fix: reduce Sub2API compact threshold and observe failures

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

@@ -1346,13 +1346,25 @@ def gateway_monitor_paths(config):
     return {"/responses", "/v1/responses", "/responses/compact", "/v1/responses/compact"}
 
 def gateway_failure_kind(message, payload, config):
-    if "openai.forward_failed" not in message or not isinstance(payload, dict):
+    if not isinstance(payload, dict):
         return None
     path = payload.get("path")
     if path not in gateway_monitor_paths(config):
         return None
     if payload.get("account_id") is None:
         return None
+    if "codex.remote_compact.failed" in message:
+        status = payload.get("status_code")
+        if isinstance(status, int) and status >= 500:
+            return "gateway-compact-final-failure"
+        return None
+    if "openai.upstream_failover_switching" in message and path in ("/responses/compact", "/v1/responses/compact"):
+        upstream_status = payload.get("upstream_status")
+        if isinstance(upstream_status, int) and upstream_status >= 500:
+            return "gateway-compact-upstream-failover"
+        return None
+    if "openai.forward_failed" not in message:
+        return None
     error_text = str(payload.get("error") or "").lower()
     fallback_written = payload.get("fallback_error_response_written") is True
     upstream_already_written = payload.get("upstream_error_response_already_written") is True
@@ -1404,7 +1416,7 @@ def gateway_failure_kind(message, payload, config):
     return None
 
 def gateway_failure_is_observe_only(failure_kind):
-    return failure_kind in {"gateway-session-affinity-failure"}
+    return failure_kind in {"gateway-session-affinity-failure", "gateway-compact-final-failure", "gateway-compact-upstream-failover"}
 
 def gateway_failure_item(ts, pod_name, payload, failure_kind):
     request_id = payload.get("request_id") or sha(json.dumps(payload, sort_keys=True, ensure_ascii=False))

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

@@ -202,6 +202,8 @@ interface CodexPoolLocalCodexConfig {
   backupSuffix: string;
   providerName: string;
   wireApi: string;
+  modelContextWindow: number;
+  modelAutoCompactTokenLimit: number;
   supportsWebSockets: boolean;
   responsesWebSocketsV2: boolean;
   responsesSmokeModel: string;
@@ -211,6 +213,8 @@ interface CodexLocalConsumerTomlOptions {
   providerName: string;
   baseUrl: string;
   wireApi: string;
+  modelContextWindow: number;
+  modelAutoCompactTokenLimit: number;
   supportsWebSockets: boolean;
   responsesWebSocketsV2: boolean;
 }
@@ -1009,6 +1013,8 @@ async function codexPoolConfigureLocal(config: UniDeskConfig, options: ConfirmOp
         baseUrl: codexConsumerBaseUrl(pool),
         providerName: pool.localCodex.providerName,
         wireApi: pool.localCodex.wireApi,
+        modelContextWindow: pool.localCodex.modelContextWindow,
+        modelAutoCompactTokenLimit: pool.localCodex.modelAutoCompactTokenLimit,
         supportsWebSockets: pool.localCodex.supportsWebSockets,
         responsesWebSocketsV2: pool.localCodex.responsesWebSocketsV2,
         responsesSmokeModel: pool.localCodex.responsesSmokeModel,
@@ -1274,6 +1280,8 @@ function defaultCodexPoolConfig(): CodexPoolConfig {
       backupSuffix: "pre-sub2api",
       providerName: "OpenAI",
       wireApi: "responses",
+      modelContextWindow: 272000,
+      modelAutoCompactTokenLimit: 240000,
       supportsWebSockets: true,
       responsesWebSocketsV2: true,
       responsesSmokeModel: "gpt-5.5",
@@ -1444,6 +1452,11 @@ function readPositiveInteger(value: unknown, key: string): number {
   return parsed;
 }
 
+function readPositiveIntegerConfig(value: unknown, key: string, fallback: number): number {
+  if (value === undefined || value === null) return fallback;
+  return readPositiveInteger(value, key);
+}
+
 function readCaddyRetryMethods(value: unknown, key: string): string[] {
   if (value === undefined || value === null) return [];
   if (!Array.isArray(value)) throw new Error(`${codexPoolConfigPath}.${key} must be a YAML array`);
@@ -1619,6 +1632,8 @@ function readLocalCodexConfig(value: unknown, defaults: CodexPoolLocalCodexConfi
     backupSuffix: stringValue(value.backupSuffix) ?? defaults.backupSuffix,
     providerName: stringValue(value.providerName) ?? defaults.providerName,
     wireApi: stringValue(value.wireApi) ?? defaults.wireApi,
+    modelContextWindow: readPositiveIntegerConfig(value.modelContextWindow, "localCodex.modelContextWindow", defaults.modelContextWindow),
+    modelAutoCompactTokenLimit: readPositiveIntegerConfig(value.modelAutoCompactTokenLimit, "localCodex.modelAutoCompactTokenLimit", defaults.modelAutoCompactTokenLimit),
     supportsWebSockets: readBooleanConfig(value.supportsWebSockets, "localCodex.supportsWebSockets", defaults.supportsWebSockets),
     responsesWebSocketsV2: readBooleanConfig(value.responsesWebSocketsV2, "localCodex.responsesWebSocketsV2", defaults.responsesWebSocketsV2),
     responsesSmokeModel: stringValue(value.responsesSmokeModel) ?? defaults.responsesSmokeModel,
@@ -3097,6 +3112,8 @@ function writeLocalCodexConfig(pool: CodexPoolConfig, apiKey: string): Record<st
       name: pool.localCodex.providerName,
       baseUrl: codexConsumerBaseUrl(pool),
       wireApi: pool.localCodex.wireApi,
+      modelContextWindow: pool.localCodex.modelContextWindow,
+      modelAutoCompactTokenLimit: pool.localCodex.modelAutoCompactTokenLimit,
       supportsWebSockets: pool.localCodex.supportsWebSockets,
       responsesWebSocketsV2: pool.localCodex.responsesWebSocketsV2,
     },
@@ -3116,6 +3133,8 @@ function updateCodexConfigToml(current: string, pool: CodexPoolConfig): string {
     providerName: pool.localCodex.providerName,
     baseUrl: codexConsumerBaseUrl(pool),
     wireApi: pool.localCodex.wireApi,
+    modelContextWindow: pool.localCodex.modelContextWindow,
+    modelAutoCompactTokenLimit: pool.localCodex.modelAutoCompactTokenLimit,
     supportsWebSockets: pool.localCodex.supportsWebSockets,
     responsesWebSocketsV2: pool.localCodex.responsesWebSocketsV2,
   });
@@ -3127,6 +3146,8 @@ function codexConsumerBaseUrl(pool: CodexPoolConfig): string {
 
 export function renderCodexLocalConsumerToml(current: string, options: CodexLocalConsumerTomlOptions): string {
   let next = upsertTopLevelTomlString(current, "model_provider", options.providerName);
+  next = upsertTopLevelTomlInteger(next, "model_context_window", options.modelContextWindow);
+  next = upsertTopLevelTomlInteger(next, "model_auto_compact_token_limit", options.modelAutoCompactTokenLimit);
   next = upsertProviderSection(next, options);
   next = upsertTomlSectionBoolean(next, "features", "responses_websockets_v2", options.responsesWebSocketsV2);
   return next.endsWith("\n") ? next : `${next}\n`;
@@ -3146,6 +3167,21 @@ function upsertTopLevelTomlString(text: string, key: string, value: string): str
   return lines.join("\n");
 }
 
+function upsertTopLevelTomlInteger(text: string, key: string, value: number): string {
+  const lines = text.split(/\r?\n/u);
+  const firstSection = lines.findIndex((line) => /^\s*\[/.test(line));
+  const end = firstSection === -1 ? lines.length : firstSection;
+  const assignment = `${key} = ${value}`;
+  for (let index = 0; index < end; index += 1) {
+    if (new RegExp(`^\\s*${escapeRegExp(key)}\\s*=`).test(lines[index])) {
+      lines[index] = assignment;
+      return lines.join("\n");
+    }
+  }
+  lines.splice(end, 0, assignment);
+  return lines.join("\n");
+}
+
 function upsertProviderSection(text: string, options: CodexLocalConsumerTomlOptions): string {
   const { providerName, baseUrl, wireApi, supportsWebSockets } = options;
   const header = `[model_providers.${providerName}]`;