fix: idle detection + exponential backoff for dev loop (#435)

Co-authored-by: Kimi Agent <kimi@timmy.local>
Co-committed-by: Kimi Agent <kimi@timmy.local>

scripts/cycle_retro.py

@@ -149,6 +149,11 @@ def update_summary() -> None:
 def main() -> None:
     args = parse_args()
 
+    # Reject idle cycles — no issue and no duration means nothing happened
+    if not args.issue and args.duration == 0:
+        print(f"[retro] Cycle {args.cycle} skipped — idle (no issue, no duration)")
+        return
+
     # A cycle is only truly successful if hermes exited clean AND main is green
     truly_success = args.success and args.main_green
 

scripts/loop_guard.py

@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+"""Loop guard — idle detection + exponential backoff for the dev loop.
+
+Checks .loop/queue.json for ready items before spawning hermes.
+When the queue is empty, applies exponential backoff (60s → 600s max)
+instead of burning empty cycles every 3 seconds.
+
+Usage (called by the dev loop before each cycle):
+  python3 scripts/loop_guard.py          # exits 0 if ready, 1 if idle
+  python3 scripts/loop_guard.py --wait   # same, but sleeps the backoff first
+  python3 scripts/loop_guard.py --status # print current idle state
+
+Exit codes:
+  0 — queue has work, proceed with cycle
+  1 — queue empty, idle backoff applied (skip cycle)
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+import time
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+QUEUE_FILE = REPO_ROOT / ".loop" / "queue.json"
+IDLE_STATE_FILE = REPO_ROOT / ".loop" / "idle_state.json"
+
+# Backoff sequence: 60s, 120s, 240s, 600s max
+BACKOFF_BASE = 60
+BACKOFF_MAX = 600
+BACKOFF_MULTIPLIER = 2
+
+
+def load_queue() -> list[dict]:
+    """Load queue.json and return ready items."""
+    if not QUEUE_FILE.exists():
+        return []
+    try:
+        data = json.loads(QUEUE_FILE.read_text())
+        if isinstance(data, list):
+            return [item for item in data if item.get("ready")]
+        return []
+    except (json.JSONDecodeError, OSError):
+        return []
+
+
+def load_idle_state() -> dict:
+    """Load persistent idle state."""
+    if not IDLE_STATE_FILE.exists():
+        return {"consecutive_idle": 0, "last_idle_at": 0}
+    try:
+        return json.loads(IDLE_STATE_FILE.read_text())
+    except (json.JSONDecodeError, OSError):
+        return {"consecutive_idle": 0, "last_idle_at": 0}
+
+
+def save_idle_state(state: dict) -> None:
+    """Persist idle state."""
+    IDLE_STATE_FILE.parent.mkdir(parents=True, exist_ok=True)
+    IDLE_STATE_FILE.write_text(json.dumps(state, indent=2) + "\n")
+
+
+def compute_backoff(consecutive_idle: int) -> int:
+    """Exponential backoff: 60, 120, 240, 600 (capped)."""
+    return min(BACKOFF_BASE * (BACKOFF_MULTIPLIER ** consecutive_idle), BACKOFF_MAX)
+
+
+def main() -> int:
+    wait_mode = "--wait" in sys.argv
+    status_mode = "--status" in sys.argv
+
+    state = load_idle_state()
+
+    if status_mode:
+        ready = load_queue()
+        backoff = compute_backoff(state["consecutive_idle"])
+        print(json.dumps({
+            "queue_ready": len(ready),
+            "consecutive_idle": state["consecutive_idle"],
+            "next_backoff_seconds": backoff if not ready else 0,
+        }, indent=2))
+        return 0
+
+    ready = load_queue()
+
+    if ready:
+        # Queue has work — reset idle state, proceed
+        if state["consecutive_idle"] > 0:
+            print(f"[loop-guard] Queue active ({len(ready)} ready) — "
+                  f"resuming after {state['consecutive_idle']} idle cycles")
+        state["consecutive_idle"] = 0
+        state["last_idle_at"] = 0
+        save_idle_state(state)
+        return 0
+
+    # Queue empty — apply backoff
+    backoff = compute_backoff(state["consecutive_idle"])
+    state["consecutive_idle"] += 1
+    state["last_idle_at"] = time.time()
+    save_idle_state(state)
+
+    print(f"[loop-guard] Queue empty — idle #{state['consecutive_idle']}, "
+          f"backoff {backoff}s")
+
+    if wait_mode:
+        time.sleep(backoff)
+
+    return 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())