The Agent Skills Directory

Indirect Prompt Injection: The skill ingests subagent execution logs (transcript.jsonl) to distill learnings. As these logs contain outputs from other agents, they constitute an external data surface where malicious instructions could potentially be embedded to influence the analysis process.
Ingestion points: Execution log files (transcript.jsonl) provided via absolute paths or resolved from harness conventions.
Boundary markers: The instructions do not specify explicit delimiters or guidelines to ignore embedded instructions within the logs being analyzed.
Capability inventory: The skill possesses the capability to execute shell commands (bash) and perform filesystem writes (workspace/learnings.jsonl).
Sanitization: There are no explicit instructions for sanitizing or escaping the content of the log entries before they are parsed or processed.
Dynamic Execution: To optimize token usage when processing large log files, the skill suggests generating short Python scripts or using command-line tools like jq and grep at runtime. Generating and executing scripts based on interpreted task requirements is a common pattern for complex analysis but involves dynamic logic creation.
Command Execution: The skill relies on shell environments to filter and parse trajectory data. Executing shell commands to process external files is a standard approach for this use case, provided the input paths and file contents are handled securely.

mantis-reflect