The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a local Python script scripts/init_review_loop.py intended to initialize the review environment. The script uses standard libraries (pathlib, json, argparse) to create directory structures and JSON state files. It does not perform any network operations or execute arbitrary system commands.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external research artifacts, experiment outputs, and drafts, which represents a potential indirect prompt injection surface.
Ingestion points: Research artifacts (PDFs, drafts, logs) and final_issues.json files generated by sibling skills are ingested into the agent's context in SKILL.md and scripts/init_review_loop.py.
Boundary markers: The workflow utilizes a "Claim Ledger" and structured "Issue Tracking" (REVIEW_STATE.json) which provides a schematic boundary for processed data. The instructions explicitly warn the agent not to authorize advancement solely based on internal agreement of artifacts.
Capability inventory: The associated script has file-read/write capabilities restricted to the project directory. The skill instructions reference other tools like research-results-auditor for validation.
Sanitization: The Python script parses input JSON using standard parsers, though no specific natural language sanitization is applied to the artifact text itself.
[DATA_EXPOSURE]: The skill manages internal state via REVIEW_STATE.json. While it resolves paths to source artifacts, it does not exfiltrate data and focuses on maintaining a local audit trail for the research project.

research-review-loop