliberately unquoted: it is a list of NAME=VALUE words to add.
box() { env -i HOME="$SB" PATH="$BOXPATH" TERM=dumb LANG=C TMPDIR="$SB/tmp" \
DOTUP_STATE="$SB/.config/dotfiles" DOTUP_MANIFEST="$M" \
DOTUP_TEST_LOG="$LOG" $BOXENV "$BOXPATH/sh" $D "$@"; }
boxrc() { c=0; box "$@" >/dev/null 2>&1 || c=$?; echo "$c"; }
pick() { printf '%s\n' "$@" > "$SB/.config/dotfiles/selected"; }
# A seal with a hole in it, for the cases whose whole point is that a tool is
# NOT there. Refuses if the host would satisfy the lookup anyway through one of
# find_tool's four absolute probes -- an absence that is not real would make the
# assertion pass for the wrong reason.
holed() {
sh ./seal.sh "$SB/seal" --without "$@" >/dev/null
BOXPATH=$SB/seal
for h in "$@"; do
grep -q "^$h " "$SB/seal/.shadowed" 2>/dev/null && return 1
done
return 0
}
# Everything a run wrote, colour stripped.
plain() { sed 's/\x1b\[[0-9;]*m//g'; }
printf '\n\033[1mtoggle rule\033[0m\n'
reset
is "package off -> group goes partial" "2/3" \
@@ -115,12 +161,6 @@ is "bws-secrets stays a non-package" "-" \
"$(awk -F'\t' '$1=="private" && $2=="bws-secrets" {print $4}' "$M")"
is "no manifest row installs bws" "0" \
"$(awk -F'\t' '!/^[#@]/ && NF>=3 && $2=="bws" {c++} END{print c+0}' "$M")"
is "the private tier installs bws itself" "1" \
"$(grep -c '^ ensure_bws$' $D)"
is "…only after a token exists to use" "1" \
"$(grep -A1 'bws token written, mode 600' $D | grep -c ensure_bws)"
is "bws is checksum-verified" "1" \
"$(grep -c 'bws checksum mismatch' $D)"
printf '\n\033[1mcredentials never reach a log\033[0m\n'
# `run` echoes its whole argv to stderr. The private init's argv ends in
@@ -136,12 +176,6 @@ is "both traced URLs are redacted" "2" \
# publishes it to /proc/<pid>/cmdline, which every account on the box can read
# for as long as the clone runs, and then into the clone's .git/config, which
# keeps it. Only the split-out p_clean may reach an argv.
is "the credential-bearing URL never reaches an argv" "0" \
"$(grep -c 'chezmoi init.*\$p_repo' $D)"
is "the token goes to a credential file instead" "1" \
"$(grep -c "umask 077; printf '%s\\\\n' \"\\\$p_cred\" > \"\\\$PRIV_CRED\"" $D)"
is "…at mode 600" "1" \
"$(grep -c 'chmod 600 "\$PRIV_CRED"' $D)"
# Exercise the real implementation lifted straight out of dotup. A copy of the
# sed expression here would keep passing after someone edited the original.
split() { p_repo=$1; eval "$(sed -n '/p_cred=\$(printf/p;/p_clean=\$(printf/p' $D)"
@@ -188,12 +222,6 @@ is "...and leaves a credential-free URL alone" "https://git.example.com/x.git" \
"$(rd 'https://git.example.com/x.git')"
# Two tiers, two configs. Sharing one meant the public installer silently ate
# the private tier's seven promptStringOnce answers.
is "private init carries its own -c" "1" \
"$(grep -c '\-c "\$PRIV_CFG"' $D)"
is "PRIV_CFG is not the default config path" "0" \
"$(grep -c 'PRIV_CFG=.*chezmoi/chezmoi.toml' $D)"
is "the private source is locked down after clone" "1" \
"$(grep -c 'chmod -R go-rwx "\$PRIV_SRC"' $D)"
printf '\n\033[1mchezmoi is not assumed to be on PATH\033[0m\n'
# dotup ARRIVES via chezmoi, so "it must be here already" is the natural
@@ -203,10 +231,6 @@ printf '\n\033[1mchezmoi is not assumed to be on PATH\033[0m\n'
# with `chezmoi: not found` AFTER writing the bws token.
is "nothing calls chezmoi by bare name" "0" \
"$(grep -cE '^[[:space:]]*chezmoi (init|apply|update)' $D)"
is "the init uses the resolved path" "1" \
"$(grep -cF '"$CHEZMOI" init --apply' $D)"
is "find_tool looks in ~/bin too" "1" \
"$(grep -cF '"$HOME/bin/$1"' $D)"
# Resolution happens before the prompt: discovering it afterwards means the
# password is spent and the token is already on disk.
is "chezmoi is resolved before the password is asked for" "yes" \
@@ -216,20 +240,6 @@ printf '\n\033[1ma wrong password is not a reinstall\033[0m\n'
# The endpoint credentials are asked for at the very END of a run, after every
# package is installed. Any non-200 used to be fatal, so one mistyped character
# meant repeating the whole install to get back to the prompt.
is "the prompt retries instead of returning" "0" \
"$(grep -c 'endpoint refused the credentials' $D)"
is "…up to a bounded number of attempts" "1" \
"$(grep -c '"$p_try" -gt 5' $D)"
is "…and points at the cheap way back in" "1" \
"$(grep -c 'Re-run only this step: dotup private' $D)"
# One message for 401, 404 and an unreachable host is how a URL-shape bug spends
# an evening looking like a password problem. Each needs a different next move.
is "a 401 names the password" "1" "$(grep -c 'wrong username or password' $D)"
is "a 404 names the route" "1" "$(grep -c 'no bootstrap.env is there' $D)"
is "an unreachable host says so" "1" "$(grep -c 'could not reach that address' $D)"
# --fail is gone, so the code must be checked explicitly or an error page would
# be parsed as the credential blob.
is "the blob is used only on 200" "1" "$(grep -cF '200) break ;;' $D)"
# The subcommand existed but was absent from --help, so there was no way to
# discover the recovery path.
is "dotup private is documented" "1" \
@@ -237,6 +247,15 @@ is "dotup private is documented" "1" \
printf '\n\033[1mrisk model — the invariants that matter\033[0m\n'
reset
# Positive controls FIRST. `picked` and `missed` both end in `; :` so a failed
# grep cannot trip `set -e` -- which also means a broken helper returns nothing
# and is indistinguishable from a clean result. These three invariants are the
# whole safety story, and all three "passed" by producing no output. Prove the
# helpers can still speak before trusting their silence.
is "control: picked can see a safe row" "core/ripgrep" \
"$(picked safe | grep -x 'core/ripgrep' || echo MISSING)"
is "control: missed can see an unticked row" "gpu/nvidia-driver" \
"$(missed invasive | grep -x 'gpu/nvidia-driver' || echo MISSING)"
is "defaults tick nothing invasive" "" "$(picked invasive)"
is "defaults tick nothing private" "" "$(picked private)"
is "defaults tick every safe package" "" "$(missed safe)"
@@ -324,7 +343,11 @@ is "no apt package falls through to brew" "brew lazygit" "$(r
is "tap-only pkg falls through to brew" "brew can1357/tap/omp" "$(rs agents/omp)"
is "special channel carries its @spec" "npm @openai/codex" "$(rs agents/codex)"
is "…and the scoped mermaid name" "npm @mermaid-js/mermaid-cli" "$(rs core/mermaid-cli)"
is "deb channel carries a source" "deb gh:mkasberg/ghostty-ubuntu:_amd64.deb" "$(rs apps/ghostty)"
# %a and %v are expanded at install time from `dpkg --print-architecture` and
# /etc/os-release. The literal `_amd64.deb` that used to be here matched no
# asset ghostty-ubuntu has ever published -- they are named
# ghostty_1.3.1-0.ppa2_amd64_24.04.deb -- so apps/ghostty could never install.
is "deb channel carries a source" "deb gh:mkasberg/ghostty-ubuntu:_%a_%v.deb" "$(rs apps/ghostty)"
is "snap channel renames to the binary" "snap bw" "$(rs core/bitwarden-cli)"
is "tarball is dispatched, not named" "tarball neovim" "$(rs core/neovim)"
is "an unknown key says so" "missing" "$(rs nope/nope | tr -d ' ')"
@@ -342,19 +365,37 @@ is "darwin: no apt package is not a reason to try apt" "unavailable" \
# Nothing in the default set may be unresolvable — that is a manifest bug, and
# it is silent until someone runs the installer on a fresh machine.
reset
# This one passed if `dotup plan` crashed: no output, no unresolved lines, green.
is "control: plan produces a table at all" "yes" \
"$(sh $D plan 2>/dev/null | grep -q 'packages' && echo yes || echo no)"
is "every default package resolves" "" \
"$(sh $D plan 2>/dev/null | sed -n '/no source on this platform/,$p' | grep -v 'no source' || true)"
printf '\n\033[1mthe installer — driven against fake package managers\033[0m\n'
FAKE=$PWD/fakebin
export DOTUP_TEST_LOG=$PWD/state/calls.log
# --print resolves everything and must call nothing at all. The fakes are first
# on PATH, so any call whatsoever leaves a trace.
reset
: > "$DOTUP_TEST_LOG"
out=$(PATH="$FAKE:$PATH" sh $D --unattended --print 2>&1 || true)
is "--print calls no package manager" "0" "$(grep -c . "$DOTUP_TEST_LOG" || true)"
printf '\n\033[1mthe installer — driven against a SEALED set of fakes\033[0m\n'
# Sealed, not prepended. This section keeps `sudo` in the seal because two of
# its assertions are about what is and is not run through it; the failure
# injection sections further down deliberately leave sudo out, so that apt-get
# is invoked directly and can be made to fail.
newbox; BOXPATH=$SEALSUDO
box preset defaults >/dev/null
: > "$LOG"
out=$(box --unattended --print 2>&1 || true)
# "Calls nothing at all" stopped being the right bar when resolution began
# asking the machine what it is: `dpkg --print-architecture` and `apt-cache
# policy` are questions. What a dry run must never do is CHANGE anything.
# "a dry run never invents a failure" lives in .tests/lab/scenarios/, NOT here,
# and the reason is worth recording. The bug was core/brew probing `have brew`
# after an install step that --print had only printed; the check then reported
# "brew still not found after installing it" for a run that installed nothing.
# It cannot be reproduced on a developer box: `find_tool` probes
# /home/linuxbrew/.linuxbrew/bin by ABSOLUTE path, this host has brew there, so
# `have brew` succeeds and the failure never happens. No $PATH can seal an
# absolute probe -- seal.sh says so in its .shadowed list. An assertion written
# here would have passed forever without ever being able to fail.
is "--print installs nothing" "" \
"$(grep -E ' (install|remove|upgrade) ' "$LOG" || true)"
is "…and does not so much as refresh a package list" "" \
"$(grep -E 'apt-get update' "$LOG" || true)"
has "--print still shows the apt batch" "apt-get install -y" "$out"
has "--print shows the npm batch" "npm install -g" "$out"
hasnt "--print never reaches the private stage" "Bootstrap URL:" "$out"
@@ -362,21 +403,11 @@ hasnt "--print never reaches the private stage" "Bootstrap URL:" "$out"
# Now the real engine, over a subset chosen to exercise every non-destructive
# channel. Deliberately not the whole default set: the tarball and script
# handlers write outside $HOME, and a test suite has no business doing that.
reset
sh $D preset none
cat > state/selected <<'EOF'
agents/omp
agents/specify-cli
apps/obsidian
core/bitwarden-cli
core/gh
core/lazygit
core/mermaid-cli
core/ripgrep
EOF
: > "$DOTUP_TEST_LOG"
PATH="$FAKE:$PATH" sh $D install >/dev/null 2>&1 || true
log=$(cat "$DOTUP_TEST_LOG")
newbox; BOXPATH=$SEALSUDO
pick agents/omp agents/specify-cli apps/obsidian core/bitwarden-cli \
core/gh core/lazygit core/mermaid-cli core/ripgrep docker/docker-ce
box install >/dev/null 2>&1 || true
log=$(cat "$LOG")
has "apt batches its packages in one call" "apt-get install -y ripgrep" "$log"
has "brew gets the tap-only package" "brew install can1357/tap/omp" "$log"
has "brew gets the apt-less package" "brew install lazygit" "$log"
@@ -386,14 +417,24 @@ has "snap gets the renamed binary" "snap install bw"
has "flatpak gets the app id" "flatpak install" "$log"
has "…with the real obsidian app id" "md.obsidian.Obsidian" "$log"
# The one that matters on Linux: apt not knowing a name is not a dead end.
has "apt probes before it batches" "apt-cache show gh" "$log"
# `apt-cache show` is not an existence test -- it exits 0 for a name that has
# no installation candidate. The probe asks for the candidate now, so the
# assertion follows it.
has "apt probes before it batches" "apt-cache policy gh" "$log"
# docker-ce is here for one reason: it is in the apt cache but has no
# installation candidate, and `apt-cache show` exits 0 for exactly that shape.
# Under the old probe it stayed in the batch and apt refused all of them at
# once -- one unavailable name taking thirty installable ones down with it.
has "…and rejects a name with no candidate" "apt-cache policy docker-ce" "$log"
hasnt "…leaving it out of the apt batch" "apt-get install -y ripgrep docker-ce" "$log"
has "…while the rest of the batch survives" "apt-get install -y ripgrep" "$log"
has "an apt name apt rejects moves to brew" "brew install gh" "$log"
hasnt "…and is not left in the apt batch" "apt-get install -y gh" "$log"
# npm cannot run before node, uv cannot run before uv. The pipeline is fixed
# rather than sorted, so assert the order it actually produces.
is "apt runs before npm" "yes" \
"$(a=$(grep -n 'apt-get install' "$DOTUP_TEST_LOG" | head -1 | cut -d: -f1)
b=$(grep -n 'npm install' "$DOTUP_TEST_LOG" | head -1 | cut -d: -f1)
"$(a=$(grep -n 'apt-get install' "$LOG" | head -1 | cut -d: -f1)
b=$(grep -n 'npm install' "$LOG" | head -1 | cut -d: -f1)
[ -n "$a" ] && [ -n "$b" ] && [ "$a" -lt "$b" ] && echo yes || echo no)"
# Nothing runs as root that does not have to.
hasnt "brew is never run through sudo" "sudo brew" "$log"
@@ -401,16 +442,10 @@ hasnt "brew is never run through sudo" "sudo brew"
printf '\n\033[1munattended — the boundary holds because of what is missing\033[0m\n'
# A stale state file is the adversary here: it ticks the two things an
# unattended run must never act on, and the run has to refuse both anyway.
rm -rf state; mkdir -p state
cat > state/selected <<'EOF'
core/ripgrep
docker/docker-ce
networking/openssh-server
private/bws-secrets
private/private-repo
EOF
: > "$DOTUP_TEST_LOG"
out=$(PATH="$FAKE:$PATH" sh $D --unattended --print 2>&1 || true)
newbox; BOXPATH=$SEALSUDO
pick core/ripgrep docker/docker-ce networking/openssh-server \
private/bws-secrets private/private-repo
out=$(box --unattended --print 2>&1 || true)
hasnt "unattended never installs a private row" "bws-secrets" "$out"
hasnt "unattended never prompts for a password" "Password:" "$out"
hasnt "unattended refuses a stale invasive tick" "docker-ce" "$out"
@@ -418,21 +453,20 @@ hasnt "…including a listening ssh port" "openssh-server" "$out"
has "…but still installs the safe defaults" "ripgrep" "$out"
# Determinism: the same command twice, on the same machine, means the same
# thing. A state file left by an interactive run must not change it.
a=$(PATH="$FAKE:$PATH" sh $D --unattended --print 2>/dev/null || true)
sh $D preset none
b=$(PATH="$FAKE:$PATH" sh $D --unattended --print 2>/dev/null || true)
a=$(box --unattended --print 2>/dev/null || true)
box preset none >/dev/null
b=$(box --unattended --print 2>/dev/null || true)
is "unattended is computed, not inherited" "same" \
"$([ "$a" = "$b" ] && echo same || echo different)"
# `set -e` would kill the subshell at the failing command, so the status is
# captured through a || branch rather than read from $? afterwards.
rc() { c=0; "$@" >/dev/null 2>&1 || c=$?; echo "$c"; }
is "unattended exits clean when nothing fails" "0" \
"$(PATH="$FAKE:$PATH" rc sh $D --unattended --print)"
is "unattended exits clean when nothing fails" "0" "$(boxrc --unattended --print)"
is "an unknown flag is refused" "2" "$(rc sh $D --nonsense)"
# The negative test in PLAN.md phase 3 runs `zsh -ic exit` after an unattended
# run. That only passes if zsh is in the unattended set.
has "unattended installs the shell it configures" "zsh" \
"$(reset; PATH="$FAKE:$PATH" sh $D --unattended --print 2>&1 | grep 'apt-get install' || true)"
"$(box --unattended --print 2>&1 | grep 'apt-get install' || true)"
printf '\n\033[1mruns where it has to run\033[0m\n'
# A machine with no sudo and no root is a real case -- a locked-down work box, a
@@ -455,15 +489,440 @@ is "runs as a non-root user with no sudo on PATH" "0" \
"$(c=0; env -i HOME="$HOME" PATH="$MIN" DOTUP_STATE="$DOTUP_STATE" DOTUP_MANIFEST="$M" \
"$MIN/sh" $D --unattended --print >/dev/null 2>&1 || c=$?
echo "$c")"
is "…and still resolves the whole default set" "1" \
"$(env -i HOME="$HOME" PATH="$MIN" DOTUP_STATE="$DOTUP_STATE" DOTUP_MANIFEST="$M" \
"$MIN/sh" $D --unattended --print 2>/dev/null | grep -c 'apt-get install -y ')"
# Two apt calls, and the distinction is the point: the manifest's packages go
# in ONE batch, and core/brew's prerequisites are a separate, named call. This
# used to assert "exactly 1" and would now read 2 with no way to tell a
# regression (the batch split apart) from the intended second call.
minprint=$(env -i HOME="$HOME" PATH="$MIN" DOTUP_STATE="$DOTUP_STATE" DOTUP_MANIFEST="$M" \
"$MIN/sh" $D --unattended --print 2>/dev/null)
# Discriminate on the exact prerequisite list, not on `build-essential` --
# that is itself a manifest package, so it appears in BOTH lines and a
# `grep -v` for it excluded the very batch this is meant to count.
is "…and still resolves the default set in ONE apt batch" "1" \
"$(printf '%s\n' "$minprint" | grep 'apt-get install -y ' \
| grep -vc 'build-essential procps curl file git$')"
is "…with brew's prerequisites as their own call" "1" \
"$(printf '%s\n' "$minprint" | grep -c 'build-essential procps curl file git$')"
unset minprint
rm -rf "$MIN"
# `sudo` must not be glued onto anything that is not a system package manager.
is "sudo is only ever used for apt, snap and dpkg" "" \
"$(grep -oE '\$\{SUDO:\+\$SUDO \}[a-z-]+' $D | sed 's/.*}//' | sort -u \
| grep -vE '^(apt-get|snap|rm|mkdir|tar|ln)$' || true)"
printf '\n\033[1mthe seal itself — a missing fake is an error, not the host tool\033[0m\n'
SB1=${TMPDIR:-/tmp}/dotup-sealtest.$$; rm -rf "$SB1"; mkdir -p "$SB1"
# The reason every section below can be believed. Under the old harness the
# fakes were PREPENDED to $PATH, so a package manager nobody had faked fell
# through to the real one; an audit of this suite invoked the developer's own
# `brew`. `env -i` plus a PATH that is only the seal makes that impossible.
is "the seal supplies the package managers" "$SEAL/brew" \
"$(env -i PATH="$SEAL" "$SEAL/sh" -c 'command -v brew')"
is "…and a tool nobody faked is simply absent" "" \
"$(env -i PATH="$SEAL" "$SEAL/sh" -c 'command -v wget || true')"
is "…so nothing in a sealed run can reach the host's own copy" "" \
"$(env -i PATH="$SEAL" "$SEAL/sh" -c 'command -v brew npm curl unzip' | grep -v "^$SEAL/" || true)"
# find_tool ALSO probes four absolute directories, which no PATH can seal. The
# seal ships a fake for every name that would be shadowed there -- `command -v`
# is consulted first, so the fake wins -- and records the rest, because a case
# whose point is that a tool is absent must not quietly pass on a box where it
# is not. seal.sh refuses to build a hole it cannot honour.
if [ -s "$SEAL/.shadowed" ]; then
printf ' \033[33mnote\033[0m this host also carries %sat find_tool'"'"'s absolute probes\n' \
"$(cut -f1 "$SEAL/.shadowed" | sort -u | tr '\n' ' ')"
fi
# A new `have foo` with no fake would silently reopen the hole. seal.sh refuses
# to build such a seal at all, so the failure lands on the harness rather than
# on an assertion that quietly starts measuring the host.
is "an unfaked probe refuses to build a seal" "1" \
"$(sed 's|^have() {.*|&\n\thave nosuchtool|' $D > "$SB1/doctored"
c=0; DOTUP_SEAL_TARGET=$SB1/doctored sh ./seal.sh "$SB1/seal" >/dev/null 2>&1 || c=$?
echo "$c")"
is "…naming the tool it has never heard of" "1" \
"$(DOTUP_SEAL_TARGET=$SB1/doctored sh ./seal.sh "$SB1/seal" 2>&1 >/dev/null \
| grep -c nosuchtool || true)"
printf '\n\033[1mprivate is never a package\033[0m\n'
# The invariant that makes --unattended safe. selected_packages drops every row
# flagged private, so a ticked private row cannot become an install command --
# and the plan, which is what the installer reads, is where that has to show.
newbox
pick core/ripgrep private/bws-secrets private/private-repo
plan=$(box plan 2>&1 | plain)
hasnt "no private row reaches the plan" "private/" "$plan"
has "…while the ordinary row does" "core/ripgrep" "$plan"
out=$(box install 2>&1 | plain || true)
hasnt "…nor the installer's own accounting" "private/" "$out"
has "…and the ordinary one still installs" "apt-get install -y ripgrep" "$(cat "$LOG")"
# A private row is not merely unresolvable -- it is not offered at all. Ticking
# only private rows must produce an empty plan rather than two failures.
newbox; pick private/bws-secrets private/private-repo
is "a private-only selection plans nothing" "0 packages" \
"$(box plan 2>&1 | plain | grep -o '[0-9]* packages' | head -1)"
printf '\n\033[1munattended refuses invasive rows a stale state file still ticks\033[0m\n'
# `dotup --unattended` recomputes the selection first, which HIDES this filter:
# cmd_install never sees a stale tick because cmd_run overwrote it a moment
# earlier. `dotup --unattended install` is the same installer with the state
# file left alone -- which is what a cron job re-running a saved selection is.
newbox; pick core/ripgrep docker/docker-ce networking/openssh-server
out=$(box --unattended install 2>&1 | plain || true)
log=$(cat "$LOG")
has "it names what it refused" "refusing invasive packages" "$out"
hasnt "the daemon never reaches a package manager" "docker-ce" "$log"
hasnt "…nor does the listening ssh port" "openssh-server" "$log"
has "…and the safe package still installs" "ripgrep" "$log"
# Conditional on the flag, or it is a broken installer rather than a boundary.
newbox; pick core/ripgrep docker/docker-ce
box install >/dev/null 2>&1 || true
has "with a human present the same tick does install" "docker-ce" "$(cat "$LOG")"
printf '\n\033[1ma package that fails is reported, and the run says so\033[0m\n'
# Reachable only because the fakes can be made to fail. Every assertion here
# survived deleting the code it is about, for want of a way to make brew lose.
newbox; BOXENV=FAKE_FAIL=brew
pick core/lazygit agents/omp core/ripgrep
out=$(box install 2>&1 | plain || true)
has "the failure is named by its manifest key" "core/lazygit" "$out"
has "…and by the second one too" "agents/omp" "$out"
has "…with what went wrong" "brew install failed" "$out"
has "…under a heading you can find" "did not install" "$out"
has "…and a count at the end" "package(s) did not install" "$out"
is "the run exits non-zero" "1" "$(boxrc install)"
has "…and the packages that CAN install still do" "apt-get install -y ripgrep" "$(cat "$LOG")"
newbox; pick core/ripgrep
is "a run with nothing failing exits 0" "0" "$(boxrc install)"
printf '\n\033[1mone unknown name does not sink the batch\033[0m\n'
# Real apt refuses the WHOLE batch when one name is unusable. The fake fails in
# exactly that shape -- any call carrying more than one package -- so the only
# way through is the one-at-a-time retry, and the only way to see the retry is
# to make the batch lose.
newbox; BOXENV=FAKE_FAIL=apt-get:batch
pick core/ripgrep core/btop core/htop
out=$(box install 2>&1 | plain || true)
is "the batch is attempted first" "1" \
"$(grep -cx 'apt-get install -y btop htop ripgrep' "$LOG" || true)"
is "…then btop on its own" "1" "$(grep -cx 'apt-get install -y btop' "$LOG" || true)"
is "…and htop" "1" "$(grep -cx 'apt-get install -y htop' "$LOG" || true)"
is "…and ripgrep" "1" "$(grep -cx 'apt-get install -y ripgrep' "$LOG" || true)"
has "it says why it is retrying" "retrying one at a time" "$out"
hasnt "…and nothing is left failed" "did not install" "$out"
is "…so the run still exits 0" "0" "$(boxrc install)"
newbox; BOXENV=FAKE_FAIL=npm:batch
pick agents/codex agents/pi core/mermaid-cli
out=$(box install 2>&1 | plain || true)
is "npm batches first" "1" \
"$(grep -c 'npm install -g @earendil-works/pi-coding-agent @mermaid-js/mermaid-cli @openai/codex' "$LOG" || true)"
is "…then @openai/codex on its own" "1" "$(grep -cx 'npm install -g @openai/codex' "$LOG" || true)"
is "…and the scoped mermaid name" "1" "$(grep -cx 'npm install -g @mermaid-js/mermaid-cli' "$LOG" || true)"
hasnt "…and nothing is left failed" "did not install" "$out"
printf '\n\033[1ma tool installed a moment ago is still found\033[0m\n'
# uv lands in ~/.local/bin, which is on no PATH this process has -- dotup will
# not rewrite PATH for its own convenience, so find_tool looks in the places
# the run just wrote to instead. Stop looking there and every uv tool is
# reported missing on a machine where uv was installed sixty seconds earlier.
newbox
if holed uv; then
ln -s "$BOXPATH/_fake" "$SB/.local/bin/uv"
pick agents/specify-cli
out=$(box install 2>&1 | plain || true)
has "uv is found in ~/.local/bin, off PATH" "uv tool install specify-cli" "$(cat "$LOG")"
hasnt "…so nothing claims uv is missing" "uv missing" "$out"
is "…and the run exits clean" "0" "$(boxrc install)"
else
printf ' \033[33mskip\033[0m this host carries uv at one of find_tool'"'"'s absolute probes\n'
fi
# ~/bin is the other one, and it is not decorative: get.chezmoi.io installs
# there when -b is not given, which is what the README one-liner does.
newbox
if holed snap; then
mkdir -p "$SB/bin"; ln -s "$BOXPATH/_fake" "$SB/bin/snap"
pick core/bitwarden-cli
out=$(box install 2>&1 | plain || true)
has "…and in ~/bin, where get.chezmoi.io puts things" "snap install bw" "$(cat "$LOG")"
hasnt "…so nothing claims snapd is absent" "snapd is not present" "$out"
else
printf ' \033[33mskip\033[0m this host carries snap at one of find_tool'"'"'s absolute probes\n'
fi
printf '\n\033[1mthe deb channel installs what it downloads\033[0m\n'
newbox; pick apps/chrome
out=$(box install 2>&1 | plain || true)
has "the vendor .deb is fetched" \
"curl -fsSL https://dl.google.com/linux/direct/google-chrome-stable_current_amd64.deb" "$(cat "$LOG")"
is "…and then handed to the package manager" "1" \
"$(grep -c 'apt-get install -y .*\.deb' "$LOG" || true)"
hasnt "…and nothing is left failed" "did not install" "$out"
# A download that fails must not be followed by an install of a file that is
# not there -- and must be reported.
newbox; BOXENV=FAKE_FAIL=curl:deb; pick apps/chrome
out=$(box install 2>&1 | plain || true)
is "a failed download installs nothing" "0" \
"$(grep -c 'apt-get install -y .*\.deb' "$LOG" || true)"
has "…and is reported against the package" "apps/chrome" "$out"
printf '\n\033[1mthe neovim tarball lands where the symlink points\033[0m\n'
# Four commands name the same directory: the rm, the mkdir, tar's -C, and the
# symlink's target. A typo in any one of them installs nothing and leaves a
# dangling /usr/local/bin/nvim -- silently, because tar succeeded. Read the
# four back out of dotup's own dry run and make them agree, rather than
# trusting four separate spellings to stay in step.
newbox; pick core/neovim
dry=$(box --print install 2>&1 | plain || true)
xt=$(printf '%s\n' "$dry" | sed -n 's/.*tar -xzf [^ ]* -C \([^ ]*\) --strip-components=1.*/\1/p' | head -1)
xl=$(printf '%s\n' "$dry" | sed -n 's|.*ln -sf \(.*\)/bin/nvim /usr/local/bin/nvim.*|\1|p' | head -1)
xr=$(printf '%s\n' "$dry" | sed -n 's/.*rm -rf \([^ ]*\) && .*mkdir -p \([^ ]*\) &&.*/\1 \2/p' | head -1)
is "the tarball is extracted where the symlink points" "$xt" "$xl"
is "…into the directory that was cleared and recreated" "$xt $xt" "$xr"
is "…and that directory is /opt/nvim" "/opt/nvim" "$xt"
printf '\n\033[1mthe picker fetches its own fzf, from the real release URL\033[0m\n'
# The stand-in for GitHub answers ONE path -- the release download URL -- and
# builds the tarball from the version named in it. So a preflight that reports
# a version is evidence the pin travelled through the URL into the binary; a
# wrong host or a wrong path shape is a failed fetch, exactly as it would be.
newbox
out=$(box preflight 2>&1 | plain || true)
is "the fetch goes to the fzf release download URL" "1" \
"$(grep -c 'curl -sfL https://github.com/junegunn/fzf/releases/download/' "$LOG" || true)"
has "preflight resolves the fzf it just cached" "using $SB/.cache/dotup/fzf" "$out"
is "the binary that landed is the version the URL asked for" \
"$(grep -o 'fzf-[0-9][0-9.]*-linux_' "$LOG" | head -1 | sed 's/^fzf-//; s/-linux_$//')" \
"$(printf '%s\n' "$out" | sed -n 's/.*fzf (\([0-9][0-9.]*\),.*/\1/p' | head -1)"
is "…and it clears the verified floor" "1" \
"$(printf '%s\n' "$out" | grep -c 'floor 0.44.0' || true)"
is "nothing was written to ~/.local/bin" "absent" \
"$([ -e "$SB/.local/bin/fzf" ] && echo present || echo absent)"
printf '\n\033[1mthe picker actually runs — every binding reaches fzf\033[0m\n'
# DU-C1. Comments were inserted BETWEEN the continued lines of the fzf
# invocation. The `\`-newline is stripped first, so the comment's own newline
# terminated the command: fzf ran with two binds, the lines below it ran as a
# command named `--bind`, and `dotup pick` returned 1 having drawn a picker
# where nothing but the cursor worked. Every other test here greps the SOURCE
# for bind strings, so all of them still passed. This one runs cmd_pick.
#
# Static half first: it needs nothing, and it catches the whole class anywhere
# in the file rather than only at the one site that was broken.
stray=$(awk '/\\$/ { cont=1; next }
cont && /^[[:space:]]*#/ { printf "%d: %s\n", NR, $0 }
{ cont=0 }' $D)
is "no comment interrupts a line continuation" "" "$stray"
if command -v python3 >/dev/null; then
# A real pty, because cmd_pick's first act is to open /dev/tty; a fake fzf,
# because the assertion is about the argv it was handed. FAKE_FZF_ARGV
# writes that argv one argument per line, so a bind that lost its
# continuation cannot be mistaken for one that survived.
newbox
BOXPATH=$SB/seal; sh ./seal.sh "$BOXPATH" --with fzf >/dev/null
ARGV=$SB/fzf.argv; : > "$SB/nothing.expect"
prc=0
python3 ./ptydrive.py --timeout 30 --script "$SB/nothing.expect" -- \
/usr/bin/env -i HOME="$SB" PATH="$BOXPATH" TERM=dumb LANG=C \
TMPDIR="$SB/tmp" DOTUP_STATE="$SB/.config/dotfiles" \
DOTUP_MANIFEST="$M" DOTUP_TEST_LOG="$LOG" FAKE_FZF_ARGV="$ARGV" \
"$BOXPATH/sh" $D pick >/dev/null 2>&1 || prc=$?
is "dotup pick exits 0" "0" "$prc"
is "…having actually invoked fzf once" "1" \
"$(grep -c '^fzf --ansi' "$LOG" || true)"
# Counted against the source, not against a literal 7: a binding added
# later must reach fzf too, and a test that hard-codes the count would not
# notice that it did not.
is "…with every --bind the source writes" \
"$(grep -c -- '--bind ' $D)" "$(grep -cx -- '--bind' "$ARGV" || true)"
nobind=$(for k in space tab ctrl-t ctrl-a ctrl-x ctrl-o enter; do
grep -q "^$k:" "$ARGV" || echo "$k"
done; :)
is "…and all seven keys among them" "" "$nobind"
# The other half of what DU-C1 cost: the marker is written after fzf
# accepts, so with the command truncated it never was, and the defaults
# re-seeded over the user's selection on every subsequent run.
is "…and the completed-pick marker is written" "present" \
"$([ -f "$SB/.config/dotfiles/picked" ] && echo present || echo absent)"
else
printf ' \033[33mskip\033[0m needs python3 to open a pty\n'
fi
printf '\n\033[1mthe private tier — the three guards, exercised\033[0m\n'
# Each of these is one line of dotup, and each of them survived being INVERTED:
# the suite only ever checked that the words were in the file. What matters is
# which of the three fires, because they say three different things and only one
# of them is true at a time.
newbox; pick private/private-repo private/bws-secrets
out=$(box --unattended private </dev/null 2>&1 | plain || true)
has "--unattended: nobody is here to type it" "nobody is here to type" "$out"
hasnt "…so it never reaches a prompt" "Bootstrap URL" "$out"
hasnt "…and does not blame the terminal instead" "no terminal" "$out"
out=$(box private </dev/null 2>&1 | plain || true)
has "no tty: it says THAT instead" "no terminal" "$out"
hasnt "…and still never prompts" "Bootstrap URL" "$out"
hasnt "…and does not blame a flag nobody passed" "nobody is here to type" "$out"
out=$(box --print private </dev/null 2>&1 | plain || true)
has "--print describes the step" "prompt for URL, username, password" "$out"
hasnt "…and asks nothing" "Bootstrap URL:" "$out"
newbox; pick core/ripgrep
is "with no private row ticked there is no tier at all" "" \
"$(box private </dev/null 2>&1 | plain || true)"
printf '\n\033[1mthe private tier — driven through a real terminal\033[0m\n'
if command -v python3 >/dev/null 2>&1; then
# Everything below runs the SHIPPING cmd_private on a pty, against the
# sealed fakes: a stand-in endpoint that checks the password it was given,
# a stand-in GitHub that serves a bws zip and a checksum file, and a
# stand-in chezmoi that produces a source tree at the caller's umask. The
# assertions are then made against the filesystem -- modes, contents -- and
# against what the run actually said, not against what the source contains.
privrun() {
sc=$1; shift
(umask 022; python3 ./ptydrive.py --timeout 40 --script "$sc" -- \
env -i HOME="$SB" PATH="$BOXPATH" TERM=dumb LANG=C TMPDIR="$SB/tmp" \
DOTUP_STATE="$SB/.config/dotfiles" DOTUP_MANIFEST="$M" \
DOTUP_TEST_LOG="$LOG" FAKE_BOOT_PW="$PW" \
FAKE_BOOT_BLOB="$SB/blob.env" "$@" \
"$BOXPATH/sh" $D private) 2>&1 | tr -d '\r' | plain
}
mode_of() { stat -c %a "$1" 2>/dev/null || echo missing; }
# ---- a wrong password costs one password, not one reinstall -------------
# These credentials are asked for at the very END of a run. Before the retry
# loop existed, one mistyped character meant repeating the entire install to
# get back to this prompt. So the first attempt here is wrong on purpose.
newbox; pick private/private-repo private/bws-secrets
PW=pw-ok-$$
printf 'PRIVATE_REPO_URL=http://git:tok-%s@example.test/dotfiles-private.git\nBWS_ACCESS_TOKEN=bwstok-%s\n' \
"$$" "$$" > "$SB/blob.env"
cat > "$SB/steps" <<STEPS
expect Bootstrap URL:
send http://example.test/boot
expect Username:
send ben
expect Password:
send WRONG-ON-PURPOSE
expect wrong username or password
expect Bootstrap URL \[
send
expect Username \[
send
expect Password:
send $PW
STEPS
sed -i 's/^\t//' "$SB/steps"
out=$(privrun "$SB/steps")
has "a wrong password names the password" "wrong username or password" "$out"
has "…and says how to get out" "type q at the URL prompt" "$out"
has "…then asks again, keeping the URL" "Bootstrap URL [http://example.test/boot]" "$out"
has "…and the username" "Username [ben]" "$out"
is "…so the endpoint is asked exactly twice" "2" \
"$(grep -c 'bootstrap.env' "$LOG" || true)"
hasnt "the password is never echoed back" "WRONG-ON-PURPOSE" "$out"
# ---- what the second attempt actually produced --------------------------
is "the bws token is written 600" "600" "$(mode_of "$SB/.config/bitwarden/bws-token")"
is "…holding what the endpoint sent" "bwstok-$$" \
"$(cat "$SB/.config/bitwarden/bws-token" 2>/dev/null || true)"
is "the private source is unreadable to anyone else" "700" \
"$(mode_of "$SB/.local/share/dotfiles-private")"
is "the credential file is 600" "600" \
"$(mode_of "$SB/.config/dotfiles/private-credentials")"
is "…and holds the token the URL carried" "http://git:tok-$$@example.test" \
"$(cat "$SB/.config/dotfiles/private-credentials" 2>/dev/null || true)"
hasnt "no credential-bearing URL reaches the transcript" "tok-$$@" "$out"
is "…nor any command line the run built" "0" \
"$(grep -c '://[^ /]*:[^ /]*@' "$LOG" || true)"
has "chezmoi is called by its resolved path" \
"chezmoi init --apply --source $SB/.local/share/dotfiles-private" "$(cat "$LOG")"
has "…against the private tier's OWN config file" \
"-c $SB/.config/chezmoi/private.toml" "$(cat "$LOG")"
is "…which is never the config the public tier rewrites" "0" \
"$(grep -c 'chezmoi/chezmoi.toml' "$LOG" || true)"
is "a later update can still authenticate" "1" \
"$(grep -c "git -C $SB/.local/share/dotfiles-private config credential.helper" "$LOG" || true)"
has "bws is verified against the published checksum" "bws checksum verified" "$out"
is "…and installed, executable" "755" "$(mode_of "$SB/.local/bin/bws")"
is "…only after a token exists to use it" "yes" \
"$(printf '%s\n' "$out" | awk '/bws token written/{t=NR} /fetching bws/{f=NR}
END{print (t && f && t < f) ? "yes" : "no"}')"
# ---- the checksum is a decision, not a decoration -----------------------
# A binary about to hold the key to every other credential. Inverting this
# one comparison installs exactly the file the check exists to reject, and
# says "verified" while doing it.
privbox() { # privbox <selected>... -- fresh box, blob, password, prompts
newbox; pick "$@"
PW=pw-$$-$BOXN
printf 'PRIVATE_REPO_URL=http://git:tok@example.test/x.git\nBWS_ACCESS_TOKEN=bt-%s\n' \
"$BOXN" > "$SB/blob.env"
{ printf 'expect Bootstrap URL:\nsend http://example.test/boot\n'
printf 'expect Username:\nsend ben\n'
printf 'expect Password:\nsend %s\n' "$PW"; } > "$SB/steps"
}
privbox private/bws-secrets
out=$(privrun "$SB/steps" FAKE_BWS_SUMS=mismatch)
has "a checksum mismatch refuses to install" "checksum mismatch" "$out"
is "…and nothing lands in ~/.local/bin" "absent" \
"$([ -e "$SB/.local/bin/bws" ] && echo present || echo absent)"
hasnt "…and it does not claim to have verified anything" "checksum verified" "$out"
privbox private/bws-secrets
out=$(privrun "$SB/steps" FAKE_BWS_SUMS=absent)
has "no checksum file says so out loud" "checksums unavailable" "$out"
is "…and installs anyway, as it says" "755" "$(mode_of "$SB/.local/bin/bws")"
# ---- three failures, three different next moves -------------------------
# One message for 401, 404 and an unreachable host is how a URL-shape bug
# spends an evening looking like a password problem.
qsteps() { printf 'expect Bootstrap URL \[\nsend q\n' >> "$SB/steps"; }
privbox private/bws-secrets; qsteps
has "a 404 names the route" "no bootstrap.env is there" \
"$(privrun "$SB/steps" FAKE_BOOT_CODE=404)"
privbox private/bws-secrets; qsteps
has "an unreachable host says so" "could not reach that address" \
"$(privrun "$SB/steps" FAKE_BOOT_CODE=000)"
privbox private/bws-secrets; qsteps
has "anything else reports its code" "answered HTTP 500" \
"$(privrun "$SB/steps" FAKE_BOOT_CODE=500)"
# An error page must never be parsed as the credential blob.
privbox private/bws-secrets; qsteps
out=$(privrun "$SB/steps" FAKE_BOOT_CODE=404)
is "a non-200 body is never read as a token" "absent" \
"$([ -e "$SB/.config/bitwarden/bws-token" ] && echo present || echo absent)"
has "…and q leaves a public-only machine" "public-only machine" "$out"
# ---- it asks again, but not forever -------------------------------------
privbox private/bws-secrets
{ printf 'expect Bootstrap URL:\nsend http://example.test/boot\n'
printf 'expect Username:\nsend ben\n'
printf 'expect Password:\nsend bad1\n'
i=2
while [ "$i" -le 5 ]; do
printf 'expect wrong username or password\n'
printf 'expect Bootstrap URL \[\nsend\nexpect Username \[\nsend\n'
printf 'expect Password:\nsend bad%s\n' "$i"
i=$((i + 1))
done
printf 'expect five failed attempts\n'; } > "$SB/steps"
out=$(privrun "$SB/steps")
has "five wrong passwords stop the loop" "five failed attempts" "$out"
has "…pointing at the cheap way back in" "dotup private" "$out"
is "…having asked the endpoint five times, not four or six" "5" \
"$(grep -c 'bootstrap.env' "$LOG" || true)"
# ---- q at the first prompt is a supported answer ------------------------
privbox private/private-repo private/bws-secrets
printf 'expect Bootstrap URL:\nsend q\n' > "$SB/steps"
out=$(privrun "$SB/steps")
has "q at the first prompt is public-only" "public-only machine" "$out"
is "…and writes nothing at all" "0" \
"$(grep -c 'bootstrap.env' "$LOG" || true)"
is "…no token" "absent" "$([ -e "$SB/.config/bitwarden/bws-token" ] && echo present || echo absent)"
is "…and no clone" "absent" "$([ -e "$SB/.local/share/dotfiles-private" ] && echo present || echo absent)"
else
printf ' \033[33mskip\033[0m needs python3 to open a pty\n'
fi
printf '\n\033[1minteractive loop (needs a pty)\033[0m\n'
if command -v fzf >/dev/null && command -v curl >/dev/null && command -v script >/dev/null; then
out=$(sh ./listen-test.sh 2>/dev/null | tr -d '\r' || true)