The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The uv.lock file and pyproject.toml reference multiple Python packages with version numbers that do not currently exist and upload timestamps set in the future (2026). Examples include certifi==2026.4.22, playwright==1.63.0, requests==2.34.0, and ruff==0.16.2. This technique is commonly used in dependency confusion or supply chain attacks to preemptively override legitimate packages with malicious versions from public registries.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill ingests raw MTG deck lists from various external sources and user-supplied power bracket labels as primary inputs for its analysis pipeline.
Boundary markers: It employs a secondary 'Rules Audit' subagent to check the final output against official rules, which provides a consistency check but does not mitigate the initial injection risk during parsing.
Capability inventory: The skill utilizes numerous subprocess calls via uv run to execute local CLI tools, performs filesystem writes, and initiates network requests.
Sanitization: There is no evidence of explicit sanitization or filtering applied to the raw deck list strings before they are processed by the underlying tools.
[DATA_EXFILTRATION]: The arena-meta tool uses browser automation (playwright) to fetch metadata from Untapped.gg. The instructions explicitly direct the user to 'sign in to Untapped once' using a browser profile. Browser automation scripts operating on authenticated sessions can be leveraged to exfiltrate session cookies, credentials, or user data if the automation logic is malicious or compromised.
[EXTERNAL_DOWNLOADS]: The skill performs remote downloads of card databases and rulebooks. While the intended targets are well-known MTG data providers, the presence of anomalous dependency versions in the lock file significantly increases the risk that these download mechanisms could be redirected or intercepted.

deck-strat