fix: use backoff for codex gateway stream freezes

config/platform-infra/sub2api-codex-pool.yaml

@@ -183,7 +183,9 @@ sentinel:
     enabled: true
     lookbackSeconds: 900
     tailLines: 4000
-    freezeTtlMinutes: 10
+    initialTtlMinutes: 5
+    maxTtlMinutes: 30
+    backoffMultiplier: 2
     paths:
       - /responses
       - /v1/responses

scripts/src/platform-infra-sub2api-codex-sentinel.ts

@@ -30,7 +30,9 @@ export interface CodexPoolSentinelConfig {
     enabled: boolean;
     lookbackSeconds: number;
     tailLines: number;
-    freezeTtlMinutes: number;
+    initialTtlMinutes: number;
+    maxTtlMinutes: number;
+    backoffMultiplier: number;
     paths: string[];
   };
   sdk: {
@@ -111,7 +113,9 @@ export function defaultCodexPoolSentinelConfig(): CodexPoolSentinelConfig {
       enabled: false,
       lookbackSeconds: 900,
       tailLines: 4000,
-      freezeTtlMinutes: 10,
+      initialTtlMinutes: 5,
+      maxTtlMinutes: 30,
+      backoffMultiplier: 2,
       paths: ["/responses", "/v1/responses"],
     },
     sdk: {
@@ -198,7 +202,9 @@ export function readCodexPoolSentinelConfig(value: unknown, defaults: CodexPoolS
       enabled: readBoolean(valueAt(gatewayFailureMonitor, "enabled"), `${sourcePath}.sentinel.gatewayFailureMonitor.enabled`, defaults.gatewayFailureMonitor.enabled),
       lookbackSeconds: readInt(valueAt(gatewayFailureMonitor, "lookbackSeconds"), `${sourcePath}.sentinel.gatewayFailureMonitor.lookbackSeconds`, defaults.gatewayFailureMonitor.lookbackSeconds, 60, 7200),
       tailLines: readInt(valueAt(gatewayFailureMonitor, "tailLines"), `${sourcePath}.sentinel.gatewayFailureMonitor.tailLines`, defaults.gatewayFailureMonitor.tailLines, 100, 50000),
-      freezeTtlMinutes: readInt(valueAt(gatewayFailureMonitor, "freezeTtlMinutes"), `${sourcePath}.sentinel.gatewayFailureMonitor.freezeTtlMinutes`, defaults.gatewayFailureMonitor.freezeTtlMinutes, 1, 1440),
+      initialTtlMinutes: readInt(valueAt(gatewayFailureMonitor, "initialTtlMinutes"), `${sourcePath}.sentinel.gatewayFailureMonitor.initialTtlMinutes`, defaults.gatewayFailureMonitor.initialTtlMinutes, 1, 1440),
+      maxTtlMinutes: readInt(valueAt(gatewayFailureMonitor, "maxTtlMinutes"), `${sourcePath}.sentinel.gatewayFailureMonitor.maxTtlMinutes`, defaults.gatewayFailureMonitor.maxTtlMinutes, 1, 1440),
+      backoffMultiplier: readInt(valueAt(gatewayFailureMonitor, "backoffMultiplier"), `${sourcePath}.sentinel.gatewayFailureMonitor.backoffMultiplier`, defaults.gatewayFailureMonitor.backoffMultiplier, 1, 10),
       paths: readPathList(valueAt(gatewayFailureMonitor, "paths"), `${sourcePath}.sentinel.gatewayFailureMonitor.paths`, defaults.gatewayFailureMonitor.paths),
     },
     sdk: {
@@ -239,6 +245,9 @@ export function readCodexPoolSentinelConfig(value: unknown, defaults: CodexPoolS
   if (config.freeze.maxTtlMinutes < config.freeze.initialTtlMinutes) {
     throw new Error(`${sourcePath}.sentinel.freeze.maxTtlMinutes must be >= initialTtlMinutes`);
   }
+  if (config.gatewayFailureMonitor.maxTtlMinutes < config.gatewayFailureMonitor.initialTtlMinutes) {
+    throw new Error(`${sourcePath}.sentinel.gatewayFailureMonitor.maxTtlMinutes must be >= initialTtlMinutes`);
+  }
   if (!/^[-0-9A-Za-z_/*,\s]+$/u.test(config.schedule)) {
     throw new Error(`${sourcePath}.sentinel.schedule has an unsupported cron format`);
   }
@@ -1383,11 +1392,23 @@ def gateway_failure_account_map(admin):
         by_id[account_id] = name
     return by_id
 
+def next_gateway_freeze_interval(account_state, config):
+    monitor_cfg = config.get("gatewayFailureMonitor") if isinstance(config.get("gatewayFailureMonitor"), dict) else {}
+    initial = int(monitor_cfg.get("initialTtlMinutes") or 5)
+    maximum = int(monitor_cfg.get("maxTtlMinutes") or 30)
+    multiplier = int(monitor_cfg.get("backoffMultiplier") or 2)
+    previous = int(account_state.get("gatewayFailureBackoffIntervalMinutes") or 0)
+    quarantine = account_state.get("quarantine") if isinstance(account_state.get("quarantine"), dict) else {}
+    if quarantine.get("reason") == "gateway-forward-failure":
+        previous = max(previous, int(quarantine.get("intervalMinutes") or 0))
+    if previous <= 0:
+        return initial
+    return min(maximum, max(initial, previous * multiplier))
+
 def apply_gateway_failure(account_name, failures, state, config, now, admin, profile):
     latest = failures[-1]
     account_state = state.setdefault("accounts", {}).setdefault(account_name, {})
-    monitor_cfg = config.get("gatewayFailureMonitor") if isinstance(config.get("gatewayFailureMonitor"), dict) else {}
-    interval = int(monitor_cfg.get("freezeTtlMinutes") or config["freeze"]["maxTtlMinutes"])
+    interval = next_gateway_freeze_interval(account_state, config)
     until = add_minutes(now, interval, int(config["freeze"]["jitterPercent"]))
     actions_enabled = bool(config["actions"]["enabled"])
     applied = False
@@ -1426,6 +1447,7 @@ def apply_gateway_failure(account_name, failures, state, config, now, admin, pro
     account_state["lastStatus"] = "quarantined"
     account_state["lastFailureAt"] = iso(now)
     account_state["lastGatewayFailureAt"] = iso(now)
+    account_state["gatewayFailureBackoffIntervalMinutes"] = interval
     account_state["lastGatewayFailure"] = {
         "accountName": account_name,
         "accountId": latest.get("accountId"),
@@ -1440,6 +1462,7 @@ def apply_gateway_failure(account_name, failures, state, config, now, admin, pro
         "countInRun": len(failures),
         "firstAt": failures[0].get("at"),
         "lastAt": latest.get("at"),
+        "intervalMinutes": interval,
         "freezeUntil": iso(until),
         "action": action,
     }