vision-manager/manager.sh

@@ -200,19 +200,27 @@ while true; do
     analysis_off=0
   fi
 
-  # Memory-yield gates (before the backlog logic). The active stop handles a
-  # running vision the idle path can't; the inhibit forces pending=0 so the
-  # existing idle-debounce path keeps it down (same shape as the crawl mutex).
+  # Memory-yield: two SEPARATE levers, never conflated.
+  #   - active stop (HIGH watermark): mem_check_and_stop SIGTERMs a RUNNING
+  #     vision when the host is critically short — the ONLY thing that stops a
+  #     vision that is actively working.
+  #   - start block (LOW watermark / cooldown): holds off STARTING a stopped
+  #     vision when memory is already tight, via the mem_block_start flag.
+  #     It must NOT zero `pending`: doing so made the idle-debounce path read
+  #     "no work" and idle-stop a vision mid-analysis once the box floated past
+  #     LOW (idle stack + LLM ≈ 80% here). LOW is a START gate, not a stop
+  #     threshold.
   mem_check_and_stop
   running="$(vision_running)"   # refresh: the active stop may have just flipped it
+  mem_block_start=0
   if [ "$(mem_yield_active)" = "1" ]; then
+    mem_block_start=1
     if [ "$mem_inhibited" != "1" ]; then
-      log "memory-yield: inhibiting vision start (used >= ${MEM_LOW_WATERMARK_PCT}% or cooldown active)"
+      log "memory-yield: blocking vision START (used >= ${MEM_LOW_WATERMARK_PCT}% or cooldown); a running vision keeps working until the ${MEM_HIGH_WATERMARK_PCT}% active stop"
       mem_inhibited=1
     fi
-    pending=0
   elif [ "$mem_inhibited" = "1" ]; then
-    log "memory-yield: pressure cleared — start inhibit lifted"
+    log "memory-yield: pressure cleared — start block lifted"
     mem_inhibited=0
   fi
 
@@ -225,7 +233,12 @@ while true; do
   if [ "$pending" -gt 0 ]; then
     idle_for=0
     if [ "$running" != "true" ]; then
-      if [ "$RESPECT_CRAWL_MUTEX" != "0" ] && [ "$(crawl_running)" != "0" ]; then
+      if [ "$mem_block_start" = "1" ]; then
+        # Memory pressure / cooldown — hold off STARTING (flip already logged).
+        # A RUNNING vision never reaches here; idle_for stayed 0 above because
+        # pending>0, so a working vision is never idle-stopped under pressure.
+        :
+      elif [ "$RESPECT_CRAWL_MUTEX" != "0" ] && [ "$(crawl_running)" != "0" ]; then
         # Log once per deferral episode, not every poll, so a long crawl
         # doesn't flood the log.
         if [ "$crawl_deferred" != "1" ]; then