fullcourtdefense-cli 1.26.13 → 1.26.14
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/commands/daemon.js +83 -6
- package/dist/commands/watchdog.js +14 -0
- package/dist/daemonForensics.d.ts +74 -0
- package/dist/daemonForensics.js +103 -0
- package/dist/integrity.d.ts +2 -0
- package/dist/integrity.js +22 -0
- package/dist/version.json +1 -1
- package/package.json +2 -1
package/dist/commands/daemon.js
CHANGED
|
@@ -109,16 +109,42 @@ const DEBOUNCE_MS = envMs('FCD_DAEMON_DEBOUNCE_MS', 2_000);
|
|
|
109
109
|
const RESCAN_INTERVAL_MS = envMs('FCD_DAEMON_RESCAN_MS', 5 * 60_000);
|
|
110
110
|
/** Heartbeat / spool flush cadence. */
|
|
111
111
|
const HEARTBEAT_INTERVAL_MS = envMs('FCD_DAEMON_HEARTBEAT_MS', 5 * 60_000);
|
|
112
|
+
/**
|
|
113
|
+
* Floor between integrity-triggered repairs. The drift watcher reacts to config
|
|
114
|
+
* CHANGES; this path reacts to a failing verdict, which persists until fixed. If
|
|
115
|
+
* a repair cannot fix it (no write permission, a config owned by another
|
|
116
|
+
* install) an uncooled retry would re-run protect-all every heartbeat forever.
|
|
117
|
+
*/
|
|
118
|
+
const INTEGRITY_HEAL_COOLDOWN_MS = envMs('FCD_INTEGRITY_HEAL_COOLDOWN_MS', 60 * 60_000);
|
|
112
119
|
/** Bundle (mode / suspension / policy version) poll cadence. */
|
|
113
120
|
const BUNDLE_POLL_MS = envMs('FCD_DAEMON_BUNDLE_POLL_MS', 60_000);
|
|
121
|
+
/**
|
|
122
|
+
* Fetch budget for the daemon's background bundle poll. Deliberately NOT the
|
|
123
|
+
* hot-path default (getRuntimeBundle's 1.5s), which exists so a hook never
|
|
124
|
+
* stalls a developer's command. Nothing waits on this poll — it runs on its own
|
|
125
|
+
* 60s timer inside a 45s deadline — so a slow-but-healthy control plane must be
|
|
126
|
+
* waited out, not reported as an outage. At 1.5s roughly 10% of polls aborted
|
|
127
|
+
* on ordinary p95 latency spikes and each one raised a `network_down` distress
|
|
128
|
+
* signal, while `doctor --health` (already passing 10s) stayed green against
|
|
129
|
+
* the same backend: alert noise with no failure behind it.
|
|
130
|
+
*/
|
|
131
|
+
const BUNDLE_FETCH_TIMEOUT_MS = envMs('FCD_DAEMON_BUNDLE_TIMEOUT_MS', 10_000);
|
|
114
132
|
/**
|
|
115
133
|
* Hard ceilings for one poll attempt (see startPollLoop). Generous multiples of
|
|
116
|
-
* the work each poll actually does — a bundle poll is a
|
|
134
|
+
* the work each poll actually does — a bundle poll is a 10s fetch plus an 8s
|
|
117
135
|
* snapshot refresh, a heartbeat is a spool flush plus a log upload — so these
|
|
118
136
|
* only ever fire on a genuinely hung socket, never on a merely slow network.
|
|
119
137
|
*/
|
|
120
138
|
const BUNDLE_POLL_DEADLINE_MS = envMs('FCD_DAEMON_BUNDLE_DEADLINE_MS', 45_000);
|
|
121
139
|
const HEARTBEAT_DEADLINE_MS = envMs('FCD_DAEMON_HEARTBEAT_DEADLINE_MS', 120_000);
|
|
140
|
+
/**
|
|
141
|
+
* Wake-from-sleep detection. Timers are the only awake clock we have: a beat
|
|
142
|
+
* that observes far more wall time than its own interval proves the machine
|
|
143
|
+
* was suspended in between. The jump bound is a generous multiple of the beat
|
|
144
|
+
* so ordinary event-loop lag or a busy CPU can never look like a resume.
|
|
145
|
+
*/
|
|
146
|
+
const RESUME_BEAT_MS = envMs('FCD_DAEMON_RESUME_BEAT_MS', 30_000);
|
|
147
|
+
const RESUME_JUMP_MS = envMs('FCD_DAEMON_RESUME_JUMP_MS', 120_000);
|
|
122
148
|
/** Delay before the one-time initial discovery sweep on a fresh machine. */
|
|
123
149
|
const INITIAL_DISCOVER_DELAY_MS = envMs('FCD_DAEMON_INITIAL_DISCOVER_MS', 2 * 60_000);
|
|
124
150
|
/** A discovery upload older than this is stale — the daemon catches up itself. */
|
|
@@ -505,6 +531,7 @@ async function runDaemon(args, config) {
|
|
|
505
531
|
let quietUntil = 0; // ignore events until this time (self-writes)
|
|
506
532
|
let debounceTimer = null;
|
|
507
533
|
let reprotecting = false;
|
|
534
|
+
let lastIntegrityHealAt = 0; // cooldown anchor for integrity-triggered repairs
|
|
508
535
|
let suspended = false;
|
|
509
536
|
let stopped = false;
|
|
510
537
|
const executingActionIds = new Set();
|
|
@@ -582,16 +609,23 @@ async function runDaemon(args, config) {
|
|
|
582
609
|
}
|
|
583
610
|
catch { /* next tick */ }
|
|
584
611
|
};
|
|
585
|
-
|
|
612
|
+
/**
|
|
613
|
+
* One repair path, two triggers: a watched config CHANGED (drift), or the
|
|
614
|
+
* integrity check reported a reason protect-all can fix (integrity). Keeping
|
|
615
|
+
* them on the same function preserves the reprotecting/quiet-window guards —
|
|
616
|
+
* a second copy would race this one and re-enter protect-all concurrently.
|
|
617
|
+
*/
|
|
618
|
+
const reprotect = async (reasonPaths, cause = 'drift') => {
|
|
619
|
+
const headline = cause === 'integrity' ? 'Integrity self-heal' : 'Config drift detected';
|
|
586
620
|
if (reprotecting || stopped)
|
|
587
621
|
return;
|
|
588
622
|
if (suspended) {
|
|
589
|
-
log(
|
|
623
|
+
log(`${headline} (${reasonPaths.join(', ')}) but machine is suspended — not re-protecting.`);
|
|
590
624
|
return;
|
|
591
625
|
}
|
|
592
626
|
reprotecting = true;
|
|
593
627
|
quietUntil = Date.now() + SELF_WRITE_QUIET_MS;
|
|
594
|
-
log(
|
|
628
|
+
log(`${headline}: ${reasonPaths.join(', ')} — re-running protect-all.`);
|
|
595
629
|
try {
|
|
596
630
|
const repaired = await (0, protectionRepair_1.repairProtection)({ ...args, dryRun: undefined }, config);
|
|
597
631
|
quietUntil = Date.now() + SELF_WRITE_QUIET_MS;
|
|
@@ -601,7 +635,9 @@ async function runDaemon(args, config) {
|
|
|
601
635
|
if (!quiet) {
|
|
602
636
|
(0, notify_1.notifyOs)({
|
|
603
637
|
title: 'FullCourtDefense re-protected this machine',
|
|
604
|
-
message:
|
|
638
|
+
message: cause === 'integrity'
|
|
639
|
+
? 'Protection was reported incomplete; the FullCourtDefense gateway was re-applied.'
|
|
640
|
+
: 'An MCP or hook config changed; the FullCourtDefense gateway was re-applied.',
|
|
605
641
|
url: (0, notify_1.consoleUrl)('/agent-security/users?view=desktop'),
|
|
606
642
|
});
|
|
607
643
|
}
|
|
@@ -990,6 +1026,7 @@ async function runDaemon(args, config) {
|
|
|
990
1026
|
machineId: identity.machineId,
|
|
991
1027
|
force: true,
|
|
992
1028
|
ttlMs: 0,
|
|
1029
|
+
timeoutMs: BUNDLE_FETCH_TIMEOUT_MS,
|
|
993
1030
|
});
|
|
994
1031
|
log(bundle.policyHash
|
|
995
1032
|
? `Policy refresh: bundle applied (hash ${bundle.policyHash.slice(0, 12)}…, ${bundle.policyCount ?? 0} policies).`
|
|
@@ -1155,6 +1192,7 @@ async function runDaemon(args, config) {
|
|
|
1155
1192
|
machineName: identity.hostname,
|
|
1156
1193
|
machineId: identity.machineId,
|
|
1157
1194
|
force: true,
|
|
1195
|
+
timeoutMs: BUNDLE_FETCH_TIMEOUT_MS,
|
|
1158
1196
|
});
|
|
1159
1197
|
if (typeof bundle.pollIntervalMs === 'number' && Number.isFinite(bundle.pollIntervalMs)) {
|
|
1160
1198
|
bundlePollBaseMs = Math.min(Math.max(bundle.pollIntervalMs, 15_000), 60 * 60_000);
|
|
@@ -1261,7 +1299,24 @@ async function runDaemon(args, config) {
|
|
|
1261
1299
|
if (!creds.shieldId)
|
|
1262
1300
|
return;
|
|
1263
1301
|
try {
|
|
1264
|
-
|
|
1302
|
+
let integrity = (0, integrity_1.getLocalIntegrityReport)({ requireDaemon: true });
|
|
1303
|
+
// A failing verdict used to be REPORTED and nothing more, so a machine
|
|
1304
|
+
// could sit "Damaged" in the console indefinitely on a reason the daemon
|
|
1305
|
+
// already knows how to fix — the drift watcher only fires when a config
|
|
1306
|
+
// file changes, and a stale gateway path changes nothing. Heal first, then
|
|
1307
|
+
// report what is true AFTER the repair rather than the verdict that
|
|
1308
|
+
// triggered it.
|
|
1309
|
+
if (!integrity.ok
|
|
1310
|
+
&& !suspended
|
|
1311
|
+
&& Date.now() - lastIntegrityHealAt >= INTEGRITY_HEAL_COOLDOWN_MS
|
|
1312
|
+
&& (0, integrity_1.hasHealableIntegrityReason)(integrity.reasons)) {
|
|
1313
|
+
lastIntegrityHealAt = Date.now();
|
|
1314
|
+
await reprotect(integrity.reasons, 'integrity');
|
|
1315
|
+
integrity = (0, integrity_1.getLocalIntegrityReport)({ requireDaemon: true });
|
|
1316
|
+
log(integrity.ok
|
|
1317
|
+
? 'Integrity self-heal: protection restored.'
|
|
1318
|
+
: `Integrity self-heal: still failing (${integrity.reasons.join(', ')}) — reported for support.`);
|
|
1319
|
+
}
|
|
1265
1320
|
// Report a previous daemon's unclean death exactly once.
|
|
1266
1321
|
const crash = (0, daemonForensics_1.readPostmortem)();
|
|
1267
1322
|
const unreportedCrash = crash && !crash.reportedAt ? crash : undefined;
|
|
@@ -1471,6 +1526,27 @@ async function runDaemon(args, config) {
|
|
|
1471
1526
|
},
|
|
1472
1527
|
onError: error => log(`Heartbeat error: ${error instanceof Error ? error.message : String(error)}`),
|
|
1473
1528
|
});
|
|
1529
|
+
// ── Wake-from-sleep recovery ──────────────────────────────────────────────
|
|
1530
|
+
// Nothing in the agent reacts to a resume, so after a laptop wakes the
|
|
1531
|
+
// machine kept reporting "daemon down" until the next SCHEDULED heartbeat —
|
|
1532
|
+
// a full cadence later. Worse, the liveness marker stayed stale for that
|
|
1533
|
+
// whole window, which is what the watchdog reads before deciding the daemon
|
|
1534
|
+
// is wedged. Refreshing the marker the moment we notice the wall-clock jump
|
|
1535
|
+
// closes the reporting gap and removes the evidence of a hang that never
|
|
1536
|
+
// happened. Both calls go through runWithDeadline, so a resume that lands on
|
|
1537
|
+
// sockets which died during sleep cannot leak an unhandled rejection.
|
|
1538
|
+
let lastResumeBeatMs = Date.now();
|
|
1539
|
+
const resumeTimer = setInterval(() => {
|
|
1540
|
+
const now = Date.now();
|
|
1541
|
+
const drift = now - lastResumeBeatMs;
|
|
1542
|
+
lastResumeBeatMs = now;
|
|
1543
|
+
if (drift < RESUME_JUMP_MS)
|
|
1544
|
+
return;
|
|
1545
|
+
log(`Resume detected: ${Math.round(drift / 1000)}s of wall time elapsed while suspended — refreshing liveness and reporting in now.`);
|
|
1546
|
+
(0, daemonForensics_1.touchDaemonAlive)();
|
|
1547
|
+
void (0, pollLoop_1.runWithDeadline)({ run: heartbeat, deadlineMs: HEARTBEAT_DEADLINE_MS });
|
|
1548
|
+
void (0, pollLoop_1.runWithDeadline)({ run: pollBundle, deadlineMs: BUNDLE_POLL_DEADLINE_MS });
|
|
1549
|
+
}, RESUME_BEAT_MS);
|
|
1474
1550
|
// PowerShell transcript retention (Windows): the Transcription policy FCD
|
|
1475
1551
|
// enables writes a file per session forever — prune anything older than the
|
|
1476
1552
|
// retention window once a day (plus once shortly after boot, so laptops
|
|
@@ -1496,6 +1572,7 @@ async function runDaemon(args, config) {
|
|
|
1496
1572
|
clearInterval(rescanTimer);
|
|
1497
1573
|
bundleLoop.stop();
|
|
1498
1574
|
heartbeatLoop.stop();
|
|
1575
|
+
clearInterval(resumeTimer);
|
|
1499
1576
|
clearTimeout(transcriptPruneBootTimer);
|
|
1500
1577
|
clearInterval(transcriptPruneTimer);
|
|
1501
1578
|
clearTimeout(discoverCatchUpBootTimer);
|
|
@@ -66,6 +66,16 @@ async function watchdogCommand(args, config) {
|
|
|
66
66
|
const state = (0, daemon_1.daemonRuntimeState)();
|
|
67
67
|
let relaunched = false;
|
|
68
68
|
let postmortem = (0, daemonForensics_1.readPostmortem)();
|
|
69
|
+
// Our own tick cadence is the only awake clock available (see
|
|
70
|
+
// detectMachineSuspension): a gap far larger than the interval proves the
|
|
71
|
+
// MACHINE was suspended, which fully explains a stale daemon marker. Read
|
|
72
|
+
// the previous tick BEFORE recording this one, and record unconditionally so
|
|
73
|
+
// an early return below still leaves the next tick a baseline.
|
|
74
|
+
const machineSuspended = (0, daemonForensics_1.detectMachineSuspension)({
|
|
75
|
+
previousTickAt: (0, daemonForensics_1.readWatchdogTick)()?.lastTickAt,
|
|
76
|
+
nowMs: Date.now(),
|
|
77
|
+
});
|
|
78
|
+
(0, daemonForensics_1.recordWatchdogTick)();
|
|
69
79
|
// Beyond "the pid exists": a wedged daemon or a recycled pid both pass a
|
|
70
80
|
// plain pid check while the machine has no working control plane. Classify
|
|
71
81
|
// with the alive-marker freshness + a verdict-pipe probe.
|
|
@@ -77,6 +87,7 @@ async function watchdogCommand(args, config) {
|
|
|
77
87
|
marker: (0, daemonForensics_1.readAliveMarker)(),
|
|
78
88
|
staleAfterMs: (0, daemonForensics_1.daemonHungAfterMs)(),
|
|
79
89
|
pipeProbe: await (0, verdictIpcClient_1.probeVerdictServer)(1_500),
|
|
90
|
+
machineSuspended,
|
|
80
91
|
});
|
|
81
92
|
}
|
|
82
93
|
if (health === 'hung' && state.pid && process.env.FCD_WATCHDOG_NO_RELAUNCH !== '1') {
|
|
@@ -85,6 +96,9 @@ async function watchdogCommand(args, config) {
|
|
|
85
96
|
// clean. The distress ledger survives; the fresh daemon's first heartbeat
|
|
86
97
|
// ships the story to the fleet.
|
|
87
98
|
(0, distress_1.reportDistress)('watchdog', distress_1.DISTRESS.DAEMON_HUNG, `daemon pid ${state.pid} unresponsive (alive-marker stale + verdict pipe silent) — recycled by watchdog`);
|
|
99
|
+
// Claim the kill first: otherwise the next daemon boots, finds a marker
|
|
100
|
+
// with no clean-exit stamp, and reports EDR/AV killed it.
|
|
101
|
+
(0, daemonForensics_1.recordWatchdogRecycle)(state.pid);
|
|
88
102
|
(0, daemon_1.forceStopPid)(state.pid);
|
|
89
103
|
}
|
|
90
104
|
// NOTE for 'pid_reused': the pid belongs to an UNRELATED process — never
|
|
@@ -98,7 +98,81 @@ export declare function classifyDaemonHealth(input: {
|
|
|
98
98
|
marker: DaemonAliveMarker | undefined;
|
|
99
99
|
staleAfterMs: number;
|
|
100
100
|
pipeProbe: 'responsive' | 'unresponsive' | 'no-endpoint' | 'not-probed';
|
|
101
|
+
/**
|
|
102
|
+
* True when the machine itself stopped running since the previous check
|
|
103
|
+
* (sleep/hibernate). Marker staleness is then not evidence of anything —
|
|
104
|
+
* see detectMachineSuspension.
|
|
105
|
+
*/
|
|
106
|
+
machineSuspended?: boolean;
|
|
101
107
|
nowMs?: number;
|
|
102
108
|
}): DaemonHealthClass;
|
|
109
|
+
/**
|
|
110
|
+
* WHY THIS EXISTS (a real developer-machine incident, 2026-08)
|
|
111
|
+
* Staleness above is a WALL-CLOCK measurement, but the daemon can only refresh
|
|
112
|
+
* its marker while the CPU is actually running. A laptop that sleeps longer
|
|
113
|
+
* than `staleAfterMs` therefore produces a stale marker through no fault of
|
|
114
|
+
* the daemon — and the watchdog read that as "wedged" and killed a perfectly
|
|
115
|
+
* healthy process on every lid-open. One developer machine was recycled 8
|
|
116
|
+
* times in 7 days, each kill also writing a post-mortem blaming EDR/AV for
|
|
117
|
+
* what was in fact our own watchdog.
|
|
118
|
+
*
|
|
119
|
+
* There is no portable awake clock to compare against. On Windows
|
|
120
|
+
* `os.uptime()` still INCLUDES sleep (measured on the incident machine: 26.66h
|
|
121
|
+
* of "uptime" spanning a 13h sleep), and reading the kernel power log would
|
|
122
|
+
* mean spawning PowerShell every 5 minutes — precisely the pattern that gets
|
|
123
|
+
* this agent killed by EDR, which the watchdog is documented to avoid.
|
|
124
|
+
*
|
|
125
|
+
* What we do have is our OWN cadence. A fixed-interval scheduled task that
|
|
126
|
+
* observes far more wall time between two of its own ticks than its interval
|
|
127
|
+
* allows has proven the machine was not running in between. That gap explains
|
|
128
|
+
* the daemon's stale marker, so staleness stops counting as evidence.
|
|
129
|
+
*/
|
|
130
|
+
/** The watchdog scheduled task's interval — its ticks are the awake clock. */
|
|
131
|
+
export declare const WATCHDOG_CADENCE_MS: number;
|
|
132
|
+
/** Gaps beyond cadence x this factor cannot be explained by jitter or a slow tick. */
|
|
133
|
+
export declare const SUSPEND_GAP_FACTOR = 3;
|
|
134
|
+
/** The watchdog's own proof that IT was running, persisted between ticks. */
|
|
135
|
+
export interface WatchdogTick {
|
|
136
|
+
lastTickAt: string;
|
|
137
|
+
}
|
|
138
|
+
export declare function watchdogTickFile(): string;
|
|
139
|
+
export declare function readWatchdogTick(): WatchdogTick | undefined;
|
|
140
|
+
export declare function recordWatchdogTick(now?: Date): void;
|
|
141
|
+
/**
|
|
142
|
+
* Pure: did the machine stop running between the previous watchdog tick and
|
|
143
|
+
* this one?
|
|
144
|
+
*
|
|
145
|
+
* A MISSING baseline answers FALSE (i.e. "not suspended", the pre-existing
|
|
146
|
+
* behavior) rather than granting grace. That looks backwards next to "never
|
|
147
|
+
* kill on a guess", but it is the safer failure mode overall: the grace
|
|
148
|
+
* depends on a tick file, and if that file can never be persisted (locked
|
|
149
|
+
* directory, full disk, AV quarantine) then every tick would see no baseline,
|
|
150
|
+
* grant grace forever, and silently disable the watchdog's ability to recycle
|
|
151
|
+
* a genuinely wedged daemon. Answering false means a persistence failure
|
|
152
|
+
* degrades to the known old behavior instead of to no protection at all, and
|
|
153
|
+
* costs at most one stale verdict on the very first tick after upgrade.
|
|
154
|
+
*
|
|
155
|
+
* A baseline that EXISTS but cannot be trusted (unparseable, or the clock
|
|
156
|
+
* moved backwards under us) still answers TRUE: there we have positive
|
|
157
|
+
* evidence that time is not measurable, which is exactly when a staleness
|
|
158
|
+
* comparison must not be believed.
|
|
159
|
+
*/
|
|
160
|
+
export declare function detectMachineSuspension(input: {
|
|
161
|
+
previousTickAt?: string;
|
|
162
|
+
nowMs: number;
|
|
163
|
+
cadenceMs?: number;
|
|
164
|
+
factor?: number;
|
|
165
|
+
}): boolean;
|
|
166
|
+
/**
|
|
167
|
+
* Stamp the marker before the watchdog force-stops a wedged daemon, so the
|
|
168
|
+
* NEXT boot's `detectUncleanDeath` does not report our own recycle as an
|
|
169
|
+
* external kill. The reason is distinct from a real shutdown's signal name,
|
|
170
|
+
* and the recycle stays visible via the `daemon_hung` distress entry and the
|
|
171
|
+
* beacon's `daemonHealth` field — this only suppresses the false EDR/AV story.
|
|
172
|
+
*
|
|
173
|
+
* Keyed on the DAEMON's pid (not `process.pid`): the watchdog is a different
|
|
174
|
+
* process, which is exactly why `recordCleanExit` cannot be reused here.
|
|
175
|
+
*/
|
|
176
|
+
export declare function recordWatchdogRecycle(daemonPid: number): void;
|
|
103
177
|
/** Stamp the post-mortem as delivered so heartbeats stop re-sending it. */
|
|
104
178
|
export declare function markPostmortemReported(): void;
|
package/dist/daemonForensics.js
CHANGED
|
@@ -33,6 +33,7 @@ var __importStar = (this && this.__importStar) || (function () {
|
|
|
33
33
|
};
|
|
34
34
|
})();
|
|
35
35
|
Object.defineProperty(exports, "__esModule", { value: true });
|
|
36
|
+
exports.SUSPEND_GAP_FACTOR = exports.WATCHDOG_CADENCE_MS = void 0;
|
|
36
37
|
exports.aliveMarkerFile = aliveMarkerFile;
|
|
37
38
|
exports.postmortemFile = postmortemFile;
|
|
38
39
|
exports.readAliveMarker = readAliveMarker;
|
|
@@ -45,6 +46,11 @@ exports.touchDaemonAlive = touchDaemonAlive;
|
|
|
45
46
|
exports.recordCleanExit = recordCleanExit;
|
|
46
47
|
exports.daemonHungAfterMs = daemonHungAfterMs;
|
|
47
48
|
exports.classifyDaemonHealth = classifyDaemonHealth;
|
|
49
|
+
exports.watchdogTickFile = watchdogTickFile;
|
|
50
|
+
exports.readWatchdogTick = readWatchdogTick;
|
|
51
|
+
exports.recordWatchdogTick = recordWatchdogTick;
|
|
52
|
+
exports.detectMachineSuspension = detectMachineSuspension;
|
|
53
|
+
exports.recordWatchdogRecycle = recordWatchdogRecycle;
|
|
48
54
|
exports.markPostmortemReported = markPostmortemReported;
|
|
49
55
|
const fs = __importStar(require("fs"));
|
|
50
56
|
const os = __importStar(require("os"));
|
|
@@ -199,12 +205,109 @@ function classifyDaemonHealth(input) {
|
|
|
199
205
|
return 'healthy';
|
|
200
206
|
if (input.pipeProbe === 'responsive')
|
|
201
207
|
return 'healthy';
|
|
208
|
+
// The marker is stale, but the daemon had no CPU with which to refresh it.
|
|
209
|
+
// Judging it now would convict a healthy daemon of the machine's sleep, so
|
|
210
|
+
// defer: the next tick runs after a full awake interval and decides fairly.
|
|
211
|
+
if (input.machineSuspended)
|
|
212
|
+
return 'indeterminate';
|
|
202
213
|
if (input.pipeProbe === 'unresponsive')
|
|
203
214
|
return 'hung';
|
|
204
215
|
if (input.pipeProbe === 'no-endpoint')
|
|
205
216
|
return 'pid_reused';
|
|
206
217
|
return 'indeterminate';
|
|
207
218
|
}
|
|
219
|
+
// ---------------------------------------------------------------------------
|
|
220
|
+
// Suspend awareness
|
|
221
|
+
// ---------------------------------------------------------------------------
|
|
222
|
+
/**
|
|
223
|
+
* WHY THIS EXISTS (a real developer-machine incident, 2026-08)
|
|
224
|
+
* Staleness above is a WALL-CLOCK measurement, but the daemon can only refresh
|
|
225
|
+
* its marker while the CPU is actually running. A laptop that sleeps longer
|
|
226
|
+
* than `staleAfterMs` therefore produces a stale marker through no fault of
|
|
227
|
+
* the daemon — and the watchdog read that as "wedged" and killed a perfectly
|
|
228
|
+
* healthy process on every lid-open. One developer machine was recycled 8
|
|
229
|
+
* times in 7 days, each kill also writing a post-mortem blaming EDR/AV for
|
|
230
|
+
* what was in fact our own watchdog.
|
|
231
|
+
*
|
|
232
|
+
* There is no portable awake clock to compare against. On Windows
|
|
233
|
+
* `os.uptime()` still INCLUDES sleep (measured on the incident machine: 26.66h
|
|
234
|
+
* of "uptime" spanning a 13h sleep), and reading the kernel power log would
|
|
235
|
+
* mean spawning PowerShell every 5 minutes — precisely the pattern that gets
|
|
236
|
+
* this agent killed by EDR, which the watchdog is documented to avoid.
|
|
237
|
+
*
|
|
238
|
+
* What we do have is our OWN cadence. A fixed-interval scheduled task that
|
|
239
|
+
* observes far more wall time between two of its own ticks than its interval
|
|
240
|
+
* allows has proven the machine was not running in between. That gap explains
|
|
241
|
+
* the daemon's stale marker, so staleness stops counting as evidence.
|
|
242
|
+
*/
|
|
243
|
+
/** The watchdog scheduled task's interval — its ticks are the awake clock. */
|
|
244
|
+
exports.WATCHDOG_CADENCE_MS = 5 * 60_000;
|
|
245
|
+
/** Gaps beyond cadence x this factor cannot be explained by jitter or a slow tick. */
|
|
246
|
+
exports.SUSPEND_GAP_FACTOR = 3;
|
|
247
|
+
function watchdogTickFile() {
|
|
248
|
+
return path.join(stateDir(), 'watchdog.tick.json');
|
|
249
|
+
}
|
|
250
|
+
function readWatchdogTick() {
|
|
251
|
+
try {
|
|
252
|
+
const tick = JSON.parse(fs.readFileSync(watchdogTickFile(), 'utf8'));
|
|
253
|
+
return tick && typeof tick.lastTickAt === 'string' ? tick : undefined;
|
|
254
|
+
}
|
|
255
|
+
catch {
|
|
256
|
+
return undefined;
|
|
257
|
+
}
|
|
258
|
+
}
|
|
259
|
+
function recordWatchdogTick(now = new Date()) {
|
|
260
|
+
writeJson(watchdogTickFile(), { lastTickAt: now.toISOString() });
|
|
261
|
+
}
|
|
262
|
+
/**
|
|
263
|
+
* Pure: did the machine stop running between the previous watchdog tick and
|
|
264
|
+
* this one?
|
|
265
|
+
*
|
|
266
|
+
* A MISSING baseline answers FALSE (i.e. "not suspended", the pre-existing
|
|
267
|
+
* behavior) rather than granting grace. That looks backwards next to "never
|
|
268
|
+
* kill on a guess", but it is the safer failure mode overall: the grace
|
|
269
|
+
* depends on a tick file, and if that file can never be persisted (locked
|
|
270
|
+
* directory, full disk, AV quarantine) then every tick would see no baseline,
|
|
271
|
+
* grant grace forever, and silently disable the watchdog's ability to recycle
|
|
272
|
+
* a genuinely wedged daemon. Answering false means a persistence failure
|
|
273
|
+
* degrades to the known old behavior instead of to no protection at all, and
|
|
274
|
+
* costs at most one stale verdict on the very first tick after upgrade.
|
|
275
|
+
*
|
|
276
|
+
* A baseline that EXISTS but cannot be trusted (unparseable, or the clock
|
|
277
|
+
* moved backwards under us) still answers TRUE: there we have positive
|
|
278
|
+
* evidence that time is not measurable, which is exactly when a staleness
|
|
279
|
+
* comparison must not be believed.
|
|
280
|
+
*/
|
|
281
|
+
function detectMachineSuspension(input) {
|
|
282
|
+
const cadence = input.cadenceMs && input.cadenceMs > 0 ? input.cadenceMs : exports.WATCHDOG_CADENCE_MS;
|
|
283
|
+
const factor = input.factor && input.factor > 0 ? input.factor : exports.SUSPEND_GAP_FACTOR;
|
|
284
|
+
if (!input.previousTickAt)
|
|
285
|
+
return false;
|
|
286
|
+
const previous = Date.parse(input.previousTickAt);
|
|
287
|
+
if (!Number.isFinite(previous))
|
|
288
|
+
return true;
|
|
289
|
+
const gap = input.nowMs - previous;
|
|
290
|
+
if (gap < 0)
|
|
291
|
+
return true;
|
|
292
|
+
return gap > cadence * factor;
|
|
293
|
+
}
|
|
294
|
+
/**
|
|
295
|
+
* Stamp the marker before the watchdog force-stops a wedged daemon, so the
|
|
296
|
+
* NEXT boot's `detectUncleanDeath` does not report our own recycle as an
|
|
297
|
+
* external kill. The reason is distinct from a real shutdown's signal name,
|
|
298
|
+
* and the recycle stays visible via the `daemon_hung` distress entry and the
|
|
299
|
+
* beacon's `daemonHealth` field — this only suppresses the false EDR/AV story.
|
|
300
|
+
*
|
|
301
|
+
* Keyed on the DAEMON's pid (not `process.pid`): the watchdog is a different
|
|
302
|
+
* process, which is exactly why `recordCleanExit` cannot be reused here.
|
|
303
|
+
*/
|
|
304
|
+
function recordWatchdogRecycle(daemonPid) {
|
|
305
|
+
const marker = readAliveMarker();
|
|
306
|
+
if (!marker || marker.pid !== daemonPid)
|
|
307
|
+
return;
|
|
308
|
+
marker.cleanExit = { at: new Date().toISOString(), reason: 'watchdog-recycle' };
|
|
309
|
+
writeJson(aliveMarkerFile(), marker);
|
|
310
|
+
}
|
|
208
311
|
/** Stamp the post-mortem as delivered so heartbeats stop re-sending it. */
|
|
209
312
|
function markPostmortemReported() {
|
|
210
313
|
const pm = readPostmortem();
|
package/dist/integrity.d.ts
CHANGED
|
@@ -1,4 +1,6 @@
|
|
|
1
1
|
export type IntegrityReason = 'mcp_gateway_missing' | 'mcp_gateway_stale_path' | 'cursor_hook_missing' | 'cursor_hook_stale_path' | 'cursor_hook_changed' | 'claude_hook_missing' | 'claude_hook_stale_path' | 'claude_hook_changed' | 'daemon_not_running' | 'runtime_bundle_stale';
|
|
2
|
+
/** Pure: does this verdict contain anything a protect-all pass could fix? */
|
|
3
|
+
export declare function hasHealableIntegrityReason(reasons: readonly string[]): boolean;
|
|
2
4
|
export interface LocalIntegrityReport {
|
|
3
5
|
ok: boolean;
|
|
4
6
|
reasons: IntegrityReason[];
|
package/dist/integrity.js
CHANGED
|
@@ -33,6 +33,7 @@ var __importStar = (this && this.__importStar) || (function () {
|
|
|
33
33
|
};
|
|
34
34
|
})();
|
|
35
35
|
Object.defineProperty(exports, "__esModule", { value: true });
|
|
36
|
+
exports.hasHealableIntegrityReason = hasHealableIntegrityReason;
|
|
36
37
|
exports.isDaemonRunning = isDaemonRunning;
|
|
37
38
|
exports.getLocalIntegrityReport = getLocalIntegrityReport;
|
|
38
39
|
const fs = __importStar(require("fs"));
|
|
@@ -42,6 +43,27 @@ const installClaudeHook_1 = require("./commands/installClaudeHook");
|
|
|
42
43
|
const installCursorHook_1 = require("./commands/installCursorHook");
|
|
43
44
|
const mcpGateway_1 = require("./commands/mcpGateway");
|
|
44
45
|
const appDetection_1 = require("./appDetection");
|
|
46
|
+
/**
|
|
47
|
+
* The reasons `protect-all` can actually repair. Detecting a failure is only
|
|
48
|
+
* half the job: the daemon heals these instead of reporting them forever, which
|
|
49
|
+
* is why the taxonomy lives beside the reason type rather than in the caller.
|
|
50
|
+
*
|
|
51
|
+
* Excluded on purpose:
|
|
52
|
+
* - `daemon_not_running` — the daemon is the thing evaluating this; re-wrapping
|
|
53
|
+
* configs cannot start a process that is already running or already dead.
|
|
54
|
+
* - `runtime_bundle_stale` — a control-plane reachability symptom. Re-wrapping
|
|
55
|
+
* would churn every config on this machine and still not fetch the bundle,
|
|
56
|
+
* hiding a network problem behind protection noise.
|
|
57
|
+
*/
|
|
58
|
+
const HEALABLE_REASONS = new Set([
|
|
59
|
+
'mcp_gateway_missing', 'mcp_gateway_stale_path',
|
|
60
|
+
'cursor_hook_missing', 'cursor_hook_stale_path', 'cursor_hook_changed',
|
|
61
|
+
'claude_hook_missing', 'claude_hook_stale_path', 'claude_hook_changed',
|
|
62
|
+
]);
|
|
63
|
+
/** Pure: does this verdict contain anything a protect-all pass could fix? */
|
|
64
|
+
function hasHealableIntegrityReason(reasons) {
|
|
65
|
+
return reasons.some(reason => HEALABLE_REASONS.has(reason));
|
|
66
|
+
}
|
|
45
67
|
function readText(file) {
|
|
46
68
|
try {
|
|
47
69
|
return fs.readFileSync(file, 'utf8');
|
package/dist/version.json
CHANGED
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "fullcourtdefense-cli",
|
|
3
|
-
"version": "1.26.
|
|
3
|
+
"version": "1.26.14",
|
|
4
4
|
"description": "Full Court Defense CLI — security scanning for AI agents from your terminal",
|
|
5
5
|
"main": "dist/index.js",
|
|
6
6
|
"bin": {
|
|
@@ -30,6 +30,7 @@
|
|
|
30
30
|
"test:audit-restore": "node scripts/test-audit-restore.js",
|
|
31
31
|
"test:shell-guard": "npm run build && node scripts/test-shell-guard.js",
|
|
32
32
|
"test:cmd-guard": "npm run build && node scripts/test-cmd-guard.js",
|
|
33
|
+
"test:cmd-guard-perf": "npm run build && node scripts/test-cmd-guard-perf.js",
|
|
33
34
|
"test:posix-guard": "npm run build && node scripts/test-posix-guard.js",
|
|
34
35
|
"test:secret-posture": "npm run build && node scripts/test-secret-posture-fixtures.js",
|
|
35
36
|
"test:agent-file-posture": "npm run build && node scripts/test-agent-file-posture.js",
|