agent-relay-runner 0.91.1 → 0.91.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-relay-runner",
-  "version": "0.91.1",
+  "version": "0.91.2",
   "description": "Unified provider lifecycle runner for Agent Relay",
   "type": "module",
   "bin": {

package/plugins/claude/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-relay-runner",
   "description": "Thin Agent Relay runner bridge for Claude Code",
-  "version": "0.91.1",
+  "version": "0.91.2",
   "agentRelayContracts": {
     "providerPluginProtocol": 1
   }

package/plugins/claude/hooks/stop-failure.sh

@@ -10,15 +10,27 @@ error="$(relay_json_string_field error_type "$payload")"
 [ -z "$error" ] && error="$(relay_json_string_field error "$payload")"
 reset_time="$(relay_json_number_field reset_time "$payload")"
 error_message="$(relay_json_string_field error_message "$payload")"
+[ -z "$error_message" ] && error_message="$(relay_json_string_field error_details "$payload")"
+[ -z "$error_message" ] && error_message="$(relay_json_string_field last_assistant_message "$payload")"
 
 # #286: a usage/rate-limit stall is transient and time-bounded — hold the agent
 # (providerState blocked) and auto-resume at reset, instead of silently going idle.
 # The disambiguator for billing_error is reset_time: a subscription rolling-window
 # limit carries one (and is auto-resumable); true credit exhaustion does not (fatal).
 case "$error" in
-  rate_limit|overloaded)
+  rate_limit|overloaded|server_error|api_error)
     relay_post_rate_limit "$error" "$reset_time" "$error_message"
     ;;
+  unknown)
+    case "$error_message" in
+      *"529"*|*"Overloaded"*|*"overloaded"*|*"temporar"*|*"server-side"*|*"server error"*)
+        relay_post_rate_limit "$error" "$reset_time" "$error_message"
+        ;;
+      *)
+        relay_post_status_clearing_subagents idle
+        ;;
+    esac
+    ;;
   billing_error)
     case "$reset_time" in
       ''|*[!0-9]*) relay_post_status_clearing_subagents error ;;

package/src/adapters/claude.ts

@@ -559,7 +559,7 @@ export function claudeRateLimitStatus(text: string, sessionName?: string): Provi
     status: "idle",
     clear: ["subagent"],
     providerState: buildRateLimitProviderState({
-      errorType: "session_limit",
+      errorType: parsed.errorType ?? "session_limit", kind: parsed.kind ?? "usage_limit",
       ...(parsed.resetAt ? { resetAt: parsed.resetAt } : {}),
       message: parsed.message,
       source: sessionName ? `claude-pane:${sessionName}` : "claude-pane",

package/src/rate-limit.ts

@@ -17,6 +17,7 @@ const MAX_RESET_AHEAD_MS = 7 * 24 * 60 * 60 * 1000;
 
 interface RateLimitHoldInput {
   errorType?: string;
+  kind?: "usage_limit" | "transient_overload";
   /** Unix ms the limit window resets, when known. */
   resetAt?: number;
   message?: string;
@@ -27,12 +28,18 @@ interface RateLimitHoldInput {
 export function buildRateLimitProviderState(input: RateLimitHoldInput): Record<string, unknown> {
   const now = Date.now();
   const resetAt = typeof input.resetAt === "number" && Number.isFinite(input.resetAt) ? input.resetAt : undefined;
-  const label = resetAt ? `usage limit · resets ${formatResetClock(resetAt)}` : "usage limit reached";
+  const kind = input.kind ?? rateLimitKindForError(input.errorType);
+  const label = kind === "transient_overload"
+    ? "provider overloaded"
+    : resetAt ? `usage limit · resets ${formatResetClock(resetAt)}` : "usage limit reached";
   return {
     state: "blocked",
     reason: RATE_LIMIT_BLOCK_REASON,
     label,
-    recommendedAction: "Holding until the usage window resets; agent-relay auto-resumes the agent.",
+    recommendedAction: kind === "transient_overload"
+      ? "Holding briefly for provider overload; agent-relay will retry with bounded backoff."
+      : "Holding until the usage window resets; agent-relay auto-resumes the agent.",
+    kind,
     source: input.source ?? "claude",
     errorType: input.errorType ?? RATE_LIMIT_BLOCK_REASON,
     enteredAt: now,
@@ -42,6 +49,10 @@ export function buildRateLimitProviderState(input: RateLimitHoldInput): Record<s
   };
 }
 
+function rateLimitKindForError(errorType: string | undefined): "usage_limit" | "transient_overload" {
+  return /^(overloaded|server_error|api_error|unknown|529|5\d\d)$/i.test(errorType ?? "") ? "transient_overload" : "usage_limit";
+}
+
 /** Host-local HH:MM for the chat notice / badge label (matches how the CLI shows "resets 3:45pm"). */
 function formatResetClock(resetAtMs: number): string {
   try {
@@ -64,6 +75,9 @@ interface PaneRateLimit {
   message: string;
   /** Unix ms reset, when parseable from the pane (else undefined → resume falls back to poll). */
   resetAt?: number;
+  /** Usage limits wait for reset; transient provider overloads retry quickly. */
+  kind?: "usage_limit" | "transient_overload";
+  errorType?: string;
 }
 
 /**
@@ -75,6 +89,8 @@ interface PaneRateLimit {
  */
 export function parseClaudeRateLimitPane(text: string, nowMs: number = Date.now()): PaneRateLimit | null {
   if (!text) return null;
+  const overload = parseClaudeTransientOverloadPane(text);
+  if (overload) return overload;
   const limit = LIMIT_PHRASE_RE.exec(text);
   if (!limit) return null;
   const message = limitLineAt(text, limit.index);
@@ -95,6 +111,21 @@ export function parseClaudeRateLimitPane(text: string, nowMs: number = Date.now(
   return null;
 }
 
+const TRANSIENT_API_ERROR_RE = /\bAPI Error:\s*((?:5\d\d)|529)\b[^\n]*(?:overload|server-side|temporar|try again|status\.claude\.com)?[^\n]*/i;
+const TRANSIENT_5XX_RE = /\b(?:5(?:00|02|03|04|20|29)|529)\b[^\n]*(?:overload|server-side|temporar|try again|status\.claude\.com)\b[^\n]*/i;
+
+function parseClaudeTransientOverloadPane(text: string): PaneRateLimit | null {
+  const match = TRANSIENT_API_ERROR_RE.exec(text) ?? TRANSIENT_5XX_RE.exec(text);
+  if (!match) return null;
+  const message = limitLineAt(text, match.index);
+  const code = match[1] ?? /\b(5\d\d|529)\b/.exec(match[0])?.[1];
+  return {
+    message: message || match[0].trim(),
+    kind: "transient_overload",
+    errorType: code === "529" ? "overloaded" : "server_error",
+  };
+}
+
 export interface ClaudeRateLimitPaneGateState {
   consecutiveDetections: number;
   reported: boolean;
@@ -116,7 +147,8 @@ export function gatedClaudeRateLimitStatus(
       status: "idle",
       clear: ["subagent"],
       providerState: buildRateLimitProviderState({
-        errorType: "session_limit",
+        errorType: parsed.errorType ?? "session_limit",
+        kind: parsed.kind ?? "usage_limit",
         ...(parsed.resetAt ? { resetAt: parsed.resetAt } : {}),
         message: parsed.message,
         source: options.sessionName ? `claude-pane:${options.sessionName}` : "claude-pane",