A user reports that Android charging notifications repeatedly woke their OpenClaw Astra agent, resulting in 127.8M recorded tokens in a single day — though ~98% were cached input tokens, not billed at the full rate. The post seeks advice on debouncing system-notification events or isolating background events to smaller contexts.
Android charging notifications repeatedly woke my OpenClaw Astra agent 127.8M tokens in one day I noticed unexpectedly high usage while I wasn’t actively interacting with OpenClaw. After investigating the local logs, the main trigger appears to have been notification forwarding from my Android phone. Setup OpenClaw 2026.9.7 on CachyOS/Linux Main agent: openai/gpt-6-astra, using the native Codex runtime Samsung Galaxy S25 with notification forwarding enabled Regular heartbeat every 30 minutes A large accumulated main-session history What the logs showed Repeated updates to Android’s charging notification (com.android.systemui, key charging_state) triggered notifications-event wakes. During some periods, these occurred approximately every 30 seconds. The agent frequently responded with no_change or NO_REPLY, but those responses still involved model calls. Many turns included a heartbeat_respond tool call followed by another model response. For October 1, 2026, the recorded usage was: 127,825,724 total tokens 125,030,400 cached input tokens 2,734,313 non-cached input tokens 61,011 output tokens Approximately 87% of the recorded tokens were associated with notification-triggered turns, and another 12% with scheduled heartbeats. One particularly clear 28-minute window contained 56 turns, 112 model responses with usage records, and approximately 14.9M tokens. The notification events repeatedly contained the same charging indicator with updated battery levels or remaining charging time. The totals were cross-checked against OpenClaw transcript records and native usage records, deduplicating response IDs and summing per-response usage rather than cumulative thread counters. Important qualification: roughly 98% of the total was cached input. This is a token-volume measurement, not a claim that 127.8M tokens were billed at the full uncached rate. I have stopped the gateway for now. An existing Python energy-display job now runs through an independent system timer. I haven’t yet completed a controlled before/after usage test. Has anyone else observed this with Android notification forwarding? Is there a recommended way to filter or debounce ongoing system-notification updates before they wake the agent? Should these background events use an isolated, smaller context? I’m sharing this as an observation from my installation; I haven’t established whether it is a configuration issue or a bug.
An OpenClaw agent with a heartbeat was consuming 50 million tokens daily due to a bloated session and a bug that kept it running despite being disabled. The author shares how to identify and fix the issue by clearing the session and configuring heartbeat settings.
This article discusses the inefficiency of polling in OpenClaw agents and introduces a plugin that moves event detection outside the agent loop, resulting in significant reductions in source calls and token usage.
The article critiques current browser AI agents for inefficiency due to repeatedly parsing and reasoning about the same websites, and proposes a model where agents reuse proven interaction paths to reduce token consumption and improve speed.
A user recounts shutting down their OpenClaw AI assistant after struggling with model access restrictions, runaway token costs, and constant bugs, concluding that the hassle outweighs the benefit.
A 13-week recap of using OpenClaw as a daily AI agent on a Raspberry Pi, highlighting strengths like cron-based automation and memory curation, and pain points like model config issues and subagent orchestration.