The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses local session logs located in ~/.claude/projects/ to extract evidence of friction or failures. These logs contain a detailed history of user interactions.
Evidence: reference.md provide explicit instructions for searching .jsonl session files in the user's project directory using grep to find past session friction.
Mitigation: The skill enforces a strict redaction policy defined in rules/evidence-no-customer-data.md that mandates the removal of hostnames, IP addresses, filesystem paths, and organization names before any data is drafted. It also instructs the agent to abort the process if sensitive secrets like tokens or JWTs are detected.
[COMMAND_EXECUTION]: Uses Bash to execute shell commands for log discovery and system interrogation.
Evidence: The skill uses grep for transcript discovery and infrahubctl info (or pip show infrahub-sdk) to retrieve version information for the Infrahub SDK and server for inclusion in reports.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from previous session logs and the current conversation to generate a draft report.
Ingestion points: Session logs (~/.claude/projects/*/*.jsonl) and the current conversation transcript.
Boundary markers: None explicitly defined during the ingestion of raw log files.
Capability inventory: Bash, Grep, WebFetch, Read, Glob.
Sanitization: Implements a comprehensive redaction protocol (rules/evidence-no-customer-data.md) that requires paraphrasing problem structures generically and includes a hard stop for high-entropy secrets. All drafts are handed to a human-in-the-loop review skill before submission.

infrahub-reporting-skill-gaps