ci(gate): the machine-health preflight and the truncation verdict guard
Some checks failed
CI / Node 22 (push) Successful in 12m14s
CI / Node 24 (push) Successful in 12m13s
CI / Integration + conformance (Node 22) (push) Failing after 13m28s
CI / Bun (latest) (push) Successful in 12m19s

Two guards for every gate lane, born from the 2026-08-13 lost-day ledger.
gate-preflight.sh refuses a lane on a machine that cannot be trusted to
produce honest numbers — co-tenant processes named by pid and command, load
average, CPU governor, disk floors — one FATAL line per violation so the
operator can act from the message alone. vitest-verdict-check.sh refuses a
suite log that cannot be trusted as a verdict — missing or mismatched
summary counts, files that never executed (a truncated run once read as
green from three files of ninety-nine), and worker-pool death signatures.

Both verified live: the preflight correctly refuses this workstation naming
its actual offenders; the verdict guard passes/fails five fixture shapes
(clean, wrong-count, truncated, worker-death, no-summary) and both CLI
modes. Wire-up into the CI lanes rides the runner program.
This commit is contained in:
David Snelling 2026-08-20 08:22:48 -07:00
parent 522b0cf827
commit 1e046aa115
3 changed files with 449 additions and 0 deletions

85
scripts/gate/README.md Normal file
View file

@ -0,0 +1,85 @@
# Gate Guards
Two standalone scripts that stand between a test/build gate and a false
verdict: one refuses to let the gate start on a noisy machine, the other
refuses to let a truncated or crashed vitest run be read as green.
## Why these exist
Both guards exist because of the 2026-08-13 lost-day ledger: a gate ran on
a machine under load, and separately a vitest worker pool died mid-suite
while still printing a plausible-looking summary line, and in both cases
the bad result was trusted and acted on for the better part of a day before
anyone noticed. Neither failure mode announces itself — a loaded machine
still finishes and reports numbers, and a truncated test run still prints a
`Test Files` / `Tests` line — so both guards check the evidence explicitly
rather than trusting that a gate finishing means the gate was valid.
## gate-preflight.sh
Run before any gate lane starts. Exits 1 the moment the machine isn't
gate-clean, with one `FATAL:` line per violation naming the exact offender
(the pid and command, the path, the measured value). Prints one `OK:` line
per check that passes. `WARNING:` lines mark checks that were skipped, not
failures.
Checks:
| # | Check | Default threshold | Override |
|---|-------|--------------------|----------|
| a | 1-minute load average | `nproc / 2` | `GATE_MAX_LOAD` |
| b | any non-allowlisted process over 50% of one core | 50% | `GATE_ALLOW_REGEX` (extra pattern matched against the process's args) |
| c | cpu0 scaling governor must be `performance` | — | none (warns and skips if the sysfs path is absent) |
| d | free space on `/` and `/tmp` | 10G each | `GATE_SKIP_DISK_CHECK=1` to skip entirely |
The allowlist for check (b) is always: this script's own process tree
(its ancestors and its direct child processes), `sshd`, `systemd`, and
kernel threads (recognizable by args wrapped in brackets, e.g.
`[kworker/0:1]`). `GATE_ALLOW_REGEX` extends it — it does not replace it.
## vitest-verdict-check.sh
Run after every vitest lane, against that lane's captured log. Fails
loudly, quoting the exact line or string that tripped it, when the log's
own summary can't be trusted:
- no `Test Files` (or, in `--count-tests` mode, `Tests`) summary line is
present at all
- the parenthesized total in that line doesn't match what was expected
- fewer files/tests are accounted for (passed + failed + skipped) than the
total claims — a truncated run
- the log contains `Unhandled Error` or `Timeout calling` anywhere — a dead
worker pool, regardless of what the summary line claims
```
vitest-verdict-check.sh <log-file> <expected-file-count>
vitest-verdict-check.sh --count-tests <log-file> <minimum-test-count>
```
The first form checks `Test Files` for an exact match. The second checks
`Tests` for a minimum (a floor, not an exact count, since the total number
of individual tests moves more often than the number of test files).
## Wiring into a CI lane
```sh
# Before any lane that will report a verdict:
scripts/gate/gate-preflight.sh || exit 1
# Run the suite, capturing its output:
npx vitest run tests/unit 2>&1 | tee /tmp/unit.log
# After every vitest lane, check the log against the actual file count:
EXPECTED_FILES=$(ls tests/unit/**/*.test.ts | wc -l)
scripts/gate/vitest-verdict-check.sh /tmp/unit.log "$EXPECTED_FILES" || exit 1
```
## Exit-code contract
| Script | Exit 0 | Exit 1 |
|--------|--------|--------|
| `gate-preflight.sh` | machine is gate-clean | one or more `FATAL:` violations printed |
| `vitest-verdict-check.sh` | log's summary is trustworthy and matches | usage error, missing/unreadable log, or one or more `FATAL:` violations printed |
Non-zero from either script means: do not trust the gate that was about to
run, or the result of the one that just ran.

206
scripts/gate/gate-preflight.sh Executable file
View file

@ -0,0 +1,206 @@
#!/bin/bash
set -euo pipefail
# Brainy Gate Preflight
# Refuses to let a test/build gate run on a machine that isn't clean enough
# to trust the numbers it produces. See scripts/gate/README.md for why (the
# 2026-08-13 lost-day ledger).
#
# Checks: 1-minute load average, any non-allowlisted process pinning a core,
# the cpu0 scaling governor, and free space on / and /tmp.
#
# Exit 0 and print one OK line per passing check when the machine is clean.
# Exit 1 and print one FATAL line per violation, naming the offender, when
# it is not.
#
# Known trap: a helper function whose last executed statement is a `while`
# (or any command whose own exit status happens to be nonzero) hands that
# status back as the function's return value. Called as a plain statement,
# that silently kills this script under `set -e`. Every helper below ends
# on an explicit `return 0` as its own statement, never on a loop or test.
#
# The same failure mode hides in plainer-looking lines too: `var=$(cmd)` is
# a bare assignment, so `set -e` DOES treat a nonzero `cmd` (or, under
# `pipefail`, a nonzero stage anywhere in `cmd`'s pipeline) as a failure of
# that statement and kills the script right there — even mid-loop, even
# when the "failure" is routine (a process that exited before a second
# lookup, a path that doesn't exist). Every such assignment below is paired
# with an explicit `|| var=""` fallback so a routine miss degrades to an
# empty value instead of an exit.
VIOLATIONS=0
ANCESTOR_PIDS=""
fatal() {
echo "FATAL: $1"
VIOLATIONS=$((VIOLATIONS + 1))
}
ok() {
echo "OK: $1"
}
# Walks this process's parent chain up to pid 1, then takes one snapshot of
# its direct children (the ps/read pipeline in check_processes), and
# records both in ANCESTOR_PIDS — so the process-scan below can recognize
# its own tree (the shell/terminal/session that launched it, plus its own
# helper commands) instead of flagging it. Children are captured once, up
# front, rather than re-queried per row later, so a helper command that has
# already exited by the time it's looked up can't be mistaken for a miss.
build_ancestor_pids() {
local pid="$$"
local ppid child
ANCESTOR_PIDS=" $pid "
while [ "$pid" != "1" ]; do
ppid=$(ps -o ppid= -p "$pid" 2>/dev/null | tr -d ' ') || ppid=""
if [ -z "$ppid" ]; then
break
fi
ANCESTOR_PIDS="${ANCESTOR_PIDS}${ppid} "
pid="$ppid"
done
while IFS= read -r child; do
[ -z "$child" ] && continue
ANCESTOR_PIDS="${ANCESTOR_PIDS}${child} "
done < <(ps --ppid "$$" -o pid= 2>/dev/null || true)
return 0
}
# (a) 1-minute load average vs. threshold (default: nproc / 2).
check_load() {
local max_load="${GATE_MAX_LOAD:-}"
if [ -z "$max_load" ]; then
max_load=$(( $(nproc) / 2 ))
if [ "$max_load" -lt 1 ]; then
max_load=1
fi
fi
local load_1m
load_1m=$(cut -d' ' -f1 /proc/loadavg)
if awk -v l="$load_1m" -v m="$max_load" 'BEGIN { exit !(l > m) }'; then
fatal "1-minute load average ${load_1m} exceeds threshold ${max_load} (GATE_MAX_LOAD=${max_load})"
else
ok "1-minute load average ${load_1m} is within threshold ${max_load}"
fi
return 0
}
# (b) any process outside the allowlist pinning more than half a core.
# Parsed with `read` into named fields, not an awk/cut chain — a fixed-column
# awk/cut split on `ps` output duplicated fields the first time this was
# tried, because process args vary in word count. `read` with a fixed list
# of variables dumps everything left over into the last one (args), which
# handles that correctly.
check_processes() {
local max_pcpu=50
local extra_regex="${GATE_ALLOW_REGEX:-}"
local violation_found=0
local line pcpu pid args pcpu_int
while IFS= read -r line; do
[ -z "$line" ] && continue
read -r pcpu pid args <<< "$line"
# Kernel threads report their comm in brackets, e.g. "[kworker/0:1]".
case "$args" in
\[*\]) continue ;;
esac
# This script's own tree: its ancestors (shell, terminal, session) and
# its direct children, both captured once by build_ancestor_pids.
case " $ANCESTOR_PIDS " in
*" $pid "*) continue ;;
esac
case "$args" in
*sshd*|*systemd*) continue ;;
esac
if [ -n "$extra_regex" ] && [[ "$args" =~ $extra_regex ]]; then
continue
fi
pcpu_int="${pcpu%.*}"
if [ -z "$pcpu_int" ]; then
pcpu_int=0
fi
if [ "$pcpu_int" -gt "$max_pcpu" ]; then
fatal "pid ${pid} ('${args}') is using ${pcpu}% of one core"
violation_found=1
fi
done < <(ps -eo pcpu,pid,args --sort=-pcpu | tail -n +2)
if [ "$violation_found" -eq 0 ]; then
ok "no process outside the allowlist exceeds ${max_pcpu}% of one core"
fi
return 0
}
# (c) cpu0 scaling governor must be "performance". Skipped with a warning
# (not a violation) when the sysfs path doesn't exist on this machine.
check_governor() {
local gov_path="/sys/devices/system/cpu/cpu0/cpufreq/scaling_governor"
if [ ! -r "$gov_path" ]; then
echo "WARNING: ${gov_path} not present; skipping governor check"
return 0
fi
local governor
governor=$(cat "$gov_path" 2>/dev/null) || governor=""
if [ "$governor" != "performance" ]; then
fatal "cpu0 governor is '${governor}', not 'performance'"
else
ok "cpu0 governor is 'performance'"
fi
return 0
}
# (d) free-space floors on / and /tmp (default 10G each). Skip entirely via
# GATE_SKIP_DISK_CHECK=1.
check_disk() {
if [ "${GATE_SKIP_DISK_CHECK:-0}" = "1" ]; then
echo "WARNING: disk free-space check skipped (GATE_SKIP_DISK_CHECK=1)"
return 0
fi
local floor_gb=10
local floor_bytes=$((floor_gb * 1024 * 1024 * 1024))
local path avail_bytes avail_gb
for path in / /tmp; do
avail_bytes=$(df --output=avail -B1 "$path" 2>/dev/null | tail -n 1 | tr -d ' ') || avail_bytes=""
if [ -z "$avail_bytes" ]; then
echo "WARNING: could not determine free space on ${path}; skipping"
continue
fi
if [ "$avail_bytes" -lt "$floor_bytes" ]; then
avail_gb=$((avail_bytes / 1024 / 1024 / 1024))
fatal "${path} has only ${avail_gb}G free, below the ${floor_gb}G floor"
else
ok "${path} has enough free space (floor ${floor_gb}G)"
fi
done
return 0
}
echo "Brainy gate preflight"
echo "----------------------"
build_ancestor_pids
check_load
check_processes
check_governor
check_disk
echo "----------------------"
if [ "$VIOLATIONS" -gt 0 ]; then
echo "FATAL: gate preflight failed with ${VIOLATIONS} violation(s) — machine is not gate-clean"
exit 1
fi
echo "gate preflight passed — machine is gate-clean"
exit 0

View file

@ -0,0 +1,158 @@
#!/bin/bash
set -euo pipefail
# Brainy Vitest Verdict Check
# Confirms a vitest run's own summary line is trustworthy before anything
# downstream treats a green run as green. See scripts/gate/README.md for why
# (the 2026-08-13 lost-day ledger).
#
# Usage:
# vitest-verdict-check.sh <log-file> <expected-file-count>
# vitest-verdict-check.sh --count-tests <log-file> <minimum-test-count>
#
# The first form checks the "Test Files" summary line's total against an
# exact expected count. The second checks the "Tests" summary line's total
# against a minimum. Both also fail on any sign the worker pool died
# mid-run, whether or not a summary line still made it into the log.
#
# Exit 0 and print one OK line per passing check when the log is clean.
# Exit 1 and print one FATAL line per violation, quoting the exact line or
# string that tripped it, when it is not.
#
# Known trap (shared with gate-preflight.sh): every helper below ends on an
# explicit `return 0` as its own statement, never on a loop or test, so a
# helper's last command can never hand its own exit status back as the
# function's under `set -e`. The same applies to `var=$(cmd)` assignments
# mid-helper: a bare assignment IS checked by `set -e`, so a `grep` that
# legitimately finds nothing (exit 1) would otherwise kill the script
# instead of just leaving the variable empty — every such assignment below
# is paired with an explicit `|| true` inside the substitution.
usage() {
echo "Usage: $0 <log-file> <expected-file-count>"
echo " $0 --count-tests <log-file> <minimum-test-count>"
exit 1
}
MODE="files"
if [ "${1:-}" = "--count-tests" ]; then
MODE="tests"
shift
fi
LOG_FILE="${1:-}"
THRESHOLD="${2:-}"
if [ -z "$LOG_FILE" ] || [ -z "$THRESHOLD" ]; then
usage
fi
if [ ! -f "$LOG_FILE" ]; then
echo "FATAL: log file '${LOG_FILE}' does not exist"
exit 1
fi
if ! [[ "$THRESHOLD" =~ ^[0-9]+$ ]]; then
echo "FATAL: threshold '${THRESHOLD}' is not a non-negative integer"
exit 1
fi
VIOLATIONS=0
fatal() {
echo "FATAL: $1"
VIOLATIONS=$((VIOLATIONS + 1))
}
ok() {
echo "OK: $1"
}
# Vitest colorizes its summary with ANSI escapes; strip them before parsing
# anything, or the color codes end up embedded in the fields we grep for.
CLEAN_LOG="$(sed 's/\x1b\[[0-9;]*m//g' "$LOG_FILE")"
# Worker-pool death: if either string appears, the run's own summary line —
# even if present and even if its numbers look fine — cannot be trusted,
# because the process died mid-suite and vitest's own accounting is what
# died with it.
check_worker_death() {
if echo "$CLEAN_LOG" | grep -q "Unhandled Error"; then
fatal "log contains 'Unhandled Error' — worker pool died mid-run"
fi
if echo "$CLEAN_LOG" | grep -q "Timeout calling"; then
fatal "log contains 'Timeout calling' — worker pool died mid-run"
fi
return 0
}
# Shared shape between the "Test Files" and "Tests" summary lines:
# <label> <n> passed | <n> failed | <n> skipped (<total>)
# `compare` is "eq" (total must equal threshold) or "min" (total must be at
# least threshold).
check_summary_line() {
local label="$1"
local threshold="$2"
local compare="$3"
local summary_line total accounted n
summary_line=$(echo "$CLEAN_LOG" | grep -E "^[[:space:]]*${label}[[:space:]]+" | tail -n 1 || true)
if [ -z "$summary_line" ]; then
fatal "no '${label}' summary line found in ${LOG_FILE}"
return 0
fi
total=$(echo "$summary_line" | grep -oE '\([0-9]+\)' | tr -d '()' | tail -n 1 || true)
if [ -z "$total" ]; then
fatal "'${label}' summary line has no parenthesized total: \"${summary_line}\""
return 0
fi
if [ "$compare" = "eq" ]; then
if [ "$total" -ne "$threshold" ]; then
fatal "'${label}' total is ${total}, expected ${threshold}: \"${summary_line}\""
else
ok "'${label}' total matches expected ${threshold}"
fi
else
if [ "$total" -lt "$threshold" ]; then
fatal "'${label}' total is ${total}, below minimum ${threshold}: \"${summary_line}\""
else
ok "'${label}' total ${total} meets minimum ${threshold}"
fi
fi
accounted=0
for n in $(echo "$summary_line" | grep -oE '[0-9]+ (passed|failed|skipped)' | grep -oE '^[0-9]+'); do
accounted=$((accounted + n))
done
if [ "$accounted" -lt "$total" ]; then
fatal "'${label}' line accounts for only ${accounted} of ${total} — truncated run: \"${summary_line}\""
else
ok "'${label}' line accounts for all ${total}"
fi
return 0
}
echo "Brainy vitest verdict check: ${LOG_FILE}"
echo "----------------------"
check_worker_death
if [ "$MODE" = "files" ]; then
check_summary_line "Test Files" "$THRESHOLD" "eq"
else
check_summary_line "Tests" "$THRESHOLD" "min"
fi
echo "----------------------"
if [ "$VIOLATIONS" -gt 0 ]; then
echo "FATAL: vitest verdict check failed with ${VIOLATIONS} violation(s) for ${LOG_FILE}"
exit 1
fi
echo "vitest verdict check passed for ${LOG_FILE}"
exit 0