Three review findings, all before the hook is wired:
is_shell_lib excluded */lib/tests/* but not */tests/lib/*. Both are test
code; the second nesting made the gate demand tests/lib/tests/run_all.sh
for 10 helper files. It would have fired precisely while editing test
helpers to write the suites this ratchet asks for. Reseeded 125 -> 115.
--seed now refuses (exit 1) when a reseed would ADD an entry, instead of
a header comment asking nicely. It reads tracked files only, so an
unstaged probe seeds normally -- staging it produces the refusal with
the allowlist left byte-identical.
The parent commit's docs claimed an exempt file loses its exemption when
next modified. It does not: check_files exempts allowlisted paths
unconditionally, verified by staging an edit and getting exit 0. Fixed
the text, not the code -- making an edit trip the gate would block
routine work across 115 of 198 libraries, which is the repo-wide gate
this ratchet exists to avoid.
Also corrects the "a percentage is unreachable" rationale: commit
53e80462 reached 100% on an entry script with no suppressions. Presence
is still the right bar, because that took per-file mount analysis.
Measurement-only. No production file changed.
The gating question was whether the campaign's constraints are jointly
satisfiable: 100% line coverage over ALL .sh files with no suppressions.
Phase 1 measured effecting code at 0% under PATH shims, which implied
either forbidden suppressions or source seams in 17+46 files.
They are satisfiable. A user+mount namespace reaches 100% with zero
source changes: --map-root-user makes require_root fall through, and
bind-mounting each write target makes installer writes land in the jail.
setup_night_lockdown.sh went 81.58% (PATH shims) -> 86.84% (userns) ->
100.00% (38/38, all mounts).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
Replaces the Q1/Q2/Q3 grilling block with the answers the ratchet was
built under, flagged as cheaply vetoable since the hook is not yet
wired. Documents the one remaining step (the .pre-commit-config.yaml
entry, denied as a sensitive file) and corrects the file-count method:
`find` reports 565 .sh because it walks .venv and node_modules, which
is how the stale "roughly 487" got in here.
A shell library under a lib/ directory must now sit beside a
tests/run_all.sh. This is a presence gate, not a coverage percentage:
Phase 1 measured these libs' decision logic at ~90% under test but
their effecting code at 0%, because it sudo-writes, pkills and installs.
A numeric bar would only be reachable by shimming every external for
every installer, or by the suppressions this repo forbids.
Seeded at 125 exempt libraries out of 208 tracked; the other 83 already
have a suite. The allowlist shrinks only — a modified library loses its
exemption and must bring tests.
Verified standalone: blocks a new untested lib (exit 1), and does NOT
block covered libs, exempt libs, or unrelated entry scripts. That last
case is the point of a ratchet over the repo-wide gate.
The .pre-commit-config.yaml entry is deliberately NOT included — the
edit was denied as a sensitive file, so the gate is runnable but not
yet enforced.
Phases 1 and 2 are done — the cap is at zero and enforced, and both CI gates
are wired — so the handoff now carries what a Phase 3 session needs instead.
Phase 3 opens with questions rather than work, because the scale (roughly 487
.sh files against a handful of lib/tests/ dirs) makes scope the user's call.
The coverage numbers Phase 1 actually measured are included so the "what bar"
question gets answered with evidence: decision logic reached 94%, while every
function that pkills, sudo-writes or installs sat at 0%. A 100% bar means
shimming sudo/pkill/systemctl/curl/npm across every installer, and that cost
belongs in front of the user before anyone starts writing tests.
Also records the traps that cost real time and are all still live: shfmt
corrupting hyphenated array subscripts, `# shellcheck` at the start of a
comment being parsed as a directive, verify_shell_split.sh being blind to
heredoc-internal functions, and the exec-bit staging rule.
Notes that Python tests has been red since 2026-08-17 from a pip
resolution-too-deep failure that predates this campaign and is unrelated to
it — which is why python-tests.yml is not in check_ci_green.sh's required list.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
The artifact was written for an earlier draft of the change and did not match
what shipped. It claimed _glyph_art.py went 253 -> 208 when HEAD was 220 and
the committed file is 238, omitted apps.py entirely, and carried the line
"Non-goal: any change to the glyph artwork itself" — false, since the whole
point of the commit was adding the note-ascender glyph and the lyricanki
AppIcon entry.
Corrected the line counts, added apps.py to scope, replaced the false non-goal
with the real one (the twelve pre-existing glyphs and style.py are untouched),
and recorded the linux=False check: ~/lyricanki has android/ and no linux/, so
there is no desktop hicolor set to install. Evidence that misdescribes its own
diff is worse than none — it is what a future reader bisects against.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two pre-commit hooks and one pre-push extension, plus the _glyph_art split
that had to land first — the cap gate cannot be wired while any file is over
it, and the lyricanki note-ascender glyph had taken _glyph_art.py to 253.
file-length-cap runs check_file_length.sh --all repo-wide rather than on
staged files, so a file that grows past the cap in some other commit is still
caught. Proven to block, not assumed: with a deliberate 261-line file staged
the hook exits 1 and names it; with the file removed it passes.
ci-baseline-green is the user's option (a): refuse to build on a main that is
already red. It fails closed on a definitively failing baseline and warns-and-
passes on anything it cannot determine — offline, no gh, unauthenticated, no
settled runs — each call capped at 5s. Blocking every commit on network
reachability buys nothing when (b) is the real per-change guarantee.
CI_GREEN_SKIP=1 is the escape for a fix-forward commit. python-tests.yml is
excluded on purpose: it is path-filtered, so a shell-only commit never
triggers it and its "latest run" would produce false reds.
This hook already earned its place. Run against a red main earlier in this
session it refused with "Shell tests (failure)", which is how the broken
SETUP_SOURCES was found instead of being pushed over.
ci_mirror gains the shell-tests workflow's side-effect-free half — every
lib/tests/run_all.sh plus the workflow's own test list, the user's option (b).
That list is parsed out of shell-tests.yml rather than copied: a hardcoded
duplicate drifts the moment someone adds a test, and a mirror that quietly
checks less than CI is worse than no mirror because it reports safe. The
Arch-container half stays unmirrored — it wants root and an Arch userland.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
The template checks were grepping an empty argument list. SETUP_SOURCES was
initialised as SETUP_SOURCES=("${SETUP_SOURCES[@]}") — a self-reference to an
array that did not exist yet — so it started empty, never contained
$SETUP_SCRIPT, and the lib-discovery loop ran dirname on nothing.
CI reported "setup_midnight_shutdown.sh should install the D-Bus monitor
helper". Reproduced locally before changing anything, so the fix is verified
against the real failure.
Seeding it with ("$SETUP_SCRIPT") restores the intent: search the entry
script together with lib/ms_*.sh, since the monitor template moved into
lib/ms_monitor.sh during the split.
test_pacman_wrapper_security.sh and test_hosts_guard_pacman_integration.sh
re-checked at the same time and both pass, so the Shell tests set is green
again before the file-length gate goes in.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
1734 -> 109 lines across eight libs plus one payload file. This was the last
file over the 250-line cap; check_file_length.sh --all now reports "all
checked files are within 250 lines".
This file needed a different proof from the others. Its product is scripts
generated into /usr/local/bin via heredocs, so "did the functions move
verbatim" is not the question that matters — "does it still emit the same
bytes" is. Every heredoc body was hashed before the split: 11 of 12 are
byte-identical afterwards, and the twelfth (the 224-line config-unlock body)
now lives in lib/payloads/shutdown-config-unlock.sh.in at hash
1a5c67cfd0244101, exactly matching what it replaced. No new bodies appeared.
The "~39 top-level functions" this file is documented as having are not real.
log(), now_epoch(), cmd_add() and friends sit at column 0 because they are
inside the generated scripts. The brace-matching extractor keeps each heredoc
with its owning generator, so they moved as part of it. That is also why
verify_shell_split.sh reports a bare "-log": it cannot tell a heredoc-internal
definition from a top-level one. log() is not lost — it is line 17 of the
payload, which is where it always ran from.
$_MS_PAYLOAD_DIR is defined in the same file as its use, directly above it,
following the _NL_PAYLOAD_DIR lesson from earlier in this campaign.
--help is byte-identical, shellcheck is clean with no suppression added.
`setup` itself is not executed — it installs systemd units, makes the config
immutable and arms a shutdown timer.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
Shell tests has been red since fbe9134c. Three assertions across two
files grep pacman_wrapper.sh for functions the split moved into the flat
pw_*.sh libs: pacman_hooks_manage_guard_lib and
should_use_wrapper_guard_lib_fallback (pw_guard.sh),
verify_policy_integrity and the two -U helpers (pw_policy.sh), and
auto_remove_virtualbox_vms plus VBoxManage (pw_cleanup.sh).
Each grep now searches the entry script and its libs together, via a
pw_*.sh glob rather than a named lib, so a future re-partition does not
break them again. The sed extraction in the security test reads
pw_policy.sh directly, since it runs those two functions in isolation.
Deliberately left narrow: the two assert_order checks still target the
entry script alone. They assert the order of two call sites, and both
calls are still there -- widening them would turn a positional check
into a meaningless one. The 'CRITICAL near verify_policy_integrity' and
auto_remove_virtualbox_vms call-site greps were also left alone after
confirming they still match in the entry script.
Both files now exit 0. Of the two other suites still failing locally,
neither is from this commit: test_security_hardening's remaining failure
is this machine's unwrapped signal-desktop/discord and passes in CI, and
test_shutdown_timer_monitor is failing on the uncommitted
setup_midnight_shutdown split in the working tree.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
929 -> 217 lines across five pw_*.sh libraries. This file IS /usr/bin/pacman
here (the symlink points at the deployed copy), so the design constraint was
not readability but that no failure mode may leave the machine unable to
install packages.
Fail OPEN, unlike the other splits in this campaign. A missing lib warns on
stderr and execs real pacman unwrapped. Failing closed would mean a missing
lib leaves no way to install anything, including the missing lib. This is the
same choice the existing pacman_lock_lib.sh block already makes, and it is
verified rather than asserted: with pw_guard.sh removed from a scratch-dir
copy, the wrapper warns, execs, and returns real `-Qi bash` output at rc=0.
The libs load at the very top, before the fast path, because the fast path
calls needs_unlock() to decide whether to exec straight through, and that
function now lives in pw_guard.sh.
They are deployed flat beside the wrapper — the shape pacman_lock_lib.sh and
heavy_job_lock.sh already use in /usr/local/bin — as required copies, and are
added to the integrity manifest. Tamper detection was round-tripped in a fake
root: 8 manifest entries verify clean, and appending one line to
pw_challenge.sh is detected and the transaction denied.
Verified by execution, not just textually: real `-Qi bash` and `-Q` through
the split wrapper are byte-identical to the pre-split baseline, same 2824
packages. verify_shell_split.sh reports all 14 top-level functions moved
verbatim.
The deployed copy is deliberately NOT updated by this commit and still hashes
548fc2b5... exactly as before the work started; live pacman is unaffected.
A live `pacman -S` remains unverified — the Bash tool deadlocks on db.lck and
a failed experiment costs the package manager.
Cap is now down to one file.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
The detection regex matched only `name() {`. bash also accepts
`function name() {` and `function name {`, and pacman_wrapper.sh uses the
keyword form for 17 of its 30 functions — so a split of that file would have
left every one of them behind in the entry script while the tool reported
success. Silent, and exactly the kind of half-done split the verifier is
meant to catch afterwards rather than the extractor causing.
Parentheses stay mandatory in the non-keyword branch. Making them optional
everywhere also matches brace groups and control constructs, which would
sweep non-function code into the library — the opposite failure, and worse.
Nine negative tests pin that boundary.
The function name now comes from a named capture group rather than
splitting on "(", which returned "function name" for the keyword form and
would have broken both --functions selection and the verifier's name matching.
Verified against the already-split files first: cco_state.sh still yields 13
functions and leechblock_fetch.sh 3, so the completed splits are unaffected.
pacman_wrapper.sh goes from 13 detected to 31.
First tests for this script: 20 cases covering every accepted spelling,
brace matching across nested bodies, and single-line functions.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
Shell tests has been red on main since 72e1b6a8. Two assertions grep
block_compulsive_opening.sh for AUTO_CLOSE_TIMEOUT_MINUTES and
launch_with_timer -- both of which the 250-line-cap split moved into
lib/cco_wrapper.sh. The code is fine; the test's idea of where it lives
was stale, and it reported "Function not found" for something that had
simply moved one file over.
Both greps now search the entry script plus every lib/cco_*.sh. The glob
means the next split does not break them again, and the assertions stay
about the feature existing, which is what they were always for.
CI's only failure on the red run was launch_with_timer, so this targets
the actual break. The one failure that remains locally, "Compulsive
block wrappers installed", is this machine's state -- signal-desktop and
discord are installed but unwrapped, which the blocker's own status
reports independently -- and it passes in CI, where neither app exists.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The block_compulsive_opening split moved launch_with_timer into
lib/cco_wrapper.sh, and tests 17 and 18 grep the entry script alone. CI went
red on "launch_with_timer function exists — Not found at expected location":
a false failure about code that is present and working, one file over.
Both assertions are about a feature existing, so they now search the entry
script together with lib/cco_*.sh. That restores what they were always
checking and does not go stale on the next split.
The lib list is built with a loop and an -f test rather than a bare glob, so
an unmatched pattern cannot land in the array as a literal and no shellcheck
suppression is needed.
Verified by running the suite: both checks PASS, 18 passed. The one remaining
local failure is test 19, which inspects /usr/bin on this machine — signal
and discord are not currently wrapped here and beeper is not installed. That
is machine state, predates this change, and is not what CI reported.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
The split referenced $_NL_PAYLOAD_DIR from both install functions but
defined it nowhere, so `setup_night_lockdown.sh setup` would have died
with an unbound variable partway through installing the lockdown. My
earlier scratch verification set the variable by hand, which is exactly
what hid the bug: the emitted artifacts hashed correctly while the real
entry path could not have reached them.
It is now defined in each lib that uses it, eight lines above the use,
via `: "${_NL_PAYLOAD_DIR:=...}"` so an explicit override still wins. A
definition placed in the entry script was silently stripped by an
autoformat pass three times during this work, and a half-run lockdown
installer is not an acceptable failure mode.
Also drops two zero-byte payload files left behind by the .sh -> .sh.in
rename.
Re-verified without presetting anything: install_enter_script and
install_unlock_script emit mode-0755 files hashing 4baaf405... and
eafce28d..., byte-identical to the live /usr/local/bin copies. help and
an unknown command match their baselines; status differs only by the
05:00 countdown ticking 13h to 12h between runs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
918 -> 138 lines, across five libs plus two data-file payloads.
The two generated scripts became data files rather than heredocs.
ENTER_EOF alone was 279 lines, so any lib containing it would have been
over the cap however the rest was arranged. Both heredocs were quoted,
so their bodies were already literal text with no build-time
interpolation, and copying them verbatim is exactly equivalent.
The payloads are named *.sh.in on purpose. The pre-commit hook globs
staged files by '\.(sh|bash|zsh)$', runs shfmt -w on them and re-stages
the result -- and shfmt does reformat these bodies. That would have
changed their bytes and destroyed the property this split is verified
by. The .in extension keeps them out of every .sh-globbing hook without
a suppression or an exclude regex.
cat + chmod 0755 became a single `install -m 0755`, closing the window
the original left between creating the file and making it executable.
The load-bearing check: running install_enter_script and
install_unlock_script against a scratch directory produces mode-0755
files whose sha256 sums are identical to the copies already deployed in
/usr/local/bin on this machine. The split provably emits the exact bytes
that are running today, and it touched no system file -- sha256sum -c on
both deployed artifacts passes unchanged.
help, status and an unknown command are byte-identical to their
pre-split baselines. A line-multiset check leaves exactly six lines
unaccounted for: the two cat/heredoc openers, the two terminators and
the two chmod calls that install -m now performs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
21 assertions for cco_state.sh, the lib that decides whether a launch is
permitted. Coverage measured: 62/66 lines = 93.94%. The only uncovered lines
are kill_app's pkill/pkill -9 pair, which must never run in a test on a
machine whose apps are live. cco_wrapper/cco_install/cco_report stay at 0%:
every function in them moves system binaries, writes /etc/pacman.d/hooks or
kills processes.
The assertion that matters most is the one in the negative direction: a
PREVIOUS hour's stamp must NOT keep blocking. If that decision inverted, a
legitimately-closed app could never be reopened, and the failure would look
like the tool working. Both directions are pinned, plus the empty-stamp case
and stale suspend-marker cleanup.
Also generalises run_all.sh from test_leechblock_*.sh to test_*.sh. This
directory now holds suites for more than one entry script, and a prefix glob
silently skips every suite added after it — including inside the CI discovery
step added in fbec4656, which runs this file.
The harness asserts its own fixture is complete rather than letting a typo in
a global surface as a confusing unbound-variable from inside the code under
test. That check is also what makes the fixture's globals legible to
shellcheck, which cannot see them being read across the source boundary — a
real check that happens to remove SC2034, not a suppression.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
705 -> 182 lines, with state, the launch gate, install/uninstall and
reporting moved into cco_state.sh, cco_wrapper.sh, cco_install.sh and
cco_report.sh.
The entry script keeps its exact deployed path and stays a real file
there, not a symlink: the app wrappers already installed on disk
hard-code /usr/local/bin/block-compulsive-opening.sh, and the pacman
hook invokes it. _CCO_LIB_DIR probes for lib/ beside the script and
falls back to the script's own directory, so the repo tree and the flat
/usr/local/bin deploy layout both work; both branches were exercised.
install_all deploys the libs and uninstall_all removes them in this same
commit, so there is never a deployed copy that is split but missing its
libs -- the failure that had to be repaired for install_leechblock.
The libs deploy flat into /usr/local/bin rather than into
/usr/local/share/digital_wellbeing/lib, because install_leechblock_payload
does an rm -rf mirror of that directory and would delete them.
Verified beyond the usual gates: a line-multiset check reports zero
missing functional lines, and all three heredoc bodies -- the generated
app wrapper, the generated pacman hook and the help text -- are
byte-identical, which verify_shell_split.sh cannot check. status output
matches the still-monolithic deployed copy byte for byte.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The pacman-wrapper installer copied install_leechblock.sh to
/usr/local/share/digital_wellbeing/ along with three data files, but no
lib/. After the split in 1dceca66 the deployed entry script sources
every install phase from lib/ beside itself, so deploying it without
that directory leaves a copy that dies on its first `source`.
This is not hypothetical: the live deployed copy was found in exactly
that state at the start of this work -- the previous session's split
entry script, deployed and then left behind when the repo was reverted.
It failed on every invocation. It has been repaired and now carries the
libs.
The glob is leechblock_*.sh rather than *.sh because lib/ is shared with
unrelated music_* helpers that have no business in /usr/local. The copy
mirrors rather than merges, so a lib deleted upstream does not linger.
Verified against a fake root and then for real: the deployed copy runs,
its sha256 matches the repo, and inject_default_config resolves the
defaults JSON and the seeder from the deploy root rather than from lib/.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four */lib/tests/run_all.sh runners existed — periodic_background, hosts,
single_use/utils and the new digital_wellbeing one — and no workflow ran any
of them. Every assertion the cap-split sessions wrote was executing only on
the author's machine.
Discovered rather than listed: each split adds another runner, and a
hardcoded list is what let the first four go unnoticed. The step fails if
discovery matches zero runners, so a future path change cannot quietly turn
this into a vacuous green step.
Verified by running the step's exact logic under bash: 4 runners discovered,
all exit 0. Checked for Arch-only dependencies first — the only `command -v`
probe in the harnesses is `unshare` (present on ubuntu-latest, and guarded),
and no suite touches pacman, chattr, systemctl or adb, so these belong in the
Ubuntu job rather than the Arch container.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
29 assertions across two files, run via lib/tests/run_all.sh.
Coverage measured, not waived: leechblock_fetch.sh 38.57% (27/70),
leechblock_browsers.sh 36.47% (31/85), config and firefox 0%. The uncovered
lines are exactly the bodies that must never run under test — curl/tar/unzip
downloads, pkill + sudo binary patching, npm install, and writes to
/etc/*/policies.json. Closing the gap needs those externals shimmed.
What the tests actually pin down:
- all three get_latest_tag fallbacks (releases API, tags API, Location
header), plus JSON-null-is-not-a-tag and the all-sources-exhausted
non-zero return. A silent failure there installs the wrong version.
- the 8+3 browser tables including every hyphenated key, and the
command -v gating that decides whether a binary is patched at all.
Also fixes a latent set -u abort: wire_up_browsers interpolates VERSION_DIR,
VERSION and TAG, which only resolve_version sets. Calling it on its own
aborted on an unbound VERSION_DIR; they are now read through locals with
defaults, leaving production output unchanged.
Harness note: PATH must be narrowed to the shim dir for "no browser
installed" to be testable on a machine that has five, but every external the
code under test calls has to be seeded there too — a missing mkdir aborted
the suite under set -e with no stderr, which under kcov looked like a
coverage-tool bug rather than a PATH bug.
The hyphenated array subscripts in the test are quoted for the same reason
they are in the lib: shfmt -w rewrites an unquoted [google-chrome-stable]
into a subtraction expression. Reproduced live in this file during the work.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
485 -> 114 lines, with the six install phases moved into four sourced
libs: leechblock_fetch.sh, leechblock_config.sh, leechblock_browsers.sh
and leechblock_firefox.sh.
detect_browsers, replace_browser_in_place and wire_up_browsers stay in
one lib on purpose: detect_browsers writes the BROWSERS/FIREFOXES arrays
that wire_up_browsers reads, and splitting them would create a
write-only cross-file global.
Four justified body changes, all recorded in the evidence:
- Quoted the six hyphenated associative-array keys. shfmt 3.13.1 parses
an unquoted [google-chrome-stable] as arithmetic and rewrites it to
[google - chrome - stable], which would break browser detection
entirely. The pre-commit hook runs shfmt -w on staged files and
re-stages them, so this corruption would have landed in the commit
unseen -- this is what caused the previous attempt at this split to be
reverted. Behaviour-preserving: quoted and unquoted are the same key.
- SCRIPT_DIR is assigned in the entry script rather than inside
inject_default_config. Left in the lib, ${BASH_SOURCE[0]} resolves to
lib/, so the defaults JSON lookup would miss and default blocking
rules would silently never be seeded.
- install_firefox_policy takes ff_found as a parameter instead of
reading it as a cross-file global, fixing a genuine SC2154 without a
suppression.
- Added require_cmd unzip. It was used bare under set -Eeuo pipefail, so
a missing unzip would abort after the version dir was populated but
before the 'current' symlink was made, leaving a half-installed tree.
verify_shell_split.sh reports additions only -- no -/+ pair -- so
get_latest_tag and replace_browser_in_place moved byte-identically.
--help and unrecognized-flag output are byte-identical to the baseline.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The three safe-tier files are done; all five remaining are deployment-trap
tier. Records what the last session actually learned rather than restating
the plan:
- the shfmt bug that killed the install_leechblock attempt (it silently
rewrites `[google-chrome-stable]` as `[google - chrome - stable]` in an
associative array, which would have broken browser detection outright), plus
the grep that catches it;
- the phase line-ranges and lib layout for install_leechblock that were
already verified byte-identical, so the next session can redo it quickly
rather than rediscovering them;
- the six-step method all three completed splits used, including the rule that
you never run an installer just to capture a baseline;
- ten sharp edges that each cost real time, from subshells stranding state to
kcov mis-attributing brace-group closers;
- the three latent defects fixed along the way, and the note that one of them
leaves two phantom entries in this machine's live unblock state file which
the next install will self-heal.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
913 lines down to 150. The ten functions were only 202 lines; the other 712
were a single top-level block, now eight phase functions across six libs. As
in the previous split the bodies were wrapped WITHOUT re-indenting, because
five heredocs use quoted terminators that must stay at column 0 (shfmt
indented them afterwards, which is safe).
The ~320-line custom-blocklist heredoc became custom_entries.hosts. It was
321 of the 370 lines of hosts_write.sh and cannot be split, so that lib could
not otherwise fit the cap; a hosts-format data file is also far easier to
audit and diff than a quoted heredoc.
That move broke a guard, which is the important part of this commit:
extract_custom_entries_from_script sliced the blocklist out of the script
between "# Custom blocking entries" and the heredoc terminator. With the
entries in a data file there is no terminator, so it would have extracted
NOTHING -- and check_custom_entries_protection, the guard whose entire job
is refusing to drop blocked domains, would have compared against an empty
list and passed everything. It now reads the data file, and the extracted
set is identical: 254 domains, diffed.
A second, pre-existing bug surfaced while testing:
extract_unblock_entries_from_script matched its OWN marker lines.
PROTECTED_UNBLOCK_LIST_START and _END are word characters, so the
'^# [a-zA-Z0-9._-]+$' pattern pulled both into the whitelist as if they
were domains. Both are in this machine's live state file right now.
Renaming either marker would have read as the whitelist growing and blocked
the install outright.
There is no dry-run or status mode, so this file cannot be proven by running
it: an aborted run leaves /etc/hosts mutable with the guard's bind mount down.
The proof is textual instead. Every one of the 8 wrapped phase bodies is
byte-identical to its original line range; a line-multiset check found 0 lines
lost; and an indentation-insensitive comparison against the pre-split file
leaves exactly 8 absent lines, each an intentional edit (the rewritten
extractor, three `readlink -f "$0"` made overridable, the /etc/hosts fallback
made overridable, and the heredoc markers). /etc/hosts is byte-unchanged
across the whole session, and `--help` output is identical.
36 assertions. Both anti-tamper guards at 100% line coverage (75/75 and
49/49), covering both directions of each: removing a blocked domain is
refused and adding one allowed; growing the whitelist is refused and shrinking
it allowed. A same-count swap is refused too, which a count-only check would
have let through.
The install phases themselves are deliberately untested -- running them would
take down the protection they install. That is weaker evidence than the two
previous splits had, and it is inherent to this file rather than a gap.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
1017 lines down to 148. Six libs: the per-language doc-URL builders (split in
two only for the cap), the lookup/dispatch layer, and one per generation phase.
The hard part was not the ten functions but the 335-line top-level block,
which extract_shell_functions.py cannot move because it only relocates
functions. It is now generate_doc_links / generate_anki_cards /
generate_llm_prompt, wrapped WITHOUT re-indenting the bodies: five heredocs
use quoted terminators, which must stay at column 0.
Proof the split is behaviour-preserving: a golden fixture captured before any
edit, re-checked after every step. All three outputs -- anki_cards.txt,
documentation_links.md, llm_anki_prompt.md -- stay byte-identical by sha256sum,
and stdout matches too.
Four latent defects surfaced while writing the tests, all fixed:
python_doc_url required a $2 it never reads, and get_doc_url has always
called it with one argument. That only worked because the entry script runs
`set -e` without `-u`, so "$2" silently expanded to empty; any caller with
`set -u` aborted. Defaulted, along with three more positionals documented as
optional but declared required.
The two imports loops hardcoded `head -20` while the other seven honoured
TOP_N, so --top was silently ignored for imports.
Only one of nine term loops skipped comment lines, so a '#' line in an
analysis file became a flashcard whose front was a comment.
Introduced and fixed within this change: the comment filter has to run
BEFORE `head -n` or comment lines count against the limit, and it needs
`|| true`, since grep exits 1 when it filters every line and the callers run
under `set -e`.
The golden fixture cannot show the last three, because it has no comments and
fewer than 20 imports -- which is exactly why they survived. A second fixture
with 25 imports and comment lines, at --top 3, emitted 20 imports and two
comment cards before, and 3 imports with zero leakage after.
166 assertions across 4 test files. 100% line coverage on study_doc_urls.sh
(38/38), study_doc_urls_more.sh (30/30), study_doc_lookup.sh (49/49) and
study_gen_llm.sh (54/54). study_gen_docs.sh 93.10% and study_gen_anki.sh
96.15%, where every remaining line is a `} >>"$FILE"` closing a brace group
whose output a passing test asserts -- the same kcov attribution artifact
recorded for services_hosts.sh.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
Eight checks assigned status="warning" unconditionally, so a check running
later in the same function overwrote the status="error" an earlier one had
set. The five other sites in this family already guard the assignment with
`if [[ $status != "error" ]]`, so this was an inconsistency, not a design.
Only half of it was load-bearing, and the two halves fail differently:
services_units.sh (4 sites) reports through report_and_fix, which repairs
ONLY on "error". A downgrade there meant no repair ran at all. The is-active
check is the LAST status assignment in each of the three functions, so it
could mask any error above it -- in check_periodic_systems, three separate
is-enabled checks. A timer that was both disabled and inactive, which is the
ordinary shape of a broken timer, was reported and then left broken.
services_hosts.sh (4 sites) gates its own repair on `status != "ok"`, so the
repair still ran. There the bug printed a real error as a yellow [!] warning
and understated the summary row.
Demonstrated rather than argued, against a simulated broken machine (timer
disabled AND inactive, unit files present):
without the guards: status=warning, repair ran 0 times
with the guards: status=error, repair ran 1 time
The seven assertions that previously pinned the broken behaviour now assert
the correct behaviour; they were rewritten first and failed, then the guards
made them pass. Two further cases cover the masking shape directly: a broken
guard-lib instance must still be repaired when a later check warns.
`--status --dry-run` output is still byte-identical to rev 6691cb26, because
this machine has no disabled timer and no missing pacman hook -- the buggy
path is never taken here, which is precisely why it went unnoticed. The A/B
above is the evidence the live diff cannot provide.
222 assertions across 10 files. services_units.sh holds 100% coverage;
services_hosts.sh rises from 86.05% to 93.02%. test_services_hosts.sh needed
splitting twice to stay under the 250-line cap (hosts_fix, resolved), and
stage_hosts_ok moved into the harness so both new files can use it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
Follow-up to 9085a4e5, which recorded services_units.sh at 84.51% because
its remaining branches sit behind [[ -f /etc/systemd/system/*.service ]] --
absolute root-owned paths that exist on this machine, so the else-branch
never runs and no PATH shim can redirect it.
All 87 such paths across the seven check libs now go through $SYSROOT, read
from SERVICES_ROOT. The prefix is deliberately NOT defaulted. Three commands
in hosts_fix_resolved run outside `run` and so are not stopped by --dry-run
(chattr -i on resolved.conf.d, find -delete over its *.conf, and an append to
resolved.conf); with an empty default, a test that forgot to set SERVICES_ROOT
would really delete this machine's resolved.conf drop-ins. Unset is therefore
a hard error and only an explicitly empty value means the real filesystem.
The harness now owns PATH outright: $FAKE_BIN for shims and staged fake apps,
$REAL_BIN for a curated set of symlinks to genuine coreutils. The machine's
/usr/bin is deliberately absent -- a really-installed discord and chromium
otherwise make the "no target apps present" branches untestable, which is
exactly what happened until these tests caught it.
Coverage, measured against run_all.sh because two test files jointly cover
services_common.sh after the 250-cap split: common 55/55, report 27/27,
units 72/72 (was 84.51%), wrappers 100/100, apps 66/66, browser 79/79 --
seven libs at 100%. services_hosts.sh reads 86.05% and services_hosts_fix.sh
97.30%, and those 14 lines are a kcov attribution artifact rather than a gap.
Falsified rather than assumed: the suite prints "51 passed, 0 failed" when run
under kcov, so no group is skipped, and every supposedly-uncovered line is an
issues+=("...") whose string can only reach stdout through the print loop
below it -- with a passing assertion matching that exact string.
Two falsification checks, both re-run after the rewrite: `--status --dry-run`
as root stays byte-identical to rev 6691cb26 modulo the Date: line, and the
sha256sums of /etc/hosts, /etc/nsswitch.conf and /etc/systemd/resolved.conf
plus the contents of resolved.conf.d are unchanged across a full suite run.
The second one is what makes a missed prefix site detectable instead of a
matter of inspection: the repair paths really do write.
220 assertions across 8 test files. Also fixes a bug in the harness's own
lsattr shim, where printf with a leading-dash format string parsed the dashes
as options and silently produced nothing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
1225 lines down to 202, with the 21 functions moved into eight sourced
libraries grouped by subject. The entry script now holds only the
configuration, the argument parsing and the order the checks run in.
check_hosts was the one function that could not move verbatim: at 260 lines
it would have stayed over the cap on its own, so its repair half is now
hosts_fix_nsswitch / hosts_fix_resolved / hosts_repair_all in
services_hosts_fix.sh. Repair order is preserved (resolver config, then the
hosts file, then the guard-lib migration), and hosts_repair_all returns
non-zero under --dry-run so a dry run can no longer promote a status to "ok"
on the strength of repairs that never happened.
SERVICE_STATUS was a write-only global across five libs, which is a real
SC2034 rather than noise: the pre-commit hook runs shellcheck with no -x and
genuinely cannot see the one reader. It now goes through set_service_status /
get_service_status, which also rejects a duplicate write — each check owns
exactly one summary row, so a second write means two checks are fighting over
it. No suppression was added anywhere.
Proof the behaviour is unchanged: `--status --dry-run` run as root against
this machine is byte-identical to the same command from a worktree at the
pre-split revision, modulo the Date: line. verify_shell_split.sh reports
check_hosts as the single expected difference; the other 20 functions hash
identically. 72 assertions across three test files pass, with 100% line
coverage on services_common.sh (54/54) and services_report.sh (27/27).
services_units.sh measures 84.51% (60/71). The 11 uncovered lines are the
else-branches of [[ -f /etc/systemd/system/*.service ]] and
[[ -f /usr/local/bin/*.sh ]] probes — absolute root-owned paths that exist on
this machine and that no PATH shim can redirect. Reaching them needs a
root-prefix rewrite of live enforcement logic, which is not a line-count fix
and would forfeit the verbatim-move proof for everything else.
Found while testing, deliberately NOT fixed here: in all three unit checks
the is-active branch assigns status="warning" unconditionally, overwriting
the status="error" the is-enabled branch set just above it. A timer that is
both disabled and inactive therefore never gets repaired, because
report_and_fix only repairs on "error". It is identical in the pre-split
file, so this split neither caused nor hides it. The tests pin the current
behaviour and name it; fixing it changes enforcement and earns its own commit.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F3PzGc75gBcdxP3PkHXGcV
Rewrites the driver doc as a self-contained fresh-session prompt:
supersedes the one-file-per-session pacing decision (user asked for
batching), adds an explicit "do not ask should I continue" preamble,
orders the remaining 8 shell files into a safe tier and a
deployment-trap tier, and points at the pacman_hook_stall_* family as
the worked reference split with harness + entry-harness pattern.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Records the three focus_owner splits (status_page_state.dart,
DevicePolicyBridge.kt, EnforcementRunner.kt) and the reusable Kotlin
split pattern -- internal helper classes as private delegate fields,
public API preserved as one-line wrappers, function references for
parent callbacks. Also flags the companion-object-called-by-test trap
and that everything remaining is shell, 4 of 8 deployment-trapped.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Extracts four cohesive concerns into standalone files:
LocationAcquisition (fix acquisition), HomeLocationStore (home
read/write), PolicyPinning (VPN/DNS pin + visibility sweep),
SweepablePackages (installed-package enumeration). Public API
unchanged -- EnforcementRunner.best(...) deliberately stays put since
EnforcementRunnerBestTest.kt calls it directly; every other public
method is a one-line delegating wrapper.
Verified with gradlew compileDebugKotlin + test --rerun-tasks
(BUILD SUCCESSFUL, 40 tests / 0 failures, including the 6
EnforcementRunnerBestTest cases unmodified).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Kotlin has no part-of like Dart, so this extracts three cohesive
method groups into standalone internal helper classes held as
private delegates: DevicePolicyVpnDns (VPN/Private DNS pinning),
DevicePolicyLocation (location permissions), DevicePolicyUninstallGuard
(uninstall blocking). DevicePolicyBridge's public API is unchanged --
every method MainActivity.kt/EnforcementService.kt call directly
still exists as a one-line delegating wrapper.
Verified with gradlew compileDebugKotlin + test --rerun-tasks
(BUILD SUCCESSFUL, 40 tests / 0 failures).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Marks focus_owner/lib/status_page_state.dart done (commit 47e05d1),
confirms focus_owner is live so it's off the dead-code-candidate list,
documents the heredoc-generated-script trap in
setup_midnight_shutdown.sh (breaks extract_shell_functions.py and
verify_shell_split.sh as-is, needs a hash-the-emitted-script
verification plan instead), and logs a dart_format_changed.sh path
bug hit this session (cd's into repo before git status, whose paths
are worktree-root-relative -- silently formats nothing).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Extracts the two AlertDialog builders and the build() body-selection
logic into a new status_page_dialogs.dart part file, following the
same part-of-main.dart split pattern status_body.dart already used.
Pure extraction, no behavior change -- verified with dart analyze and
the full flutter test suite (94/94 passing).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Pre-commit checks, Shell tests, and Python tests are all green on
4a35be7. Documents the fix (librsvg2-bin) and the gh-run-rerun trap
hit along the way so a future session doesn't repeat it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
The rsvg-convert fix landed in a commit that only touched workflow
files, which python-tests.yml's path filter doesn't watch -- so the
workflow never re-ran and gh run rerun replays the old, pre-fix
definition instead of the current one. Adding the workflow's own
path to its filter and workflow_dispatch closes both gaps.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
ubuntu-latest lacks librsvg2-bin, so test_render_writes_png was
skipped in CI, leaving vector.py:156 uncovered and both Pre-commit
checks and Python tests red at 99.98% (1 line miss). Installing the
real dependency lets the test run instead of excluding the line.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Rewritten as a 4-phase recurring driver per user request: (0) fix current
CI red state — found Pre-commit checks AND Python tests both failing with
an identical, pre-existing coverage gap (python_pkg/artgate/routes/vector.py:156,
99.98% total, rsvg-convert missing on the GitHub Actions runner) that
predates this session by at least a day; Shell tests was also red from
this session's own archive commit, already fixed in e48f253 and confirmed
green — (1) drive the 11 remaining over-cap files to zero, same
dead-code-check-in-first pattern as every prior split — (2) only then wire
check_file_length.sh into pre-commit as a hard gate, plus a CI-green check
(mechanism deliberately left as an open question for the user: check the
latest main run, or extend the pre-push ci-mirror hook) — (3) a new,
separately-scoped repo-wide 100% shell coverage gate, measured at 487 .sh
files against only 14 existing lib/tests/ directories, explicitly flagged
as needing its own user confirmation and ratchet-vs-hard-gate decision
before starting given the scale.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Regression from this session's earlier commit (327d48b): the test still
referenced pacman-hooks/*.sh and install_pacman_hooks.sh, which that commit
archived to testsAndMisc-archive. shell-tests.yml runs this test directly
(not scoped by pre-commit's --files), so it went red on CI without being
caught locally before the push.
Dropped the assertions tied to the archived files; kept and extended the
ones checking pacman_wrapper.sh's guard-lib fallback logic, which was
already fully guard-lib-native and unaffected by the archive.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
setup_hosts_guard.sh was NEXT_SESSION_PROMPT.md's next split target, but
live-machine inspection showed it (and 16 supporting files - enforce-*.sh,
6 systemd units, install_pacman_hooks.sh, pacman-hooks/, psychological/,
both READMEs) is fully retired in favor of guard-lib, migrated 2026-07-04.
User confirmed archive over split, then confirmed the larger 17-file scope
once the directory was surveyed. Full history preserved via git filter-repo,
live at github.com/kuhyx/testsAndMisc-archive/tree/main/linux_configuration/scripts/periodic_background/hosts/guard.
plugins/{nsswitch,resolved}-plugin.sh stay - they're guard-lib's live
plugins, referenced by absolute path from /etc/guard-lib/targets/*.conf.
check_and_enable_services.sh's check_hosts() rewritten to check guard-lib
state (guardctl file-guard status) instead of the retired units/scripts it
used to check and self-repair - verified live, now reports the hosts
service ok against the actual migrated machine instead of a false error.
copilot-instructions.md and one install.sh comment updated to match.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Confirmed live (not dead code) via the dead-code check-in before any code
was touched -- the last two candidates this pattern was tried on were both
archived instead. Kept run_one/cleanup/log_size/main together in the entry
script: a prior session (2527fb3) already proved splitting run_one out
produces a real SC2153/SC2034 pair since PACMAN_PID/LAST_ELAPSED cross the
EXIT trap, not a suppressible false positive.
Fixed a real bug found while writing tests: print_summary's inline awk had
a dead `NR == 0` guard (printf on an empty array still emits one blank
line), reachable via `-n 0`. Refactored mem_available_mb, descendant_pids/
kill_tree, capture_stall's two output blocks, and print_summary's median
calc after confirming empirically (bash -x tracing) that kcov's
xtrace-based coverage can never see a `done <redirect>`, `} >redirect`, or
inline-awk line -- user chose to fix the code over the coverage tool.
descendant_pids specifically needed an explicit-fd process substitution
(not a pipe) to keep its array mutation in the current shell.
Made PACMAN_BIN/PACMAN_LOG/CACHE_DIR/PACMAN_LOCK/LOAD_FLOOR_MB/
LOAD_MIN_FREE_MB env-overridable for testing. Caught and fixed, before it
ran unbounded, a --with-load test that would have dd'd 13-20+ GB into
/dev/shm using this box's real available memory.
56 lib-level + 13 entry-level PATH-shim/subprocess assertions, 0 failed.
The entry test itself grew past 250 lines while adding coverage and had to
be split into a shared harness plus two test files, mid-session.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Evaluates github.com/RyanAlberts/best-of-Agent-Harnesses (160 projects)
against the existing ~/.claude/ setup: hooks, skills, session-autopsy
pipeline, and prior independent arrival at the token-spend thesis.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Full history preserved via git filter-repo, including both pre-reorg
root paths (libre_translate.sh and the subtree-imported
Bash/libre_translate.sh) since both were genuine ancestors of the
current file. Live at
github.com/kuhyx/testsAndMisc-archive/tree/main/linux_configuration/scripts/single_use/misc/testsAndMisc-bash.
Asked the dead-code check-in before any split work started this time,
per the lesson from the plagiarism-tools session. NEXT_SESSION_PROMPT.md
now points at diagnose_pacman_hook_stall.sh.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Full history preserved via git filter-repo, including the pre-rename path
from before the single_use/periodic_background split. Live at
github.com/kuhyx/testsAndMisc-archive/tree/main/linux_configuration/scripts/single_use/utils.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
Wraps the four install stages into named functions, one per lib/plagiarism_*.sh
sibling (matching install_exercism.sh's convention), and extracts the two
embedded Python heredocs into real, lint-clean check_plagiarism.py/sherlock.py
files -- check_plagiarism.py needed a second split into plagiarism_similarity.py
to also clear the 250-line cap. Every function reaches 100% shell coverage via
a PATH-shim test harness (fake python3/pip/git/java/wine); the two Python CLIs
were separately verified against a real venv with actual nltk/scikit-learn.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
