diff --git a/tests/voice/100-cycle-soak.sh b/tests/voice/100-cycle-soak.sh
new file mode 100755
index 00000000..b0b82983
--- /dev/null
+++ b/tests/voice/100-cycle-soak.sh
@@ -0,0 +1,333 @@
+#!/usr/bin/env bash
+# M1 exit soak harness — 100 consecutive voice cycles (issue #109).
+#
+# Usage:
+#   ./tests/voice/100-cycle-soak.sh run [--cycles N] [--config PATH]
+#   ./tests/voice/100-cycle-soak.sh analyze LOG [--expect-cycles N] [--ledger PATH]
+#
+# `run` starts genie-core --voice and tees stderr to a timestamped log, then
+# analyzes it. Complete N wake→STT→LLM→TTS cycles manually (push-to-talk Enter)
+# or via wake-word before the analyzer runs.
+#
+# `analyze` parses a voice-loop log and emits pass/fail plus a JSON ledger.
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
+OUT_DIR="${SCRIPT_DIR}/out"
+
+DEFAULT_CYCLES=100
+# Upper bound per cycle: record_secs (3) + STT + LLM + TTS + post-TTS silence.
+MAX_CYCLE_MS=180000
+
+usage() {
+    cat <<'EOF'
+Usage:
+  100-cycle-soak.sh run   [--cycles N] [--config PATH] [--log PATH]
+  100-cycle-soak.sh analyze LOG [--expect-cycles N] [--ledger PATH] [--max-cycle-ms MS]
+
+Commands:
+  run      Start genie-core --voice; log stderr; analyze when the process exits.
+  analyze  Parse an existing voice-loop log and print pass/fail + JSON ledger.
+
+Environment:
+  GENIEPOD_CONFIG   Used when --config is omitted (run mode).
+EOF
+}
+
+require_cmd() {
+    if ! command -v "$1" >/dev/null 2>&1; then
+        echo "error: required command not found: $1" >&2
+        exit 1
+    fi
+}
+
+timestamp_utc() {
+    date -u +"%Y-%m-%dT%H:%M:%SZ"
+}
+
+analyze_log() {
+    local log_path=$1
+    local expect_cycles=$2
+    local ledger_path=$3
+    local max_cycle_ms=$4
+
+    if [[ ! -f "${log_path}" ]]; then
+        echo "error: log file not found: ${log_path}" >&2
+        exit 1
+    fi
+
+    mkdir -p "$(dirname "${ledger_path}")"
+
+    require_cmd python3
+
+    LOG_PATH="${log_path}" \
+        LEDGER_PATH="${ledger_path}" \
+        EXPECT_CYCLES="${expect_cycles}" \
+        MAX_CYCLE_MS="${max_cycle_ms}" \
+        python3 - <<'PY'
+import json
+import os
+import re
+import sys
+
+log_path = os.environ["LOG_PATH"]
+ledger_path = os.environ["LEDGER_PATH"]
+expect = int(os.environ["EXPECT_CYCLES"])
+max_ms = int(os.environ["MAX_CYCLE_MS"])
+
+recording_re = re.compile(r"\[voice\] Recording ")
+you_said_re = re.compile(r'\[voice\] You said: "([^"]*)"')
+geniepod_re = re.compile(r"\[voice\] GeniePod:")
+total_re = re.compile(r"\[voice\] Total cycle: (\d+) ms")
+no_speech_re = re.compile(r"\[voice\] No speech detected\.")
+fail_re = re.compile(r"\[voice\] (?:STT failed:|Recording failed:)")
+
+in_cycle = False
+saw_recording = False
+saw_you_said = False
+saw_geniepod = False
+transcript = ""
+cycles = []
+failures = 0
+silent_drops = 0
+stalls = 0
+
+with open(log_path, encoding="utf-8", errors="replace") as fh:
+    for line in fh:
+        if recording_re.search(line):
+            if not in_cycle:
+                in_cycle = True
+                saw_recording = True
+                saw_you_said = False
+                saw_geniepod = False
+                transcript = ""
+            continue
+
+        m = you_said_re.search(line)
+        if m:
+            saw_you_said = True
+            transcript = m.group(1)
+            continue
+
+        if geniepod_re.search(line):
+            saw_geniepod = True
+            continue
+
+        if in_cycle and no_speech_re.search(line) and saw_recording and not saw_you_said:
+            failures += 1
+            continue
+
+        if in_cycle and fail_re.search(line):
+            failures += 1
+            continue
+
+        m = total_re.search(line)
+        if not m:
+            continue
+
+        total_ms = int(m.group(1))
+        if in_cycle and saw_you_said and transcript and not saw_geniepod:
+            silent_drops += 1
+        if total_ms > max_ms:
+            stalls += 1
+
+        cycles.append(
+            {
+                "cycle": len(cycles) + 1,
+                "total_ms": total_ms,
+                "had_transcript": saw_you_said,
+                "spoke": saw_geniepod,
+                "transcript": transcript,
+            }
+        )
+        in_cycle = False
+        saw_recording = False
+        saw_you_said = False
+        saw_geniepod = False
+        transcript = ""
+
+missing = max(0, expect - len(cycles))
+stalls += missing
+passed = len(cycles) >= expect and stalls == 0 and silent_drops == 0
+
+ledger = {
+    "log": log_path,
+    "expect_cycles": expect,
+    "completed_cycles": len(cycles),
+    "stalls": stalls,
+    "silent_drops": silent_drops,
+    "partial_failures": failures,
+    "max_cycle_ms_budget": max_ms,
+    "pass": passed,
+    "cycles": cycles,
+}
+
+with open(ledger_path, "w", encoding="utf-8") as out:
+    json.dump(ledger, out, indent=2)
+    out.write("\n")
+
+sys.exit(0 if passed else 1)
+PY
+
+    local pass=0
+    if jq -e '.pass == true' "${ledger_path}" >/dev/null 2>&1; then
+        pass=1
+    fi
+
+    echo "=== 100-cycle voice soak (issue #109) ==="
+    jq -r '
+        "Log:               \(.log)",
+        "Expected cycles:   \(.expect_cycles)",
+        "Completed cycles:  \(.completed_cycles)",
+        "Stalls:            \(.stalls)",
+        "Silent drops:      \(.silent_drops)",
+        "Partial failures:  \(.partial_failures)",
+        "Result:            \(if .pass then "PASS" else "FAIL" end)"
+    ' "${ledger_path}"
+    echo "Ledger: ${ledger_path}"
+
+    if [[ "${pass}" -eq 1 ]]; then
+        return 0
+    fi
+    return 1
+}
+
+cmd_run() {
+    local cycles="${DEFAULT_CYCLES}"
+    local config="${GENIEPOD_CONFIG:-}"
+    local log_path=""
+
+    while [[ $# -gt 0 ]]; do
+        case "$1" in
+            --cycles)
+                cycles="$2"
+                shift 2
+                ;;
+            --config)
+                config="$2"
+                shift 2
+                ;;
+            --log)
+                log_path="$2"
+                shift 2
+                ;;
+            *)
+                echo "error: unknown run option: $1" >&2
+                usage
+                exit 1
+                ;;
+        esac
+    done
+
+    require_cmd jq
+    local binary="${REPO_ROOT}/target/release/genie-core"
+    if [[ ! -x "${binary}" ]]; then
+        binary="${REPO_ROOT}/target/debug/genie-core"
+    fi
+    if [[ ! -x "${binary}" ]]; then
+        echo "error: build genie-core first (cargo build -p genie-core --release)" >&2
+        exit 1
+    fi
+
+    mkdir -p "${OUT_DIR}"
+    if [[ -z "${log_path}" ]]; then
+        log_path="${OUT_DIR}/soak-$(date -u +%Y%m%dT%H%M%SZ).log"
+    fi
+    local ledger_path="${log_path%.log}.json"
+
+    echo "=== M1 voice soak — run mode ==="
+    echo "Target cycles:  ${cycles}"
+    echo "Log file:       ${log_path}"
+    echo "Binary:         ${binary}"
+    if [[ -n "${config}" ]]; then
+        echo "Config:         ${config}"
+        export GENIEPOD_CONFIG="${config}"
+    else
+        echo "Config:         (GENIEPOD_CONFIG unset — genie-core default)"
+    fi
+    echo
+    echo "Complete ${cycles} voice cycles (wake word or push-to-talk), then exit genie-core (Ctrl+C)."
+    echo "stderr is tee'd to the log; analysis runs on exit."
+    echo
+
+  set +e
+  "${binary}" --voice 2>&1 | tee "${log_path}"
+  local core_status=${PIPESTATUS[0]}
+  set -e
+
+    echo
+    echo "genie-core exited with status ${core_status}"
+    analyze_log "${log_path}" "${cycles}" "${ledger_path}" "${MAX_CYCLE_MS}"
+}
+
+cmd_analyze() {
+    local log_path=""
+    local expect_cycles="${DEFAULT_CYCLES}"
+    local ledger_path=""
+    local max_cycle_ms="${MAX_CYCLE_MS}"
+
+    if [[ $# -lt 1 ]]; then
+        echo "error: analyze requires a LOG path" >&2
+        usage
+        exit 1
+    fi
+    log_path=$1
+    shift
+
+    while [[ $# -gt 0 ]]; do
+        case "$1" in
+            --expect-cycles)
+                expect_cycles="$2"
+                shift 2
+                ;;
+            --ledger)
+                ledger_path="$2"
+                shift 2
+                ;;
+            --max-cycle-ms)
+                max_cycle_ms="$2"
+                shift 2
+                ;;
+            *)
+                echo "error: unknown analyze option: $1" >&2
+                usage
+                exit 1
+                ;;
+        esac
+    done
+
+    require_cmd jq
+    if [[ -z "${ledger_path}" ]]; then
+        ledger_path="${log_path%.log}.json"
+        if [[ "${ledger_path}" == "${log_path}" ]]; then
+            ledger_path="${log_path}.ledger.json"
+        fi
+    fi
+
+    analyze_log "${log_path}" "${expect_cycles}" "${ledger_path}" "${max_cycle_ms}"
+}
+
+main() {
+    if [[ $# -lt 1 ]]; then
+        usage
+        exit 1
+    fi
+
+    local cmd=$1
+    shift
+
+    case "${cmd}" in
+        run) cmd_run "$@" ;;
+        analyze) cmd_analyze "$@" ;;
+        -h | --help | help) usage ;;
+        *)
+            echo "error: unknown command: ${cmd}" >&2
+            usage
+            exit 1
+            ;;
+    esac
+}
+
+main "$@"