Stale OpenClaw bot accounts kept sending hourly Telegram messages after the channel was disabled

The setting was three lines of YAML:

```yaml channels: telegram: enabled: false ```

The author runs OpenClaw, an npm package plus a native Mac app, pinned at [email protected], and wrote none of it ^[1]. They also run Hermes, a deployed copy of Nous Research's MIT-licensed hermes-agent ^[2]. Their diagnosis of the Telegram failure is narrow. "It said nothing about whether anything could still go out. Same boolean, two different claims," the author wrote of the flag ^[8].

Two trigger paths sat behind that one switch. Inbound polling runs because the gateway is listening, and the flag stopped the listening ^[4]. A cron job runs because the clock fired it. The log fits a scheduled job that looked up an account and its token and sent, without consulting the flag ^[4][9]. The post does not include the scheduler's code or the flag's documentation, so which check the job skipped is not established.

The author declines to call it a bug. They wrote that "the flag almost certainly does exactly what it promises for the path it governs" ^[10]. I think that is the right reading of one operator's logs from one version. Their advice is broader: "If you operate anything with both a request-driven path and a schedule-driven path gated by the same nominal switch, don't assume they obey it identically." ^[11]

The token move is the part I would copy. A config flag is a check that each code path has to remember to make. A missing credential fails every path at authentication, including paths the operator forgot existed. "No token, no code path (scheduled or otherwise) can authenticate and send," the author wrote ^[9]. Handing Telegram to Hermes alone ended a setup of "two systems each partially able to trigger sends" ^[7].

The author folds all four incidents in the post into one rule: "a system's report of its own state is a claim, not a fact" ^[15]. In the second, an upgrade from 2026.7.1-2 to 2026.8.1 left twelve stale files that OpenClaw's doctor --fix and --repair commands were supposed to clear ^[12]. The author tried every flag combination the tool exposed, and the files stayed ^[12]. They renamed the files to match the tool's own remediation convention, backed up the originals first, and approved each batch individually from a phone ^[13].

Hermes produced the third. Its persona spec went through at least seven dated revisions, growing from one sentence to 364 lines ^[14]. One example logged as "Bad" has the model suggesting a clarifying question instead of asking one ^[16]. Enforcement then moved out of the prose and into code that runs before the model gets a turn ^[14].