diff --git a/skills/continuous-learning-v2/hooks/observe.sh b/skills/continuous-learning-v2/hooks/observe.sh
index 4fc51458..65d2f1d2 100755
--- a/skills/continuous-learning-v2/hooks/observe.sh
+++ b/skills/continuous-learning-v2/hooks/observe.sh
@@ -477,21 +477,82 @@ fi
 # which caused runaway parallel Claude analysis processes.
 SIGNAL_EVERY_N="${ECC_OBSERVER_SIGNAL_EVERY_N:-20}"
 SIGNAL_COUNTER_FILE="${PROJECT_DIR}/.observer-signal-counter"
+SIGNAL_COUNTER_LOCK="${SIGNAL_COUNTER_FILE}.lock"
 ACTIVITY_FILE="${PROJECT_DIR}/.observer-last-activity"
 
 touch "$ACTIVITY_FILE" 2>/dev/null || true
 
+# Serialize the throttle-counter read-modify-write. observe.sh runs on every
+# tool call (which can fire every second), so concurrent invocations previously
+# raced on this counter: both read the same value, both incremented, and one
+# write was lost, signaling the observer at unpredictable intervals (#2296).
+# Prefer flock (a kernel advisory lock the OS releases automatically if the hook
+# is killed); fall back to the atomic mkdir lock this script already uses for
+# the lazy-start path above. Both wrap the same read-modify-write below.
 should_signal=0
-if [ -f "$SIGNAL_COUNTER_FILE" ]; then
-  counter=$(cat "$SIGNAL_COUNTER_FILE" 2>/dev/null || echo 0)
-  counter=$((counter + 1))
-  if [ "$counter" -ge "$SIGNAL_EVERY_N" ]; then
-    should_signal=1
-    counter=0
+
+_ecc_bump_signal_counter() {
+  if [ -f "$SIGNAL_COUNTER_FILE" ]; then
+    counter=$(cat "$SIGNAL_COUNTER_FILE" 2>/dev/null || echo 0)
+    # Guard against a corrupt counter file: a non-integer value would abort the
+    # hook under `set -e` at the arithmetic below.
+    case "$counter" in
+      ''|*[!0-9]*) counter=0 ;;
+    esac
+    counter=$((counter + 1))
+    if [ "$counter" -ge "$SIGNAL_EVERY_N" ]; then
+      should_signal=1
+      counter=0
+    fi
+    echo "$counter" > "$SIGNAL_COUNTER_FILE"
+  else
+    echo "1" > "$SIGNAL_COUNTER_FILE"
   fi
-  echo "$counter" > "$SIGNAL_COUNTER_FILE"
+}
+
+if command -v flock >/dev/null 2>&1 && exec 8>"$SIGNAL_COUNTER_LOCK" 2>/dev/null; then
+  # flock is auto-released when fd 8 closes or the process dies, so there is no
+  # stale lock and no lost increment. Use a bounded -w wait so the hook never
+  # blocks indefinitely, and only bump the counter while the lock is held -- on
+  # a timeout we skip the tick rather than doing an unlocked read-modify-write.
+  if flock -w 2 8 2>/dev/null; then
+    _ecc_bump_signal_counter
+    flock -u 8 2>/dev/null || true
+  fi
+  exec 8>&- 2>/dev/null || true
 else
-  echo "1" > "$SIGNAL_COUNTER_FILE"
+  # No flock (e.g. macOS): atomic mkdir lock with a bounded spin so the hook
+  # never blocks indefinitely. A trap releases the lock on every exit path --
+  # including the async-timeout SIGTERM -- so a killed hook does not strand the
+  # directory. We deliberately do NOT hand-roll PID-based stale reclaim:
+  # re-verifying then removing another process's lock is racy and can delete a
+  # live re-acquirer's directory, reintroducing the very race this fixes.
+  _signal_lock_held=0
+  _signal_lock_spins=0
+  while [ "$_signal_lock_spins" -lt 100 ]; do
+    if mkdir "$SIGNAL_COUNTER_LOCK" 2>/dev/null; then
+      # EXIT cleans up on normal completion. INT/TERM must release AND exit:
+      # a signal trap that only released the lock would otherwise fall through
+      # and continue the read-modify-write without ownership.
+      trap 'rmdir "$SIGNAL_COUNTER_LOCK" 2>/dev/null || true' EXIT
+      trap 'rmdir "$SIGNAL_COUNTER_LOCK" 2>/dev/null || true; exit 130' INT
+      trap 'rmdir "$SIGNAL_COUNTER_LOCK" 2>/dev/null || true; exit 143' TERM
+      _signal_lock_held=1
+      break
+    fi
+    _signal_lock_spins=$((_signal_lock_spins + 1))
+    sleep 0.02
+  done
+  if [ "$_signal_lock_held" -eq 1 ]; then
+    # Bump only under the held lock -- never an unlocked read-modify-write.
+    _ecc_bump_signal_counter
+    rmdir "$SIGNAL_COUNTER_LOCK" 2>/dev/null || true
+    trap - EXIT INT TERM
+  fi
+  # If the lock could not be acquired within the spin budget we skip this tick
+  # rather than racing on an unlocked counter. Dropping one throttle tick under
+  # extreme contention only delays the next observer signal slightly; it never
+  # corrupts the counter or signals spuriously.
 fi
 
 # Signal observer if running and throttle allows (check both project-scoped and global observer, deduplicate)
diff --git a/tests/hooks/observe-signal-counter-race.test.js b/tests/hooks/observe-signal-counter-race.test.js
new file mode 100644
index 00000000..2fb978aa
--- /dev/null
+++ b/tests/hooks/observe-signal-counter-race.test.js
@@ -0,0 +1,271 @@
+/**
+ * Regression tests for the SIGUSR1 throttle-counter race in observe.sh (#2296)
+ *
+ * observe.sh runs on every tool call and bumps a throttle counter in
+ * ${PROJECT_DIR}/.observer-signal-counter so the observer is signaled only
+ * every N observations (#521). The bump used a plain read-modify-write with no
+ * locking, so concurrent hook invocations could read the same value, both
+ * increment, and lose a write — the observer then fired at unpredictable
+ * intervals. The fix serializes the read-modify-write with an atomic mkdir
+ * lock.
+ *
+ * These tests drive the real observe.sh (reusing the stub harness from
+ * observer-memory.test.js) and assert the lock's invariant: with the reset
+ * threshold set high enough that no reset fires, the final counter must equal
+ * the number of invocations — i.e. no increment is ever lost, even under heavy
+ * concurrency.
+ *
+ * Run with: node tests/hooks/observe-signal-counter-race.test.js
+ */
+
+const assert = require('assert');
+const path = require('path');
+const fs = require('fs');
+const os = require('os');
+const { spawn, spawnSync } = require('child_process');
+
+let passed = 0;
+let failed = 0;
+
+function test(name, fn) {
+  try {
+    fn();
+    console.log(`  ✓ ${name}`);
+    passed++;
+  } catch (err) {
+    console.log(`  ✗ ${name}`);
+    console.log(`    Error: ${err.message}`);
+    failed++;
+  }
+}
+
+async function asyncTest(name, fn) {
+  try {
+    await fn();
+    console.log(`  ✓ ${name}`);
+    passed++;
+  } catch (err) {
+    console.log(`  ✗ ${name}`);
+    console.log(`    Error: ${err.message}`);
+    failed++;
+  }
+}
+
+function createTempDir() {
+  return fs.mkdtempSync(path.join(os.tmpdir(), 'ecc-signal-race-'));
+}
+
+function cleanupDir(dir) {
+  try {
+    fs.rmSync(dir, { recursive: true, force: true });
+  } catch {
+    // ignore cleanup errors
+  }
+}
+
+const repoRoot = path.resolve(__dirname, '..', '..');
+const observeShPath = path.join(repoRoot, 'skills', 'continuous-learning-v2', 'hooks', 'observe.sh');
+
+const isWindows = process.platform === 'win32';
+const hasPython = !isWindows && spawnSync('python3', ['--version']).status === 0;
+// When the runner has flock the lock is exact (blocking, kernel auto-release);
+// without it observe.sh uses a best-effort mkdir spin that may drop at most one
+// increment under pathological contention.
+const hasFlock = !isWindows && spawnSync('bash', ['-c', 'command -v flock']).status === 0;
+
+// Build a self-contained observe.sh sandbox (stub detect-project.sh +
+// homunculus-dir.sh, SKILL_ROOT patched to the sandbox) and return its paths.
+function buildSandbox() {
+  const testDir = createTempDir();
+  const projectDir = path.join(testDir, 'project');
+  fs.mkdirSync(projectDir, { recursive: true });
+
+  const skillRoot = path.join(testDir, 'skill');
+  const scriptsDir = path.join(skillRoot, 'scripts');
+  const scriptsLibDir = path.join(scriptsDir, 'lib');
+  const hooksDir = path.join(skillRoot, 'hooks');
+  fs.mkdirSync(scriptsDir, { recursive: true });
+  fs.mkdirSync(scriptsLibDir, { recursive: true });
+  fs.mkdirSync(hooksDir, { recursive: true });
+
+  fs.writeFileSync(
+    path.join(scriptsDir, 'detect-project.sh'),
+    [
+      '#!/bin/bash',
+      'PROJECT_ID="test-project"',
+      'PROJECT_NAME="test-project"',
+      `PROJECT_ROOT="${projectDir}"`,
+      `PROJECT_DIR="${projectDir}"`,
+      'CLV2_PYTHON_CMD="python3"',
+      ''
+    ].join('\n')
+  );
+  fs.writeFileSync(
+    path.join(scriptsLibDir, 'homunculus-dir.sh'),
+    [
+      '#!/bin/bash',
+      '_ecc_resolve_homunculus_dir() { printf "%s\\n" "$HOME/.local/share/ecc-homunculus"; }',
+      ''
+    ].join('\n')
+  );
+
+  let observeContent = fs.readFileSync(observeShPath, 'utf8');
+  const skillRootMarker = 'SKILL_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"';
+  // Fail fast if observe.sh's SKILL_ROOT definition drifts; otherwise the
+  // no-op replace would leave the sandbox pointing at the real skill tree and
+  // the test could pass spuriously.
+  assert.ok(
+    observeContent.includes(skillRootMarker),
+    'observe.sh SKILL_ROOT definition changed; update the sandbox rewrite'
+  );
+  observeContent = observeContent.replace(
+    skillRootMarker,
+    `SKILL_ROOT="${skillRoot}"`
+  );
+  const testObserve = path.join(hooksDir, 'observe.sh');
+  fs.writeFileSync(testObserve, observeContent, { mode: 0o755 });
+
+  return { testDir, projectDir, testObserve };
+}
+
+// Run observe.sh once against the sandbox. Resolves when the process exits.
+function runObserve(testObserve, projectDir) {
+  const input = JSON.stringify({
+    tool_name: 'Read',
+    tool_input: { file_path: '/tmp/test.txt' },
+    session_id: 'test-session',
+    cwd: projectDir
+  });
+  return new Promise((resolve, reject) => {
+    const child = spawn('bash', [testObserve, 'post'], {
+      env: {
+        ...process.env,
+        HOME: projectDir,
+        CLAUDE_CODE_ENTRYPOINT: 'cli',
+        ECC_HOOK_PROFILE: 'standard',
+        ECC_SKIP_OBSERVE: '0',
+        CLAUDE_PROJECT_DIR: projectDir,
+        // Reset threshold far above the invocation count, so no reset fires and
+        // the final counter equals the number of invocations.
+        ECC_OBSERVER_SIGNAL_EVERY_N: '100000'
+      },
+      stdio: ['pipe', 'ignore', 'pipe']
+    });
+    let stderr = '';
+    // Fail the test on a hung hook rather than waiting forever.
+    const timer = setTimeout(() => {
+      child.kill('SIGKILL');
+      reject(new Error('observe.sh timed out'));
+    }, 20000);
+    child.stderr.on('data', (chunk) => { stderr += chunk; });
+    // A broken observe.sh must fail the test, not be silently swallowed.
+    child.on('close', (code, signal) => {
+      clearTimeout(timer);
+      if (code === 0 && signal === null) {
+        resolve();
+      } else {
+        reject(new Error(`observe.sh failed code=${code} signal=${signal}: ${stderr.trim()}`));
+      }
+    });
+    child.on('error', (err) => {
+      clearTimeout(timer);
+      reject(err);
+    });
+    child.stdin.end(input);
+  });
+}
+
+function readCounter(projectDir) {
+  const counterFile = path.join(projectDir, '.observer-signal-counter');
+  if (!fs.existsSync(counterFile)) {
+    return null;
+  }
+  return parseInt(fs.readFileSync(counterFile, 'utf8').trim(), 10);
+}
+
+console.log('\n=== observe.sh signal-counter race regression (#2296) ===\n');
+
+test('observe.sh uses a lock around the throttle-counter update', () => {
+  const content = fs.readFileSync(observeShPath, 'utf8');
+  assert.ok(
+    content.includes('SIGNAL_COUNTER_LOCK'),
+    'observe.sh should define a lock for the signal counter'
+  );
+  assert.ok(
+    /flock 8\b/.test(content) || /mkdir "\$SIGNAL_COUNTER_LOCK"/.test(content),
+    'observe.sh should acquire the counter lock via flock or an atomic mkdir'
+  );
+});
+
+test('observe.sh guards against a corrupt (non-integer) counter file', () => {
+  const content = fs.readFileSync(observeShPath, 'utf8');
+  assert.ok(
+    /''\|\*\[!0-9\]\*\) counter=0/.test(content),
+    'observe.sh should reset a non-integer counter to 0 before incrementing'
+  );
+});
+
+async function runSequential() {
+  const { testDir, projectDir, testObserve } = buildSandbox();
+  try {
+    const N = 5;
+    for (let i = 0; i < N; i++) {
+      await runObserve(testObserve, projectDir);
+    }
+    const counter = readCounter(projectDir);
+    assert.notStrictEqual(counter, null, 'counter file should exist after runs');
+    assert.strictEqual(counter, N, `sequential counter should be ${N}, got ${counter}`);
+  } finally {
+    cleanupDir(testDir);
+  }
+}
+
+async function runConcurrent() {
+  const { testDir, projectDir, testObserve } = buildSandbox();
+  try {
+    const K = 20;
+    // Spawn all K before awaiting any, so they genuinely contend on the counter.
+    const runs = [];
+    for (let i = 0; i < K; i++) {
+      runs.push(runObserve(testObserve, projectDir));
+    }
+    await Promise.all(runs);
+    const counter = readCounter(projectDir);
+    assert.notStrictEqual(counter, null, 'counter file should exist after concurrent runs');
+    if (hasFlock) {
+      // flock serializes every invocation, so no increment is ever lost. The
+      // pre-fix unlocked code drops increments under this same contention.
+      assert.strictEqual(
+        counter,
+        K,
+        `with flock the counter must be exactly ${K}, got ${counter}`
+      );
+    } else {
+      // mkdir fallback is best-effort: it may drop at most one increment if its
+      // bounded spin is exhausted, but never the multi-increment loss the
+      // unlocked code exhibited.
+      assert.ok(
+        counter >= K - 1,
+        `mkdir fallback should keep the counter >= ${K - 1}, got ${counter}`
+      );
+    }
+  } finally {
+    cleanupDir(testDir);
+  }
+}
+
+(async () => {
+  if (!isWindows && hasPython) {
+    await asyncTest('sequential invocations increment the counter exactly once each', runSequential);
+    await asyncTest('concurrent invocations never lose a counter increment', runConcurrent);
+  } else {
+    console.log('  - skipping shell-execution tests (requires non-Windows + python3)');
+  }
+
+  console.log('\n=== Test Results ===');
+  console.log(`Passed: ${passed}`);
+  console.log(`Failed: ${failed}`);
+  console.log(`Total:  ${passed + failed}`);
+
+  process.exit(failed > 0 ? 1 : 0);
+})();