Files
Linux_post_install/tests/t-ai-server-bus.sh
T
Your Name 0b5043a9f3
gates / consistency-and-conventions (push) Successful in 26s
fix: llama-server start breakage — version detection, flag-validation race, model dir resolution, user-bus pre-flight
User report after the llamacpp app install: 'installed llama.cpp unknown',
valid flags rejected (randomly per run), 'Model not found' for the HF
downloader's own layout, and a systemd user-bus failure over SSH. Detective
(real b10822 binary, FACT) found four independent causes:

- version: llama-server --version prints to STDERR; detect_llama_version's
  2>/dev/null swallowed it -> always 'unknown'. Now captures 2>&1 + accepts
  semver/build tokens (incl. build 1.2.3 edge)
- validation: printf|grep -q under pipefail -> SIGPIPE rc=141 race randomly
  rejected flags present in the 59 KB --help. Now pipe-less grep (no race);
  20x determinism regression test
- model resolution: resolve_model accepted files only, but the HF downloader
  creates <models>/<repo>/file.gguf dirs. Now expands a dir with exactly one
  *.gguf (never silently picks; multi-gguf lists + errs)
- port: llama.cpp default 8080 vs tool/adapter 8088; validation reliability
  means --port is now always pinned in the unit
- user bus: headless/SSH sessions lack XDG_RUNTIME_DIR -> ensure_user_bus in
  lib/common.sh pre-flights all three systemctl --user tools with remediation
  text; pos ai server --no-unit direct-run escape hatch (pidfile) for boxes
  with no bus
- find_llamacpp narrowed to llama-server/llama-server-cuda (bare 'server'
  fallback hazard); installer post-install sanity (version+help execute,
  symlink targets resolve)

Architect decisions DQ1-DQ6 recorded. Tester: 4 new regression files
(version-from-stderr, 25x flag-validation determinism, model dir expansion,
bus pre-flight + E2E) + 3 fixture updates; suite 16 files / 269 checks.

Verified: make gen idempotent; make check OK; make lint 0 FAIL, 0 WARN;
make test 269/269 (~49s); bash -n clean; git diff --check clean.
2026-09-06 09:25:52 -04:00

225 lines
10 KiB
Bash

#!/usr/bin/env bash
set -euo pipefail
# t-ai-server-bus.sh — F4 + E2E regression for the user systemd-bus pre-flight.
#
# F4 contract:
# - `pos ai server start` (unit path) calls ensure_user_bus BEFORE the unit
# is written; with no user bus (SSH/headless: XDG_RUNTIME_DIR and
# DBUS_SESSION_BUS_ADDRESS unset, systemctl --user fails) it errs with BOTH
# remediation lines and leaves NO orphaned unit file.
# - with a reachable bus the command proceeds past the pre-flight.
# - `--no-unit` is the bus-less escape hatch: it skips the pre-flight, runs
# llama-server directly under nohup with a pidfile, and stop/status can
# find it via the pidfile.
#
# E2E regression (the reported user scenario):
# - install-shaped stubs + working bus: `start Qwen-Qwen3-1.7B-GGUF` resolves
# the model dir (F3), writes a correct unit (F6: all 5 flags + --port
# 8088), reports a resolvable version (F1: 0.4.0, not "unknown"), and the
# unit passes `systemd-analyze verify`.
# - SSH-shaped no-bus env: clean pre-flight err, remediation lines, NO unit.
run_test() {
local sandbox stubs models unitdir runtime
sandbox="$(mksandbox ai-server-bus)"
stubs="$sandbox/stubs"
models="$sandbox/models"
unitdir="$sandbox/userunits"
runtime="$sandbox/runtime"
mkdir -p "$stubs" "$models" "$unitdir" "$runtime"
# ── User's real HF layout (F3 fixture) ──
mkdir -p "$models/Qwen-Qwen3-1.7B-GGUF"
touch "$models/Qwen-Qwen3-1.7B-GGUF/Qwen3-1.7B-Q8_0.gguf"
# ── Stubs ──
# llama-server: b10822-shaped (version → stderr, full-support help).
cat > "$stubs/llama-server" <<'STUB'
#!/usr/bin/env bash
case "$1" in
--version) printf 'version: 0.4.0-dev (build 10822, commit c457e3bf7)\n' >&2 ;;
--help)
cat <<'HELP'
usage: llama-server [options]
options:
--host <addr> bind address
--port <port> server port
--n-gpu-layers <n> layers to offload
--ctx-size <n> context size
--threads <n> cpu threads
HELP
;;
*) exit 0 ;;
esac
STUB
# systemctl: FAILS exactly like the real bus-missing case (SSH).
cat > "$stubs/systemctl" <<'STUB'
#!/usr/bin/env bash
echo "Failed to connect to user scope bus via local transport: \$DBUS_SESSION_BUS_ADDRESS and \$XDG_RUNTIME_DIR not defined" >&2
exit 1
STUB
# nvidia-smi: no GPU → deterministic CPU path.
printf '#!/usr/bin/env bash\nexit 1\n' > "$stubs/nvidia-smi"
# curl: healthy JSON (health + models probes).
printf '#!/usr/bin/env bash\nprintf "%%s" '"'"'{"status":"ok"}'"'"'\n' > "$stubs/curl"
# loginctl: Linger=no (so the linger warn path is deterministic).
printf '#!/usr/bin/env bash\nprintf "Linger=no\\n"\n' > "$stubs/loginctl"
chmod +x "$stubs/llama-server" "$stubs/systemctl" "$stubs/nvidia-smi" "$stubs/curl" "$stubs/loginctl"
local server="$ROOT/bin/pos-ai-server"
local nobus_env=(-u XDG_RUNTIME_DIR -u DBUS_SESSION_BUS_ADDRESS
PATH="$stubs:/usr/bin:/bin" HF_DOWNLOAD_DIR="$models"
USER_SYSTEMD_DIR="$unitdir" NO_UNIT_PIDFILE="$sandbox/pos-ai-server.pid"
NO_UNIT_LOG="$sandbox/pos-ai-server.log")
local bus_env=(-u XDG_RUNTIME_DIR -u DBUS_SESSION_BUS_ADDRESS
PATH="$stubs:/usr/bin:/bin" HF_DOWNLOAD_DIR="$models"
USER_SYSTEMD_DIR="$unitdir" XDG_RUNTIME_DIR="$runtime"
DBUS_SESSION_BUS_ADDRESS="unix:path=$runtime/bus"
NO_UNIT_PIDFILE="$sandbox/pos-ai-server.pid" NO_UNIT_LOG="$sandbox/pos-ai-server.log")
# ═══ F4 pre-flight: bus missing → err + remediation + NO unit ═══
test_run_env "${nobus_env[@]}" -- timeout 30 "$server" start Qwen-Qwen3-1.7B-GGUF
check_contains "F4 bus-missing start names the bus problem" "user systemd bus" "$TR_OUT"
check_contains "F4 remediation: export XDG_RUNTIME_DIR" "export XDG_RUNTIME_DIR=/run/user/" "$TR_OUT"
check_contains "F4 remediation: sudo loginctl enable-linger" "sudo loginctl enable-linger" "$TR_OUT"
check_contains "F4 states no unit was written" "no unit was written" "$TR_OUT"
[ "$TR_RC" -ne 0 ] && printf ' PASS F4 bus-missing start exits nonzero\n' || printf ' FAIL F4 bus-missing start exited 0\n'
check_file_absent "F4 no orphaned unit written" "$unitdir/pos-ai-server.service"
# ═══ F4 pre-flight: bus present → proceeds past pre-flight (DRY_RUN) ═══
# Replace the failing systemctl with a succeeding one.
printf '#!/usr/bin/env bash\nexit 0\n' > "$stubs/systemctl"
chmod +x "$stubs/systemctl"
test_run_env "${bus_env[@]}" DRY_RUN=1 -- timeout 30 "$server" start Qwen-Qwen3-1.7B-GGUF
check_rc "F4 bus-present DRY_RUN start exits 0" 0 "$TR_RC"
check_contains "F4 bus-present proceeds past pre-flight (ExecStart)" "ExecStart:" "$TR_OUT"
# ═══ F4 --no-unit escape hatch (bus-less direct run) ═══
# A real --no-unit start must skip the bus check and launch the stub under
# nohup with a pidfile; status finds it; stop kills it. The stub sleeps via
# `exec sleep` so the recorded pid IS the sleeper (no orphan children).
cat > "$stubs/llama-server" <<'STUB'
#!/usr/bin/env bash
case "$1" in
--version) printf 'version: 0.4.0-dev (build 10822, commit c457e3bf7)\n' >&2 ;;
--help)
cat <<'HELP'
usage: llama-server [options]
options:
--host <addr> bind address
--port <port> server port
--n-gpu-layers <n> layers to offload
--ctx-size <n> context size
--threads <n> cpu threads
HELP
;;
*) exec sleep 300 ;;
esac
STUB
chmod +x "$stubs/llama-server"
# failing systemctl stub again: --no-unit must NOT care.
cat > "$stubs/systemctl" <<'STUB'
#!/usr/bin/env bash
echo "no bus" >&2
exit 1
STUB
chmod +x "$stubs/systemctl"
test_run_env "${nobus_env[@]}" -- timeout 30 "$server" start Qwen-Qwen3-1.7B-GGUF --no-unit
check_rc "F4 --no-unit bus-less start exits 0" 0 "$TR_RC"
check_file_exists "F4 --no-unit writes pidfile" "$sandbox/pos-ai-server.pid"
check_contains "F4 --no-unit prints log path hint" "log:" "$TR_OUT"
local direct_pid
direct_pid="$(cat "$sandbox/pos-ai-server.pid" 2>/dev/null || true)"
if [[ "$direct_pid" =~ ^[0-9]+$ ]] && kill -0 "$direct_pid" 2>/dev/null; then
printf ' PASS F4 --no-unit launched a live server process (pid %s)\n' "$direct_pid"
else
printf ' FAIL F4 --no-unit pid %s not alive\n' "${direct_pid:-EMPTY}"
fi
# status via the pidfile reports running (deps + curl stubs present).
test_run_env "${nobus_env[@]}" -- timeout 30 "$server" status
check_rc "F4 status exits 0 after --no-unit start" 0 "$TR_RC"
check_contains "F4 status finds direct-run server running" "service: running" "$TR_OUT"
# stop kills the pid and removes the pidfile.
test_run_env "${nobus_env[@]}" -- timeout 30 "$server" stop
check_rc "F4 --no-unit stop exits 0" 0 "$TR_RC"
check_file_absent "F4 stop removes pidfile" "$sandbox/pos-ai-server.pid"
if [[ "$direct_pid" =~ ^[0-9]+$ ]] && ! kill -0 "$direct_pid" 2>/dev/null; then
printf ' PASS F4 stop killed the direct-run process\n'
else
printf ' FAIL F4 direct-run process still alive after stop\n'
fi
# ═══ E2E positive: user scenario with a working bus ═══
# Restore full-support stub + succeeding systemctl.
cat > "$stubs/llama-server" <<'STUB'
#!/usr/bin/env bash
case "$1" in
--version) printf 'version: 0.4.0-dev (build 10822, commit c457e3bf7)\n' >&2 ;;
--help)
cat <<'HELP'
usage: llama-server [options]
options:
--host <addr> bind address
--port <port> server port
--n-gpu-layers <n> layers to offload
--ctx-size <n> context size
--threads <n> cpu threads
HELP
;;
*) exit 0 ;;
esac
STUB
printf '#!/usr/bin/env bash\nexit 0\n' > "$stubs/systemctl"
chmod +x "$stubs/llama-server" "$stubs/systemctl"
test_run_env "${bus_env[@]}" -- timeout 60 "$server" start Qwen-Qwen3-1.7B-GGUF
check_rc "E2E user scenario start exits 0" 0 "$TR_RC"
check_file_exists "E2E unit written" "$unitdir/pos-ai-server.service"
local unit="$unitdir/pos-ai-server.service"
local unit_exec
unit_exec="$(grep '^ExecStart=' "$unit")"
check_contains "E2E unit: model expanded from dir" "Qwen3-1.7B-Q8_0.gguf" "$unit_exec"
check_contains "E2E unit: --port 8088 pinned" "--port 8088" "$unit_exec"
check_contains "E2E unit: --host present" "--host 127.0.0.1" "$unit_exec"
check_contains "E2E unit: --n-gpu-layers present" "--n-gpu-layers 0" "$unit_exec"
check_contains "E2E unit: --ctx-size present" "--ctx-size 4096" "$unit_exec"
check_contains "E2E unit: --threads present" "--threads" "$unit_exec"
# Version resolvable through the real CLI on the same run's stub.
test_run_env "${bus_env[@]}" -- timeout 30 "$server" status
check_contains "E2E version is resolvable (0.4.0)" "version: 0.4.0" "$TR_OUT"
check_not_contains "E2E version is NOT unknown" "version: unknown" "$TR_OUT"
# Unit must pass systemd-analyze verify.
if command -v systemd-analyze >/dev/null 2>&1; then
test_run systemd-analyze verify "$unit"
check_rc "E2E systemd-analyze verify accepts unit" 0 "$TR_RC"
else
skip_case "E2E systemd-analyze verify" "systemd-analyze not available"
fi
# stop cleans up (unit removed).
test_run_env "${bus_env[@]}" -- timeout 30 "$server" stop
check_rc "E2E stop exits 0" 0 "$TR_RC"
check_file_absent "E2E stop removes unit" "$unitdir/pos-ai-server.service"
# ═══ E2E SSH-shaped: no bus → clean pre-flight err, remediation, NO unit ═══
cat > "$stubs/systemctl" <<'STUB'
#!/usr/bin/env bash
echo "Failed to connect to user scope bus via local transport: \$DBUS_SESSION_BUS_ADDRESS and \$XDG_RUNTIME_DIR not defined" >&2
exit 1
STUB
chmod +x "$stubs/systemctl"
test_run_env "${nobus_env[@]}" -- timeout 30 "$server" start Qwen-Qwen3-1.7B-GGUF
check_rc "E2E SSH-shaped start exits nonzero" 1 "$TR_RC"
check_contains "E2E SSH-shaped: bus problem named" "user systemd bus" "$TR_OUT"
check_contains "E2E SSH-shaped: export XDG_RUNTIME_DIR remediation" "export XDG_RUNTIME_DIR=/run/user/" "$TR_OUT"
check_contains "E2E SSH-shaped: loginctl enable-linger remediation" "sudo loginctl enable-linger" "$TR_OUT"
check_contains "E2E SSH-shaped: no unit was written" "no unit was written" "$TR_OUT"
check_file_absent "E2E SSH-shaped: no orphaned unit" "$unitdir/pos-ai-server.service"
}