scheduled-tasks

Scheduled tasks fail because sandboxed sessions lack inherited configuration and have unpredictable filesystem contexts. Hardcoded local paths and missing dependencies cause execution errors when the scheduled Claude task attempts to access repo files.

Reduce your scheduled task to a single shell command that invokes a standalone Python helper. This helper discovers the repo root dynamically at runtime, ensuring reliable execution across fresh sandboxes without hardcoded paths.

Minimize tool approvals by wrapping the scheduled task in a single Bash command that delegates to a Python script. This limits the approval surface primarily to Bash, avoiding complex multi-tool operations inside the sandbox.

Avoid running full pytest or complex multi-tool test workflows inside the sandbox. These operations are brittle in scheduled sessions; use safe checks like import-level verification when needed to prevent CI-style failures.

Hardcoded paths fail across fresh sandboxes and changing session IDs. You must discover the repo dynamically at runtime within your Python helper script so scheduled runs work regardless of the sandbox filesystem context.

The best approach uses a single-command shell wrapper that changes into a dynamically discovered repo path and calls a standalone Python helper. This enforces reliable execution patterns for recurring operations while avoiding sandbox-specific failures.