The Agent Skills Directory

[OBFUSCATION]: The skill instructs users to deliberately evade platform security observability (specifically redaction filters) by splitting sensitive environment variable names into multiple string fragments. This technique is designed to prevent the platform from identifying and redacting secrets in generated scripts and log outputs.
[PROMPT_INJECTION]: The skill documentation mandates the use of the --ignore-rules flag when invoking headless reviewer instances. This flag disables system-level safety instructions and project rules to achieve 'lens independence,' effectively bypassing platform safety guardrails.
[DYNAMIC_EXECUTION]: The orchestration logic relies on runtime generation of shell (.sh) and Python (.py) scripts. These scripts are written to temporary files and executed as background processes to manage parallel model calls, which increases the execution complexity and attack surface.
[REMOTE_CODE_EXECUTION]: The skill performs remote model invocations using the hermes -z utility and direct Python urllib requests to external providers (e.g., OpenRouter, OmniRoute). When combined with the safety bypass flags, this allows for unconstrained model behavior.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface by inlining untrusted artifact text (such as code or external plans) directly into reviewer prompts across multiple model instances.
Ingestion points: Artifact text is ingested via the SKILL.md synthesis workflow and references/delegate-task-panel-review.md context inlining.
Boundary markers: The templates use text-based delimiters like 'Review target:' but do not implement robust structural isolation for the untrusted content.
Capability inventory: The skill utilizes terminal, write_file, and background process execution.
Sanitization: There is no evidence of automated sanitization or escaping of the processed artifacts to prevent prompt injection attacks from malicious inputs.

multi-review