The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection because its core functionality involves ingesting untrusted data from external sources, such as issue trackers (GitHub, Linear, Jira), emails, and verbal requests, as detailed in the Intake (Phase 1) and Discovery (Phase 2) sections of the change-request journey. The skill mitigates this risk through a rigorous 'evidence-driven' doctrine, which requires the agent to verify all claims against primary artifacts (code, tests, and runtime output) and maintain a mandatory 'evidence ledger' to record assumptions and unverified boundaries.
[DYNAMIC_EXECUTION]: The skill includes a local Python script, eval/run_eval.py, used for validating task fixtures and generating evaluation reports. Analysis of the source code confirms it uses only standard library modules (pathlib, re, argparse, sys) and does not employ dangerous functions such as eval(), exec(), or subprocess to execute arbitrary or untrusted code.
[COMMAND_EXECUTION]: The skill documentation (e.g., references/lifecycle.md) references standard development CLI tools such as git and gh for repository management, commit tracking, and issue snapshotting. These are legitimate, intended functionalities for a software delivery skill and are used within a structured process that enforces human authority gates before sensitive actions like merging or releasing.

neckbeard