refactor(tests): discover suites instead of naming them

run-tests.fish globs tests/test-*.fish and runs each in the mode the suite
declares in its own header. Isolated is the default and no typo can
promote a suite to in-session: detection is case-insensitive so a
near-miss is caught, the value comparison is exact.

In-session suites share one sandboxed session; isolated suites each get a
--no-config child with temp XDG dirs. Counts come back through a file so
output keeps streaming.

functional.fish becomes test-session.fish, its 15 predicates rewritten as
check calls. 332 assertions, unchanged per-phase.
This commit is contained in:
2026-09-07 20:02:59 -04:00
parent d3028d7703
commit 0a9cfebcad
3 changed files with 212 additions and 209 deletions
+150 -95
View File
@@ -4,15 +4,19 @@
#
# CI test runner for this fish configuration.
# 1. Syntax-lints every tracked .fish file (fish -n).
# 2. Copies the config-relevant files into a throwaway sandbox (never
# the live checkout -- this repo doubles as a real ~/.config/fish,
# so a symlinked sandbox would let universal-variable writes like
# first-run's escape into the real, gitignored fish_variables file)
# and loads it as an isolated interactive session.
# 3. Runs the functional checks in tests/functional.fish inside that
# loaded session.
# 4. Runs tests/test-agents-vault.fish as its own process; that suite
# builds its own throwaway repos and needs no loaded config.
# 2. Discovers tests/test-*.fish and reads the mode each suite declares
# in its own header (`# MODE: isolated` or `# MODE: in-session`).
# 3. Runs each isolated suite as its own --no-config fish process with
# throwaway XDG dirs.
# 4. Sources every in-session suite into ONE sandboxed interactive fish
# session built from a copy of this config (never the live checkout
# -- this repo doubles as a real ~/.config/fish, so a symlinked
# sandbox would let universal-variable writes like first-run's escape
# into the real, gitignored fish_variables file).
# 5. Sums the per-suite assertion counts and reports a total.
#
# Assertions come from tests/lib.fish (section/check/report); suites are
# never special-cased here by name.
#
# Usage: fish tests/run-tests.fish
@@ -42,104 +46,155 @@ if test $lint_failed -ne 0
set overall_failed 1
end
# ---- Phase 2: isolated load + functional checks --------------------------
echo ""
echo "== Sandboxed load + functional checks =="
set -l sandbox (mktemp -d)
set -l sandbox_cfg $sandbox/xdgcfg/fish
mkdir -p $sandbox_cfg
# path-setup only adds directories that already exist (fish_add_path is a
# no-op on missing paths), so give it $HOME/.local/bin to find.
mkdir -p $sandbox/home/.local/bin
# Every utility below goes through `command`. This driver runs under the
# very config it tests, which shadows these: `cp` is an alias for `cp -i`,
# `rm` is a trash wrapper, `cat` resolves to bat. Only `cp` is an actual
# hazard today -- `-i` on a non-empty destination reads EOF in a
# non-interactive runner and SILENTLY SKIPS the copy while exiting 0,
# which would leave the sandbox missing config files and report success.
# `rm -rf` and `cat` were measured and behave correctly as-is (the rm
# wrapper bails to `command rm` on any non-recursive flag, so -rf really
# deletes and does not trash). Prefixed anyway: a test runner must not
# depend on the configuration under test.
command cp $repo_root/config.fish $sandbox_cfg/
test -f $repo_root/fish_plugins
and command cp $repo_root/fish_plugins $sandbox_cfg/
for d in functions conf.d completions integrations themes data
test -d $repo_root/$d
and command cp -r $repo_root/$d $sandbox_cfg/
# ---- Phase 2: discover suites --------------------------------------------
# Mode is declared by the suite, not by this driver. Detection is
# case-insensitive so a near-miss like "# Mode: in-session" is caught rather
# than silently read as "no marker"; the comparison is exact so only the two
# real spellings are accepted. Absence means isolated, the safe default -- a
# suite that forgets the marker gets its own clean process instead of being
# injected into a loaded session, and no typo can ever promote a suite into
# in-session. Duplicate markers resolve first-match-wins.
set -l isolated_suites
set -l session_suites
for f in (find $script_dir -name 'test-*.fish' | sort)
set -l decl (grep -im1 '^# *mode:' $f)
if test -z "$decl"
set -a isolated_suites $f
else if test "$decl" = "# MODE: in-session"
set -a session_suites $f
else if test "$decl" = "# MODE: isolated"
set -a isolated_suites $f
else
echo " FAIL "(basename $f)": unrecognized mode declaration: $decl" >&2
set overall_failed 1
end
end
set -l err_file (mktemp)
env -i \
HOME=$sandbox/home \
XDG_CONFIG_HOME=$sandbox/xdgcfg \
PATH="$PATH" \
TERM=xterm \
__fish_config_op_autoexec=off \
fish -i -c "source $repo_root/tests/functional.fish; functional_test_main" \
2>$err_file
set -l functional_status $status
set -l counts (mktemp)
set -l stderr_out (command cat $err_file)
command rm -rf $sandbox $err_file
if test -n "$stderr_out"
# Diagnostic only, not a gate: on machines with vendor fish configs
# (e.g. CachyOS's cachyos-fish-config, which this repo's config.fish
# sources when present) unrelated vendor warnings can land here. Real
# breakage in this repo's own code is caught by the assertions below.
echo " Session stderr output (informational):"
printf '%s\n' $stderr_out
end
if test $functional_status -ne 0
set overall_failed 1
end
# ---- Phase 3: hermetic vault helper tests --------------------------------
# Run as its own fish process rather than inside the sandboxed session:
# the suite builds its own throwaway git repos and binds the vault, claude
# and agy roots to them, so it needs no loaded config and must never see
# the real ~/.claude.
#
# ---- Phase 3: isolated suites --------------------------------------------
# HOME is deliberately NOT overridden here. Read this before "improving" it.
#
# Overriding XDG_CONFIG_HOME/XDG_DATA_HOME plus --no-config is what makes
# this run isolated: the universal-variable file fish can reach is a fresh
# empty one, and no config.fish/conf.d is loaded. Without that, an
# "isolated" suite runs against the user's LIVE config and real universal
# variables -- this repo doubles as a real ~/.config/fish -- so a test
# doing `set -e __fish_config_op_logging` would erase a real universal
# variable out of the running shell. Measured:
# $__fish_config_op_registry_keys has 65 entries under a plain `fish`, 0
# under `fish --no-config`.
# Overriding XDG_CONFIG_HOME/XDG_DATA_HOME plus --no-config is what makes these
# runs isolated: the universal-variable file fish can reach is a fresh empty
# one, and no config.fish/conf.d is loaded. Without that, an "isolated" suite
# runs against the user's LIVE config and real universal variables -- this repo
# doubles as a real ~/.config/fish -- so a guard test doing
# `set -e __fish_config_op_logging` would erase a real universal variable out of
# the running shell. Measured: $__fish_config_op_registry_keys has 65 entries
# under a plain `fish`, 0 under `fish --no-config`.
#
# `env -i HOME=$sandbox` was tried and REJECTED. It looks strictly more
# hermetic, but test-agents-vault.fish's hermeticity floor snapshots the
# real $HOME/.claude/memory and $HOME/.gemini/antigravity-cli and asserts
# them unchanged at the end. Point HOME at a sandbox and both snapshots
# read "absent" before and after: the assertions still pass while
# asserting nothing. A change that turns a real assertion into a tautology
# without turning anything red is the worst failure mode a test harness
# has. Keeping HOME real is what keeps those two assertions biting.
# hermetic, but test-agents-vault.fish's hermeticity floor snapshots the real
# $HOME/.claude/memory and $HOME/.gemini/antigravity-cli and asserts them
# unchanged at the end. Point HOME at a sandbox and both snapshots read "absent"
# before and after: the assertions still pass while asserting nothing. A change
# that turns a real assertion into a tautology without turning anything red is
# the worst failure mode a test harness has. Keeping HOME real is what keeps
# those two assertions biting.
#
# Overriding XDG_DATA_HOME is a hermeticity gain on top of the isolation:
# _agents_vault_dir falls back to
# ${XDG_DATA_HOME:-$HOME/.local/share}/agent-vault, so a vault path that
# no test overrode lands in a temp dir instead of the user's real
# ~/.local/share/agent-vault.
# ${XDG_DATA_HOME:-$HOME/.local/share}/agent-vault, so a vault path that no test
# overrode lands in a temp dir instead of the user's real ~/.local/share.
for suite in $isolated_suites
echo ""
echo "== "(string replace $repo_root/ '' $suite)" =="
set -l xdg (mktemp -d)
env XDG_CONFIG_HOME=$xdg/cfg XDG_DATA_HOME=$xdg/data \
FISH_CONFIG_TEST_ROOT=$repo_root FISH_CONFIG_TEST_COUNTS=$counts \
fish --no-config $suite
if test $status -ne 0
set overall_failed 1
end
command rm -rf $xdg
end
# ---- Phase 4: in-session suites ------------------------------------------
# All in-session suites share ONE sandboxed interactive session: building it
# (copying the config, starting fish -i) is the expensive part.
#
# The config is COPIED, never symlinked. This repo doubles as a real
# ~/.config/fish, so a symlinked sandbox would let universal-variable writes
# like first-run's escape into the real, gitignored fish_variables file.
if test (count $session_suites) -gt 0
echo ""
echo "== Sandboxed load + session checks =="
set -l sandbox (mktemp -d)
set -l sandbox_cfg $sandbox/xdgcfg/fish
mkdir -p $sandbox_cfg
# path-setup only adds directories that already exist (fish_add_path is a
# no-op on missing paths), so give it $HOME/.local/bin to find.
mkdir -p $sandbox/home/.local/bin
# Every utility below goes through `command`. This driver runs under the
# very config it tests, which shadows these: `cp` is an alias for `cp -i`,
# `rm` is a trash wrapper, `cat` resolves to bat. Only `cp` is an actual
# hazard today -- `-i` on a non-empty destination reads EOF in a
# non-interactive runner and SILENTLY SKIPS the copy while exiting 0,
# which would leave the sandbox missing config files and report success.
# `rm -rf` and `cat` were measured and behave correctly as-is (the rm
# wrapper bails to `command rm` on any non-recursive flag, so -rf really
# deletes and does not trash). Prefixed anyway: a test runner must not
# depend on the configuration under test.
command cp $repo_root/config.fish $sandbox_cfg/
test -f $repo_root/fish_plugins
and command cp $repo_root/fish_plugins $sandbox_cfg/
for d in functions conf.d completions integrations themes data
test -d $repo_root/$d
and command cp -r $repo_root/$d $sandbox_cfg/
end
set -l srcs
for s in $session_suites
set -a srcs "source $s;"
end
set -l err_file (mktemp)
env -i \
HOME=$sandbox/home \
XDG_CONFIG_HOME=$sandbox/xdgcfg \
PATH="$PATH" \
TERM=xterm \
__fish_config_op_autoexec=off \
FISH_CONFIG_TEST_ROOT=$repo_root \
FISH_CONFIG_TEST_COUNTS=$counts \
fish -i -c "source $repo_root/tests/lib.fish; $srcs report" \
2>$err_file
set -l session_status $status
set -l stderr_out (command cat $err_file)
command rm -rf $sandbox $err_file
if test -n "$stderr_out"
# Diagnostic only, not a gate: on machines with vendor fish configs
# (e.g. CachyOS's cachyos-fish-config, which this repo's config.fish
# sources when present) unrelated vendor warnings can land here. Real
# breakage in this repo's own code is caught by the assertions.
echo " Session stderr output (informational):"
printf '%s\n' $stderr_out
end
if test $session_status -ne 0
set overall_failed 1
end
end
# ---- Phase 5: totals -----------------------------------------------------
set -l total_run 0
set -l total_failed 0
for line in (command cat $counts)
set -l parts (string split ' ' -- $line)
set total_run (math $total_run + $parts[1])
set total_failed (math $total_failed + $parts[2])
end
command rm -f $counts
echo ""
echo "== Vault helper tests =="
set -l vault_xdg (mktemp -d)
env XDG_CONFIG_HOME=$vault_xdg/cfg XDG_DATA_HOME=$vault_xdg/data \
fish --no-config $repo_root/tests/test-agents-vault.fish
if test $status -ne 0
echo "TOTAL: "(math $total_run - $total_failed)"/$total_run assertions passed"
if test $total_failed -ne 0
set overall_failed 1
end
# `command rm`, not bare `rm`: this driver runs under the config it tests,
# which shadows rm/cp/cat. See the Phase 2 note for the full reasoning.
command rm -rf $vault_xdg
exit $overall_failed