The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/model-error-wrapper.sh executes arbitrary shell commands provided via the --command argument using bash -c "$COMMAND". This is a core feature of the utility but allows for the execution of any shell-compatible code passed to the wrapper.
[COMMAND_EXECUTION]: The skill relies extensively on executing external system commands and scripts, including openclaw, notify-send, and custom scripts located in ~/.openclaw/scripts/ to manage model status and alerts.
[INDIRECT_PROMPT_INJECTION]: The skill captures error messages from model API responses or command outputs and passes them to scripts/auto-switch-handler.sh. These externally sourced messages are logged and processed to determine switching logic.
Ingestion points: Standard error and output streams of the model commands are captured in model-error-wrapper.sh and passed to the handler via the --error-message flag.
Boundary markers: No explicit delimiters or sanitization routines are used when processing the captured error strings.
Capability inventory: The skill has access to shell execution (bash -c), file system writes for logging, and potentially network access if the user implements the suggested Telegram alert scripts.
Sanitization: The error messages are interpolated into log commands and passed between scripts without escaping, which could lead to issues if an attacker can control the error output of a failed model request.

model-fallback