The Agent Skills Directory

[COMMAND_EXECUTION]: The orchestrator component of the skill executes the gh command-line tool to create issues on GitHub repositories based on findings generated by the auditor sub-agent. This is documented in Step 3 of the process using the gh issue create command.
[DATA_EXFILTRATION]: The skill accesses sensitive local files, including session transcripts located in ~/.claude/projects/ and commit grant logs in ~/.claude/logs/commit-grants.jsonl. It then transmits summaries of these findings to external GitHub repositories or appends them to local backlog files. Although this is the primary purpose of the skill, it involves the movement of internal session metadata to external destinations.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) because it ingests and processes untrusted session transcript data which could contain instructions designed to influence the behavior of the auditor sub-agent.
Ingestion points: Internal session transcripts from ~/.claude/projects/ and grant logs from ~/.claude/logs/ are ingested into the auditor sub-agent's prompt in Step 2.
Boundary markers: The instructions define a structured rules block for the auditor but do not specify the use of delimiters or escaping mechanisms to isolate the untrusted transcript content from the auditor's system instructions.
Capability inventory: The orchestrator possesses capabilities to execute shell commands (gh), use the Agent tool to launch sub-agents, and perform file-write operations (appending to backlog files).
Sanitization: The process does not include explicit sanitization or validation of the transcript content to prevent embedded instructions from being interpreted as commands by the auditing agent.

session-audit