The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a 'Self-Evolving' design where the agent is instructed to modify the skill's own TRIGGERS list and routing table based on user interactions.
Ingestion points: User intent signals provided during session interactions enter the context to evaluate routing accuracy.
Boundary markers: No explicit delimiters or instructions are provided to the agent to distinguish between valid user intent and adversarial instructions intended to poison the routing table.
Capability inventory: The skill allows Read, Glob, and Grep tools. While the markdown instructions direct the agent to 'Edit this file in-place', the YAML configuration restricts tools to read-only operations, mitigating the risk of persistent instruction poisoning.
Sanitization: No sanitization or validation logic is defined for the updates made to the routing table.
[SAFE]: The skill references local repository landmarks (e.g., findings/evolution/evolution.jsonl, findings/evolution/audits/) to maintain a research ledger. These operations are consistent with the stated purpose of research campaign management and do not involve sensitive system paths or credentials.

crucible-navigator