testsAndMisc/NEXT_SESSION_PROMPT.md

STATUS 2026-08-23 (later session): jobs 1 and 2 are DONE

Read this box before anything below it. The rest of this file is the ORIGINAL handoff, kept because its kcov/jail notes are still the reference for measuring a lib on purpose. Its SCOPE is now obsolete.

Job 1 — the push gate — DONE (commit 8ab15446). ~130s to: docs-only 7.6s, python 25.3s, one shell suite 43.6s, unchanged-tree re-push 0.005s. The guarantee is intact; a broken python assertion and a broken shell assertion were each refused with exit 1.

Found while measuring it: ci_mirror.sh called pytest_changed_packages.py with no arguments, and that script returns 0 on empty argv. The pre-push python stage had never run a single test. Behind it sat a real failing test (fixed in ff3ea99a).

Job 2 — the 100% shell bar — DELIBERATELY DROPPED (commit 578c4a6c). The user's call on 2026-08-23, after being shown the real cost. The bar had been raised from a presence check only the day before (bca70b67), and pricing it out honestly: one hand-written test file per lib (fixes/lib took 34 files for 22 libs), a serial jail at 37-90s a pass, and a production seam needed in most libs before they are testable at all — several days for a repo of personal automation scripts. is_covered is a presence check again.

The ratchet is now WIRED into pre-commit (commit 46c8f065) — it never was before, which is why it rotted. 0 BLOCK entries, 0.16s repo-wide. The three meta/scripts/lib coverage-tooling libs got a 48-assertion suite so they stop blocking it.

Python needs nothing: 2101 tests, 100.00%, zero missing statements or branches, measured in a clean worktree. File-length cap needs nothing: 0 violations, hook-enforced.

What is left, if anyone wants it: 8 directories still have no suite at all (.githooks/lib, meta/lib, mtk_root/lib, digital_wellbeing/{pacman, virtualbox}/lib, system-maintenance/bin/lib, single_use/{fresh-install,}/lib) — 20 libs, ~1,980 lines. All are allowlisted, so nothing is blocked. This is optional work now, not a campaign.

Traps worth carrying forward, beyond the list below:

• Never git reset --hard to undo a probe commit. It discards UNSTAGED working-tree edits. That cost a full rewrite of ci_mirror.sh this session. Recovery: pre-commit stashes unstaged files to ~/.cache/pre-commit/patch*; git apply on the newest one restored it exactly. Use git revert --no-commit <sha>.
• Read the predicate, not the header comment. check_shell_coverage.sh's header still described a presence gate for a full day after the gate became a 100% bar. Acting on the prose produced a confidently wrong scope estimate.
• A sourced-only lib takes # shellcheck shell=bash, no shebang, no exec bit, or check-shebang-scripts-are-executable fails the commit.

---

Next session: make the push gate fast, then close the coverage holes

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.

There are two jobs here. Do them in order — job 1 is small and makes job 2 bearable, because job 2 involves many commits.

---

JOB 1: get git push under 10 seconds

What is slow

git push currently takes ~130 seconds. Nearly all of it is the ci-mirror pre-push hook (meta/scripts/ci_mirror.sh), which deliberately re-runs CI locally so a red push cannot reach GitHub. It does four things inside a fresh git worktree checkout of HEAD:

stage	what it costs
venv install from requirements.txt	~0s when cached, ~40s on a change
pre-commit run --all-files	the bulk of it
pytest_changed_packages.py in that venv	seconds to minutes, package-dependent
run_shell_test_gate (the shell suites)	~40s+

For comparison: pre-commit is already fast — 4.8s measured. Do not touch it. The whole problem is pre-push.

The target, and the honest tradeoff

The user asked for under 10 seconds, ideally ~10s or less, total.

Be straight with them about this: a clean-venv install plus --all-files plus two test suites cannot fit in 10s. Something must actually be given up. The decision is theirs, so measure first, then present the options and their real costs, then implement the one they pick.

Step 1 — measure each stage separately

Do this before proposing anything. Numbers, not guesses:

cd /home/kuhy/testsAndMisc
SP=$(mktemp -d); git worktree add --detach --quiet "$SP/co" HEAD

time (cd "$SP/co" && SKIP=pytest-coverage pre-commit run --all-files)
time (cd "$SP/co" && python meta/scripts/pytest_changed_packages.py)
time bash linux_configuration/fixes/lib/tests/run_all.sh

git worktree remove --force "$SP/co"

Then find which individual pre-commit hooks dominate --all-files:

cd "$SP/co" && for h in $(grep -oP '(?<=id: ).*' .pre-commit-config.yaml | sort -u); do
  s=$(date +%s%N); SKIP='' pre-commit run "$h" --all-files >/dev/null 2>&1
  printf '%6d ms  %s\n' $(( ($(date +%s%N)-s)/1000000 )) "$h"
done | sort -rn | head -15

Step 2 — the options, in the order I would try them

1. Only check what changed. The single biggest win and the least loss. --all-files is the expensive part; pre-commit run --from-ref origin/main --to-ref HEAD checks the pushed range instead. The clean worktree already guarantees you are checking committed content. Similarly, only run a shell suite when the push touches shell files.
2. Parallelise the stages. pre-commit --all-files, pytest and the shell gate are independent. Run them concurrently and wait. Roughly the cost of the slowest rather than the sum. Note: two coverage jails must never run at once (see the warning in job 2), so if the shell gate is jailed, keep it serial with respect to any other jail.
3. Cache on tree state. Skip the whole mirror when git rev-parse HEAD: (the tree hash) matches the last green run, recorded in a dotfile. A re-push of an unchanged tree then costs ~0s.
4. Move ci-mirror off pre-push entirely. Pre-push drops to ~5s. The cost is real and must be stated plainly: a red push can then reach GitHub, and CI catches it minutes later instead of the hook catching it instantly. The repo's refuse-red-ci hook already blocks the NEXT commit on a red baseline, so this is less dangerous than it sounds — but it is a genuine change in guarantees, not a free win.

Options 1-3 are optimisations that keep the guarantee. Option 4 trades the guarantee for speed. 1+2+3 together may well reach ~10s; try them before proposing 4. If they land at, say, 25s, say so honestly rather than reaching for 4 without asking.

Step 3 — prove it

time git push        # the number that matters

And prove the gate still catches things: deliberately break something (a failing assertion in a shell test), confirm the push is refused, then revert. A fast gate that no longer gates is worse than a slow one.

Do not

• Do not weaken or delete a test to make a gate faster.
• Do not add --no-verify anywhere, or suggest it. It is forbidden by the user's rules.
• Do not touch pre-commit's 4.8s — it is already at target.

---

JOB 2: actually get the repo to 100% coverage

The user asked to "make the repo truly 100% test covered". Be careful with that phrasing: it is a large campaign, not one task. Scope it, report the real size, and then work through it. Do not report it done until it is.

Where things actually stand

Shell: meta/shell-coverage-allowlist.txt has 84 exempt libs

grep -vc '^#' meta/shell-coverage-allowlist.txt     # 84

Each line is a shell library allowed to sit below 100% line coverage. The just-finished campaign took this from 91 to 84 by clearing seven libs in fixes/lib. The remaining 84 are the job.

~12 gated libs OUTSIDE fixes/lib fail the gate RIGHT NOW

check_shell_coverage.sh is not wired into pre-commit, so these have been failing silently. Confirmed pre-existing (verified against an earlier commit — they are not caused by recent work):

lib/common_packages.sh                                        FAIL
periodic_background/digital_wellbeing/lib/ms_override_mgr.sh  FAIL
periodic_background/digital_wellbeing/lib/ms_scripts.sh       FAIL
periodic_background/lib/periodic_browser.sh                   FAIL
single_use/utils/lib/analyze_repo_symbols.sh                  FAIL
single_use/utils/lib/bl9000_backup.sh                         FAIL
single_use/utils/lib/download_media.sh                        FAIL
single_use/utils/lib/idle_inhibit.sh                          FAIL
single_use/utils/lib/offline_index.sh                         FAIL
single_use/utils/lib/steam_api.sh                             FAIL
single_use/utils/lib/steam_report.sh                          FAIL

hosts_protect_custom.sh and hosts_protect_unblock.sh PASS.

Start here. These are worse than an exemption: they are libs the gate believes it is checking and which do not meet the bar. single_use/utils/lib/ is most of the list and shares one tests/ directory, so one harness covers several.

Python: python_pkg/ is 25 packages at fail_under = 100

Branch coverage is already enforced at 100% by meta/pyproject.toml, so this is likely in decent shape. Verify before assuming — and note the user's standing rule: never exclude a package from coverage, write the tests.

timeout 1800 python -m pytest python_pkg --cov=python_pkg --cov-report=term-missing -q 2>&1 | tail -30

Other languages

billsplit/ (Dart), reverse_survivors/, poker-stakes/ (TS) have their own CI. Check their gates before claiming anything about them.

Report the real scope first

Before writing a single test, produce a count: N shell libs exempt, M gated libs failing, Python status, other languages. Give the user that number and a suggested order. "100% covered" spans several days of work; they should see the shape before you spend it.

The order I would work in

1. The ~12 gated-but-failing shell libs (worst: the gate is lying today).
2. Wire check_shell_coverage.sh into pre-commit so it cannot rot again — but only after 1, or every commit is blocked. Mind job 1: this hook is jailed and slow, so scope it to changed files.
3. The 84 allowlist entries, cheapest first.
4. Python, if the run above shows gaps.

---

How to do shell coverage work here (read before starting job 2)

The suite and its harnesses

linux_configuration/fixes/lib/tests/ is the reference suite: 30 test files, 645 assertions, all green, zero suppressions, 37s. Copy its shape. Layers:

• run_all.sh — the runner, and the subject the gate measures through.
• lib_test_core.sh (194 lines) — assertions and stubs.
• lib_test_path.sh (67 lines) — _t_hide and the PATH shim farm.
• Six family harnesses, one per entry script.

Key primitives:

• _t_eq / _t_contains / _t_lacks
• _t_stub NAME BODY, plus _t_stub_stdin (quoted heredoc), _t_stub_cat, _t_stub_writes. Use these rather than _t_stub foo '...$1...', which trips SC2016 every time.
• _t_run FUNC [ARG...] — runs FUNC in the CURRENT shell, output in $out. Never out="$(func)": command substitution forks a subshell and kcov then does not register the sourced lib at all (measured: 0 lines via $( ), 54 via a direct call).
• _t_hide TOOL... — makes a tool genuinely unfindable. It KEEPS $FAKE_BIN on PATH, so _t_unstub the tool first or the harness's own default stub still satisfies has_cmd and the branch never runs.
• _t_unstub (runs hash -r; deleting the file alone does nothing).

Production seams

Most libs write to /etc. The idiom is an override defaulting to the real path, so production behaviour is byte-identical:

local dropin="${SYSCTL_DROPIN_DIR:-/etc/sysctl.d}/99-performance-tuning.conf"

Reuse existing names — SYSTEMD_UNIT_DIR, JOURNALD_CONF_DIR, UDEV_RULES_DIR, SYSCTL_DROPIN_DIR — and grep the entry script before picking one; a collision is silent.

Watch for globals ASSIGNED inside a function (XORG_CONF="/etc/..." with no local). Exporting from the harness does not survive the assignment; the default must live at the assignment itself.

After the first run, ls -la the real paths and confirm the mtimes did not move.

The measurement loop

# 1. write the tests, chmod +x
# 2. run bare first -- far faster than the jail
bash <suite>/run_all.sh

# 3. measure through the RUNNER, in a detached worktree
timeout 900s bash meta/scripts/shell_coverage_jail.sh --timeout 600s \
  --subject <suite>/run_all.sh --measure <lib>.sh --min 1 \
  --fail-on-case-error -- ""

Then the three-step bar to clear an exemption:

1. 100.00% through run_all.sh.
2. Again at --min 100, so a shortfall is a non-zero exit.
3. Delete the allowlist line, then run bash meta/scripts/check_shell_coverage.sh <lib> and confirm exit 0. Running it before deleting proves nothing — it passes via the exemption.

Traps that cost real hours last session

• Verify an edit landed. A seam edit reported success, was measured, and never reached the commit; ci-mirror caught it only because it runs from a clean checkout. After editing, grep the file for what you just wrote.
• Run the suite from a clean checkout before pushing. It catches exactly the above:

  SP=$(mktemp -d); git worktree add --detach "$SP/c" HEAD
  bash "$SP/c/<suite>/run_all.sh"; git worktree remove --force "$SP/c"
  

• NEVER run two coverage jails at once. Every false FAIL last session was CPU contention — a jail plus another jail, or a jail plus run_all.sh. Libs "failed" a sweep and passed solo. While a sweep runs, do zero-CPU work.
• Another agent may be committing to this repo concurrently. The jail fingerprints git status --porcelain and aborts with "a write escaped the jail" on THEIR edits. Measure in a detached worktree, and re-check a transient gate failure before believing it.
• Adding a test file can break an ALREADY-CLEARED lib. Unexplained: test_arch_sysctl.sh made kcov stop registering bt_audio.sh entirely ("instrumented no lines"), turning the gate red. Ruled out: export out, the _t_hide cache, and runtime (still fails at a 600s timeout). The file is quarantined as arch_sysctl_quarantined_cases.sh, outside the test_*.sh glob, with 23 still-green assertions. After adding any test file, re-measure an already-cleared lib — bt_audio.sh is the canary.
• shfmt reformats staged shell files, moving the denominator and invalidating a measurement. Run shfmt -l first and land reformats separately.
• The 250-line cap aborts commits, and a task notification can report "completed (exit code 0)" for a commit that actually failed. Check.
• Stage narrowly. A git add of a directory swept another agent's untracked file into a commit and broke the push.
• mapfile is bash-only; the tool shell is zsh. Wrap in bash -c.
• git stash is blocked — even git stash list inside a compound command trips the guard. Use a detached worktree.
• Do not pipe jail output through grep 'lines =' — that hides the warn: case timed out line, which is the diagnosis you need.
• Do not pgrep -f <thing> in a wait loop — the loop's own command line matches, so it never exits.

The coverage instrument

coverage = (PS4-traced | kcov hits) & (kcov line set - continuation lines)

continuation_lines() in meta/scripts/lib/shell_coverage_lines.py excludes lines kcov instruments but bash attributes no statement to: multi-line quoted arguments, multi-line array literals, } | / } >>file, and (added last session) done >file / done <file.

If you touch it: the acceptance pair is dot_resolver_install.sh = 39/39 and rpi_nc_ca.sh = 73/73 through features/lib/tests/run_all.sh with its jail_args file. And audit any new predicate by printing every newly excluded line with its text, across the WHOLE repo — scoping that audit to one directory last session undercounted the blast radius by 35x (2 lines vs ~75).

A lib already at 100% cannot regress from excluding more lines. Over-exclusion is the only real risk.

Known-unfixed artifact

bt_adapter.sh is capped at 98.55% (68/69) by line 97: the opening line of a backslash-continued assignment, which bash attributes to the LAST line of the run. A backslash-continued plain command is attributed to its first line and covers normally. Find them with:

grep -nE '^[[:space:]]*(local[[:space:]]+)?[A-Za-z_][A-Za-z0-9_]*=.*\\$' <lib>

Libs that do not measure at all

arch_sysctl.sh and nc_php.sh both report "kcov instrumented no lines" on every version of the instrument. Treat as unverified; they stay exempt.

---

State as of 2026-08-23

• Allowlist 84 (was 91 at the start of the previous session).
• fixes/lib cleared: arch_cpu 57/57, arch_hardware 94/94, thorium_repairs 75/75, hosts_guard_migrate 81/81, hosts_guard_rollback 65/65, ubuntu_perf_fixes 81/81, nvidia_config 76/76.
• The fixes/lib suite: 30 files, 645 assertions, 37s, zero suppressions.
• Three production bugs were found by writing tests, all dead branches: bt_audio.sh:21 (timeout cannot invoke a shell function), arch_perf_report.sh:36,65 and arch_hardware.sh (journal-size regex demanded a space before the unit, but journalctl prints 305.5M, so the vacuum was unreachable). Expect more. When a branch looks untestable, check whether it is reachable in production first — the user's standing answer is "fix it in this commit".

Still exempt and measured by an older session; all predate the instrument fixes and may now measure higher — re-measure before writing any test: dwm_config.sh 98.48%, pacman_hook_stall_setup.sh 96.30%, aw_autostart.sh 97.56%, rpi_nc_install.sh 96.59%, transcribe_pkgmgr.sh 98.00%, clean_audio_filters.sh 91.40%, transcribe_deps.sh 92.31%.

rpi_nc_ca.sh (73/73) and dot_resolver_install.sh (39/39) are already at 100% and can come off the allowlist whenever you want, by the three-step bar. That is two free entries.