The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute a local script (audit.ts or audit.py) to process transcript data. This is the primary function of the skill and is triggered by the user via the /harness-doctor command.
[DATA_EXPOSURE]: The tool reads session transcripts from local storage, which may contain sensitive information or credentials. To mitigate risk, the skill implements a redaction engine (scripts/lib/redact.ts and scripts/lib/redact.py) that filters out common secret patterns (API keys, SSH keys, etc.) by default. A --no-redact flag exists to disable this behavior, but the skill instructions warn to use it only if explicitly requested by the user.
[INDIRECT_PROMPT_INJECTION]: The skill processes historical transcript data which may contain untrusted content from past sessions. There is a theoretical risk that malicious content within past transcripts could influence the agent's current analysis.
Ingestion points: scripts/lib/scan.ts and scripts/lib/scan.py read JSONL transcript files from the user's local projects root.
Boundary markers: The agent is instructed to recompute findings from generated fact tables rather than relying on inferred text, and it is told to verify every number before reporting.
Capability inventory: The agent can execute the audit scripts and, with explicit user permission, modify project configuration files such as CLAUDE.md, settings.json, and package.json.
Sanitization: A best-effort redaction mechanism is used when writing diagnostic fact tables to disk, unless the user intentionally disables it.

harness-doctor