The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where an agent records its decisions in a TSV log and later audits that log. Because the log entries (specifically the 'evidence' and 'result' fields) may contain data from untrusted external sources like PR titles, file contents, or commit messages, there is a risk that malicious instructions could be embedded in the log. During the 'Audit' or 'Cross-model review' phases, the agent reading these logs could be influenced by such instructions.
Ingestion points: scripts/log.sh accepts arguments that can include external data; SKILL.md instructs the agent to read the resulting TSV log.
Boundary markers: The skill uses a structured TSV format, but lacks explicit prompt-level delimiters within the TSV cells to warn the agent about untrusted content.
Capability inventory: The skill utilizes t3_thread_read to ingest history and delegate_task to spawn review processes.
Sanitization: The scripts/log.sh script includes robust sanitization that strips control characters (tabs, newlines) to preserve TSV structure and prefixes formula-triggering characters (=, +, -, @) with a single quote to prevent injection when the log is viewed in spreadsheet software.

show-me-your-work