agent-relay-runner 0.14.0 → 0.15.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-relay-runner",
-  "version": "0.14.0",
+  "version": "0.15.1",
   "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": "0.14.0",
+  "version": "0.15.1",
   "agentRelayContracts": {
     "providerPluginProtocol": 1
   }

package/plugins/claude/hooks/relay-status.sh

@@ -96,7 +96,10 @@ relay_pending_reply_stop_decision() {
   local port="${AGENT_RELAY_RUNNER_PORT:-}"
   [ -z "$port" ] && return 0
   local response
-  response="$(curl -fsS "http://127.0.0.1:${port}/reply-obligations/claude-stop" 2>/dev/null || true)"
+  # --max-time guards the Claude Stop hook's 5s budget: a slow runner/server (e.g. an
+  # un-indexed obligation query) must never block past the timeout, or Claude SIGKILLs
+  # the hook before it clears the turn -> stuck "busy" (#199). On timeout: no block.
+  response="$(curl -fsS --max-time 2 "http://127.0.0.1:${port}/reply-obligations/claude-stop" 2>/dev/null || true)"
   case "$response" in
     *'"decision":"block"'*|*'"decision": "block"'*) ;;
     *) return 0 ;;

package/plugins/claude/hooks/stop.sh

@@ -7,7 +7,9 @@ stop_hook_active="$(relay_json_bool_field stop_hook_active "$payload")"
 if [ "$stop_hook_active" != "true" ]; then
   last_assistant_msg="$(echo "$payload" | jq -c '.last_assistant_message // empty' 2>/dev/null || true)"
   relay_post_session_turn "$(relay_json_string_field transcript_path "$payload")" "$last_assistant_msg"
-  stop_decision="$(relay_pending_reply_stop_decision)"
+  # `|| true`: under `set -e`, a non-zero from the obligation check must never abort
+  # the hook before the idle-clear below — clearing the turn is the critical path (#199).
+  stop_decision="$(relay_pending_reply_stop_decision || true)"
   if [ -n "$stop_decision" ]; then
     printf '%s\n' "$stop_decision"
     exit 0

package/src/runner.ts

@@ -76,12 +76,18 @@ const LOG_TAIL_BYTES = 128 * 1024;
 const PROMPT_ECHO_DEDUP_MS = 30_000;
 // Busy reconciler: a conservative LAST-RESORT backstop for a turn that ended
 // without the provider's Stop hook clearing busy (e.g. ESC straight into the web
-// terminal). It must never fire during a live turn, so it (a) only counts idle
-// after it has actually observed the provider busy, and (b) requires a long,
+// terminal). It must never fire during a live turn, so it requires a long,
 // unbroken idle streak — an active turn shows its working spinner well within
 // this window, which resets the streak. ~32s of uninterrupted idle = really done.
 const BUSY_RECONCILE_POLL_MS = 4_000;
 const BUSY_RECONCILE_IDLE_CONFIRM = 8;
+// When the reconciler never observed the provider busy this turn (a turn faster
+// than the 4s poll — common for short voice/autosend replies), it can't trust a
+// quick idle the way it does after seeing the spinner. But refusing forever wedged
+// fast turns in "busy" when the Stop hook's idle was lost (#199). So we still
+// force-clear, just after a much longer unbroken-idle window — an active turn would
+// have flashed its spinner into at least one of these probes and reset the streak.
+const BUSY_RECONCILE_IDLE_CONFIRM_NO_BUSY = 15;
 // After a dashboard interrupt, give the provider a moment to drop out of its turn,
 // then reconcile immediately so the user sees "stopped" without waiting for the backstop.
 const INTERRUPT_RECONCILE_DELAY_MS = 1_500;
@@ -1146,18 +1152,21 @@ export class AgentRunner {
     let activity: "busy" | "idle" | "unknown";
     try { activity = await this.options.adapter.probeActivity(this.process); } catch { return; }
     if (activity === "busy") this.busyReconcileSawBusy = true;
-    // Reset the streak on anything that isn't a confident idle — and never start
-    // counting until we've actually observed the provider busy this turn.
-    if (activity !== "idle" || !this.busyReconcileSawBusy) {
-      if (activity !== "idle") this.busyReconcileIdleStreak = 0;
+    // Reset the streak on anything that isn't a confident idle.
+    if (activity !== "idle") {
+      this.busyReconcileIdleStreak = 0;
       this.sessionDebug(`reconcile probe=${activity} sawBusy=${this.busyReconcileSawBusy} streak=${this.busyReconcileIdleStreak}`);
       return;
     }
     this.busyReconcileIdleStreak += 1;
-    this.sessionDebug(`reconcile probe=idle streak=${this.busyReconcileIdleStreak}/${BUSY_RECONCILE_IDLE_CONFIRM}`);
-    if (this.busyReconcileIdleStreak < BUSY_RECONCILE_IDLE_CONFIRM) return;
+    // Confirm faster once we've seen the spinner this turn; otherwise demand a much
+    // longer all-idle window before trusting it (rescues fast turns without
+    // false-clearing a live turn that simply hasn't flashed busy into a probe yet).
+    const confirm = this.busyReconcileSawBusy ? BUSY_RECONCILE_IDLE_CONFIRM : BUSY_RECONCILE_IDLE_CONFIRM_NO_BUSY;
+    this.sessionDebug(`reconcile probe=idle sawBusy=${this.busyReconcileSawBusy} streak=${this.busyReconcileIdleStreak}/${confirm}`);
+    if (this.busyReconcileIdleStreak < confirm) return;
     this.disarmBusyReconciler();
-    this.forceClearProviderTurn("backstop reconciler");
+    this.forceClearProviderTurn(this.busyReconcileSawBusy ? "backstop reconciler" : "backstop reconciler (no-busy-observed)");
   }
 
   // Force-clear a stuck provider-turn claim directly. Unlike the idle status path