bash-script-generator

Shell script generation

Before repository work, read and apply Mandatory agent tools in the repository root at docs/agent-tools.md: Caveman for every task, CocoIndex Code before broad code navigation, and Crawl4AI for web-page research. Retain actual execution evidence.

Determine the intended interpreter, inputs, outputs, failure behavior and target environment from the request and existing project. Default to Bash only when no different shell is required. Use the project's conventions and dependencies; do not promise POSIX sh portability for Bash-specific syntax or a Bash scaffold. For a POSIX sh target, write for that shell and check the actual supported implementations; a syntax check alone does not establish portability.

Implement the requested behavior with appropriate argument handling and meaningful tests. Validate the generated script itself, using the intended interpreter's syntax check and, when available, ShellCheck configured for that interpreter. Run representative behavior and failure cases in an isolated test workspace. Report commands, results, unavailable checks and remaining limits. A syntax check does not run the script or establish functional correctness.

Do not assume set -e handles every failure. In Bash, a function called as an if condition can continue after a failing command even when errexit is enabled; explicitly propagate failures that matter. With Bash pipefail, check the pipeline result and do not interpret it as identifying the command that finished last. Test the actual control flow and shell versions used by the script.

The upstream scripts/run_ci_checks.sh checks the skill's bundled files and its own regression tests; it does not accept the generated script as an input. Its passing result is not validation of the user's output. The upstream standard template uses Bash features and is not a POSIX sh scaffold.

The original upstream instructions are preserved at provenance/original.SKILL.md. This overlay addresses selected interpreter and validation issues; its component controls do not measure coding-agent effectiveness.