The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local script scripts/agent-workflow-roast.mjs using the Node.js runtime. This script is responsible for processing session data and generating reports.
[DATA_EXFILTRATION]: The skill accesses sensitive local information, including session JSONL files, memory files, and history files. While the documentation states that the analyzer 'redacts obvious secrets,' it explicitly mentions the ability to read raw secrets if requested for a fresh run. Additionally, the skill includes a workflow to upload generated archives (agent-workflow-roast-site.tgz) to an external 'Sites' service.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from local session history and memory to generate qualitative synthesis and coaching reports. This creates an attack surface where malicious content previously logged in history could influence the agent's coaching output.
Ingestion points: Reads raw session JSONL, memory files, and history files as specified in SKILL.md.
Boundary markers: None identified in the skill instructions to delimit untrusted history data from agent instructions.
Capability inventory: The skill can execute shell commands via node and potentially perform network uploads of archives.
Sanitization: The skill claims to redact 'obvious secrets' before synthesis, but does not specify any sanitization or filtering for instructional content embedded in history.

roast