The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local shell command to update the skill index after modifications: uv run --directory tools/skill-router python indexer.py.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and parses untrusted subagent logs (*.jsonl) that contain raw tool outputs and reasoning steps from previous engagements. These logs could potentially contain malicious payloads or instructions from a compromised target.
Ingestion points: engagement/evidence/logs/*.jsonl, engagement/scope.md, and engagement/state.db.
Capability inventory: The skill has the ability to write to the skills/ and agents/ directories and execute shell commands via uv run.
Boundary markers: The instructions do not define specific delimiters or "ignore" instructions for the subagent tasked with parsing the logs, relying on the subagent's general-purpose prompt.
Sanitization: While manual "CTF sanitization" is mentioned for the final output, there is no automated sanitization of the log data during the parsing phase.
[DYNAMIC_EXECUTION]: The skill automates the creation and modification of other skill definitions (SKILL.md) and agent templates based on analyzed engagement data, effectively modifying the agent's logic for future sessions.

retrospective