The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses historical agent session logs located at ~/.pi/agent/sessions/**/*.jsonl. These logs contain a record of all past interactions, which may include sensitive user data, private information, or credentials. The skill extracts this data and sends it to external backend providers during the replay phase.\n- [DYNAMIC_EXECUTION]: The workflow involves generating a TypeScript file (replay.mts) from a provided template and executing it at runtime. This dynamically created script can perform arbitrary file system and network operations.\n- [COMMAND_EXECUTION]: The replay runner template uses Node.js execFileSync to call external CLI tools. This allows for the execution of arbitrary system commands defined during the script's creation to simulate backend behavior.\n- [EXTERNAL_DOWNLOADS]: The skill's environment check and execution steps rely on npx to run the tsx package. If the package is not already cached, it is downloaded from the npm registry at runtime.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests and replays historical user queries. If those queries contained malicious instructions or were designed to trigger specific agent behaviors, they would be re-processed by the tools under test.\n
Ingestion points: Historical logs at ~/.pi/agent/sessions/**/*.jsonl processed by scripts/extract-queries.py.\n
Boundary markers: None; extracted queries are passed directly to tools without delimiters or safety warnings.\n
Capability inventory: execFileSync, fetch, and writeFileSync (found in scripts/replay-runner.template.mts).\n
Sanitization: Limited to a basic regex filter for short or specific 'junk' strings; the content of the queries is not sanitized or validated.

replay-compare