bash-watchdog-presence-gate-false-alarm-elf-pi-layout

Answer

Diagnosis complete and fix verified. Full solution written to /workspace/solution.md; a runnable harness is at /workspace/verify_presence_gate.sh (4/4 scenarios pass).

Root cause

The wrapper ~/.local/bin/pi-agent resolves pi ELF-first ($PI_BIN → <wrapper_dir>/pi → <wrapper_dir>/../lib/pi/pi → /tmp/pi/pi), then the npm JS layout. scripts/pi-agent-watchdog.sh stage-1 instead asserted the npm-only tree (packages/coding-agent/dist/cli.js + node_modules/.bin + package.json). On the 2026-09-05 ELF release, cli=0 bin=0 but leftover package.json gives pkg=1, and the probe is aimed at the missing npm entrypoint → resolve=probe-error, so a healthy host is flagged hollow-wipe. The same assumption also ran stage-2 npm workspace resolution unconditionally.

The fix (scripts/pi-agent-watchdog.sh)

1. Add resolve_pi_bin() mirroring findPiBin's ELF-first order (regular-file check so a directory named pi can't match).
2. Gate on the resolved layout, keeping cli/pkg/bin/wrapper as informational-only log fields (old alert string preserved). Hollow alert only when neither layout resolves.
3. Skip stage-2 workspace resolution when PI_LAYOUT == "elf".

WATCHDOG_WRAPPER_DIR="${PI_AGENT_DIR:-$HOME/.local/bin}"

resolve_pi_bin() {
  local c
  local elf_candidates=("${PI_BIN:-}" "$WATCHDOG_WRAPPER_DIR/pi" \
                        "$WATCHDOG_WRAPPER_DIR/../lib/pi/pi" "/tmp/pi/pi")
  for c in "${elf_candidates[@]}"; do
    [[ -n "$c" ]] || continue
    if [[ -f "$c" && ! -d "$c" ]]; then PI_RESOLVED_BIN="$c"; PI_LAYOUT="elf"; return 0; fi
  done
  local js_candidates=()
  [[ -n "${PI_HOME:-}" ]] && js_candidates+=("$PI_HOME/packages/coding-agent/dist/cli.js")
  js_candidates+=("/tmp/pi/packages/coding-agent/dist/cli.js")
  js_candidates+=("$HOME/.local/share/pi/packages/coding-agent/dist/cli.js")
  js_candidates+=("$HOME/.pi/packages/coding-agent/dist/cli.js")
  for c in "${js_candidates[@]}"; do
    [[ -f "$c" ]] && { PI_RESOLVED_BIN="$c"; PI_LAYOUT="npm"; return 0; }
  done
  PI_RESOLVED_BIN=""; PI_LAYOUT="hollow"; return 1
}

# stage-1: npm legs informational, resolve ELF-first
PI_RESOLVED_BIN=""; PI_LAYOUT="hollow"
resolve_pi_bin || true
resolve="probe-error"
if [[ -n "$PI_RESOLVED_BIN" ]]; then
  if [[ "$PI_LAYOUT" == "elf" ]]; then
    "$PI_RESOLVED_BIN" --version >/dev/null 2>&1 && resolve="ok"
  else
    resolve="ok"
  fi
fi
if [[ "$PI_LAYOUT" == "hollow" ]]; then
  echo "ALERT: pi-agent binary UNHEALTHY (hollow-wipe class) — cli=$cli pkg=$pkg bin=$bin wrapper=$wrapper resolve=$resolve" >&2
  exit 1
fi

# stage-2: only for npm layout; ELF is self-contained
if [[ "$PI_LAYOUT" == "npm" ]]; then
  : # existing workspace/node_modules resolution
fi
exit 0

Verification

bash /workspace/verify_presence_gate.sh:

PASS real-elf-layout        rc=0   cli=0 pkg=1 bin=0 wrapper=1 layout=elf resolve=ok
PASS synthetic-npm-layout   rc=0   cli=1 ... layout=npm resolve=ok
PASS hollow-layout-alert    rc=1   layout=hollow resolve=probe-error  (original ALERT preserved)
PASS elf-npm-legs-absent    rc=0   (regression case)
TOTAL: 4 passed, 0 failed

Acceptance: healthy ELF host (cli=0 pkg=1 bin=0 wrapper=1 + /tmp/pi/pi) exits 0; npm layout still runs stage-2; neither layout yields the unchanged hollow-wipe class alert rc=1; a directory named pi never resolves. A lockstep grep "tmp/pi/pi" across the watchdog and wrapper is recommended as a CI tripwire, since layout detection is duplicated.