testsAndMisc/NEXT_SESSION_PROMPT.md
Krzysztof kuhy Rudnicki 53e80462c5 Phase 3: prove 100% shell coverage is reachable via a namespace jail
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
2026-08-21 17:08:05 +02:00

9.5 KiB

Next session: Phase 3 — finish wiring the shell-coverage ratchet

Paste this whole file into a fresh Claude session opened in ~/testsAndMisc. It is self-contained. Do not go looking for the previous session's context.

Phases 1 and 2 are DONE. The 250-line cap is at zero and enforced; the CI gates are wired.

Phase 3 is BUILT but NOT ENFORCED (commit 3278bebd). The three scoping questions below were answered without the user, because the session that built it was non-interactive — a grilling round would have ended the turn with nothing delivered. The user can still veto any of these cheaply, since the gate is not yet in the hook chain:

  • Q1 Scope → ratchet. 115 pre-existing libraries are exempt via meta/shell-coverage-allowlist.txt; only new or modified libraries must be covered. Unrelated commits are never blocked (verified).
  • Q2 Bar → presence, not a percentage. A numeric bar is unreachable here: Phase 1 measured the effecting code at 0% because it sudos, pkills and installs, so hitting a number needs either mass shimming or the suppressions this repo forbids. The gate asserts a tests/run_all.sh exists beside the lib.
  • Q3 Which files → .sh under a lib/ dir, excluding lib/tests/ and lib/payloads/. Entry scripts are orchestration whose bodies are untestable.

The one thing left to do

meta/scripts/check_shell_coverage.sh works standalone but is not in .pre-commit-config.yaml — that edit was denied as a sensitive file. Add this block after the file-length-cap hook (note files:, not --all; per-file is what makes it a ratchet):

- id: shell-coverage-ratchet
  name: Shell libraries must have a test suite
  entry: bash meta/scripts/check_shell_coverage.sh
  language: system
  files: \.sh$

Then re-verify the negative case before trusting it: bash meta/scripts/check_shell_coverage.sh meta/scripts/check_file_length.sh must exit 0 (an unrelated entry script must never block a commit).

Do not run --seed after adding an untested library — it would rewrite the allowlist and silently exempt it. The list is shrink-only; nothing mechanically enforces that yet.

Chipping away at the 115

198 in-scope libraries: 115 exempt, 83 already covered. The allowlist is concentrated: single_use/features/lib (45), single_use/fixes/lib (22), phone_focus_mode/lib (9), scripts/lib (9). Adding one tests/run_all.sh to a directory enforces every file in it at once, so expect the work to arrive in batches rather than one file at a time.

--seed is shrink-only and now enforces it — it exits 1 rather than adding an entry, so you cannot exempt a new library by reseeding. It reads tracked files only, so stage before trusting its output.

What is already true (verify, do not redo)

bash meta/scripts/check_file_length.sh --all   # "all checked files are within 250 lines"
bash meta/scripts/check_ci_green.sh            # exit 0 while main is green
gh run list --limit 5
  • Every file is under 250 lines. The five deployment-trap files were split: install_leechblock.sh (485→119), block_compulsive_opening.sh (705→182), setup_night_lockdown.sh (918→127), pacman_wrapper.sh (929→217), setup_midnight_shutdown.sh (1734→109), plus _glyph_art.py (253→208).
  • file-length-cap runs check_file_length.sh --all on every commit.
  • ci-baseline-green refuses to commit onto a red main (fails closed on red, warns and passes when it cannot tell). Bypass: CI_GREEN_SKIP=1.
  • ci-mirror on pre-push now also runs every */lib/tests/run_all.sh and the side-effect-free half of shell-tests.yml.
  • shell-tests.yml discovers and runs every */lib/tests/run_all.sh. Four such suites existed before and none of them ran in CI until this was added.

Measuring the scale — use git ls-files, not find

git ls-files '*.sh' | wc -l                          # 560 tracked
git ls-files '*/lib/tests/run_all.sh'                # 4 suites
bash meta/scripts/check_shell_coverage.sh --all      # 115 uncovered, 115 exempt

A bare find reports 565 .sh files because it walks .venv/, .ci-mirror-venv/ and node_modules/. The earlier "roughly 487" figure in this file came from an unfiltered walk of a different tree state. Seeding an allowlist from an unfiltered find would bake vendored venv scripts into a tracked file, and check_file_length.sh --all passing would not reveal it, because that script carries its own separate exclusion list.

What Phase 1 measured — why the bar is presence, not a percentage

Coverage of the new libs, measured with bash meta/scripts/shell_coverage.sh <lib/tests/run_all.sh> <lib-basename> 0:

lib coverage what is uncovered
cco_state.sh 93.94% (62/66) only kill_app's two pkill lines
leechblock_fetch.sh 38.57% (27/70) download_extension: curl/tar/unzip
leechblock_browsers.sh 36.47% (31/85) replace_browser_in_place: sudo binary patching
cco_wrapper/install/report, leechblock_config/firefox 0% every function pkills, sudo-writes or installs

The pattern: the decision logic reaches ~90%+, the effecting code reaches 0%. A 100% bar means shimming sudo, pkill, systemctl, curl and npm for every installer in the repo. That is the real cost to put in front of the user.

The harness that works — extend it, do not clone it

linux_configuration/scripts/periodic_background/digital_wellbeing/lib/tests/ holds leechblock_harness.sh and cco_harness.sh plus a run_all.sh whose glob is test_*.sh. One harness per directory, not per file — jscpd fails above 2% duplication and five near-identical harnesses is exactly how that trips.

Two patterns worth copying:

  • _t_isolate_path narrows PATH to a shim dir. It is the only way to make command -v jq or command -v chromium fail on a machine that has them. Every external the code under test calls must be seeded into that dir — a missing mkdir aborted the suite under set -e with no stderr, which under kcov looked like a coverage-tool bug for twenty minutes.
  • A fixture self-check. cco_harness.sh asserts its own globals are populated and calls notify for real. This is not linter appeasement: it is what turns "a typo in a global name" from a silent skipped assertion into a loud failure, and it happens to satisfy SC2034 honestly.

Traps that cost real time in Phase 1 — all still live

  • shfmt -w corrupts [hyphenated-keys] in associative arrays, turning ${BROWSERS[google-chrome-stable]} into a subtraction expression. Quote the subscripts. Reproduced live in a test file during this campaign. Canary: grep -nE '\[[a-z0-9]+ - ' <file> must be empty after every shfmt and every pre-commit run.
  • A comment line starting with # shellcheck is parsed as a directive and fails with SC1072/SC1073. Write "the linter", not "shellcheck", at the start of a comment line.
  • verify_shell_split.sh cannot see heredocs. In setup_midnight_shutdown.sh it reported a bare -log, because log() sits at column 0 inside a generated script. For any file that emits scripts, the real check is hashing the emitted content, not the function list.
  • New sourced libs need a shebang AND the exec bit, staged with git add --chmod=+x. A plain git add afterwards resets it. Payload files under lib/payloads/ are 100755 too — the shebang hook rejects 100644.
  • gh and git push race. Several pushes failed with "cannot lock ref"; git fetch then re-push. Not a divergence.

Known-broken, NOT caused by the campaign

Python tests has been failing since 2026-08-17 — before any of this work. The cause is pip hitting error: resolution-too-deep while installing meta/requirements.txt, on commits that touch no Python at all. It needs dependency pinning, and it is why python-tests.yml is deliberately not in check_ci_green.sh's required list. Worth fixing on its own; do not fold it into Phase 3.

Rules that will bite you

  • No suppressions, ever. No # noqa, no shellcheck disable, no per-file-ignore without asking every single time.
  • Every commit touching code needs evidence in docs/superpowers/evidence/<slug>-<date>.json, plus a contract in docs/superpowers/contracts/ once ≥4 code files are staged. Put the measured number in it, never a rounded or hoped-for one.
  • A test file broken by your commit is a same-session bug. Three were broken this way in Phase 1 (security_apps.sh, test_shutdown_timer_monitor.sh, and the pacman pair) — each asserted on code a split had moved. Fix the assertion to search the entry script and its libs; do not weaken it.
  • pre-commit run --files <changed> after git add. prettier and ci-mirror run on pre-push. npx prettier --write any .md/.json.
  • Work directly on main. git stash and branch creation are blocked.
  • Verify what actually got committed. Re-grep the file after committing; an autoformat pass stripped a load-bearing variable definition three times in Phase 1, once landing a set -u abort in setup_night_lockdown.sh that had to be fixed in a follow-up commit.