Deliver scheduled article digests to Telegram with OpenClaw cron jobs
After deploying OpenClaw and documenting 16 incidents in one day, I kept getting questions about specific errors. Task: The author deployed OpenClaw to deliver scheduled article digests to Telegram, encountering 16 incidents in one day. The workflow involved configuring cron agents to scan RSS feeds and announce agents to forward summaries. Key fixes included disabling 'thinking' mode for cron jobs to prevent empty outputs and using isolated sessions to stop DeepSeek hallucinations about phantom files. The author manually diagnosed logs to identify that 'delivered: true' status did not guarantee message arrival.; tools: OpenClaw. Human role: The original post does not say how a person stayed involved.; limitations: Small models like Mistral Small or Phi-3 may fail to produce useful announce output from long summaries. The announce agent model is hardcoded to the primary default and cannot be configured independently. 'Delivered: true' status only indicates process completion, not actual message arrival, requiring manual verification in Telegram..
Tools
Published on Agent Work Atlas:
Overview
After deploying OpenClaw and documenting 16 incidents in one day, I kept getting questions about specific errors.
Workflow
The author deployed OpenClaw to deliver scheduled article digests to Telegram, encountering 16 incidents in one day. The workflow involved configuring cron agents to scan RSS feeds and announce agents to forward summaries. Key fixes included disabling 'thinking' mode for cron jobs to prevent empty outputs and using isolated sessions to stop DeepSeek hallucinations about phantom files. The author manually diagnosed logs to identify that 'delivered: true' status did not guarantee message arrival.
Reported result
Disabling thinking mode reduced cron job execution time from an average of 88 seconds to 30 seconds and enabled successful delivery of six articles on the first attempt. The initial 24-hour deployment cost $1.50, largely due to retry loops caused by using a small model for complex tasks. The author did not report long-term stability metrics beyond these immediate fixes.
Human role
The original post does not say how a person stayed involved.
Limitations
Small models like Mistral Small or Phi-3 may fail to produce useful announce output from long summaries. The announce agent model is hardcoded to the primary default and cannot be configured independently. 'Delivered: true' status only indicates process completion, not actual message arrival, requiring manual verification in Telegram.
Try it
I need to configure OpenClaw to send daily article digests to Telegram. First, set up a cron agent to scan [replaceable RSS feed URLs] and generate summaries. Ensure the config sets "thinking": "off" for the cron agent to prevent empty outputs. Use a large model like DeepSeek Chat for heartbeat/cron tasks and a smaller model for chat if needed. Configure the announce agent to forward these summaries to [replaceable Telegram Chat ID]. Before enabling the schedule, run a single test job manually. Check the logs for execution time and verify the message appears in Telegram. If the status says 'delivered: true' but no message arrives, inspect the summary field in DEBUG logs for empty content. Do not publish the schedule until I confirm the test message is correct and free of control tokens like <｜tool▁calls▁begin｜>. Report the cost and execution time of the test run.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.