The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted natural-language feedback, attachments, and system logs provided by users to route bug reports. * Ingestion points: SKILL.md Step 1.1 explicitly instructs the agent to preserve 'quoted feedback, attachments, paths, URLs, logs, screenshots' from the user input. * Boundary markers: While the skill uses quoting to identify feedback, it lacks robust boundary markers or specific instructions to ignore embedded prompts within the preserved data when forwarding to downstream skills. * Capability inventory: The skill acts as a gateway to $task-cycle, $create-task, and $implement-task, which have capabilities to modify the repository, create files, and execute implementation logic. * Sanitization: Input is preserved 'byte-for-byte' without sanitization, increasing the risk that malicious instructions in a bug report could influence the agent's behavior during task resolution or implementation.
[COMMAND_EXECUTION]: The skill performs shell and Version Control System (VCS) operations to inspect the local environment and resolve task identities. * Evidence: SKILL.md Section 1, Item 4: 'Use only read-only shell/VCS inspection here.'

task-bug