The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes terminal scrollback logs and agent traces which may contain untrusted data or instructions from previous agent sessions. Ingestion points: Terminal scrollback data is ingested via herdr pane read in scripts/audit-worker-pane.sh and scripts/read-herdr-panes.sh. Boundary markers: The audit templates in references/forensic-subagent-briefs.md use markdown structure but lack explicit delimiters or instructions to ignore potential commands embedded in the processed logs. Capability inventory: The auditor agent possesses capabilities including run_command, write_to_file, and agent management tools via herdr. Sanitization: The skill reads raw terminal output without sanitizing or escaping content before presenting it to the auditor subagent.
[COMMAND_EXECUTION]: Multiple utility scripts (scripts/audit-worker-pane.sh, scripts/read-herdr-panes.sh, scripts/launch-derailment.sh) utilize shell commands to interact with the herdr CLI for agent lifecycle management and terminal inspection.
[DYNAMIC_EXECUTION]: The scripts/parse-derailment-trace.sh utility executes an embedded Python heredoc to parse and report on agent trace files. Additionally, scripts/launch-derailment.sh executes a user-specified agent command through bash -lc.
[EXTERNAL_DOWNLOADS]: The README.md file contains installation instructions that fetch the skill and associated tools from the author's public GitHub repository (yigitkonur/skills-by-yigitkonur).

audit-skill-by-derailment