The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to spawn an interactive shell to execute a specific tool command.
Evidence: In DEEP-DFS.md, Section 6, the agent is directed to use zsh -ic 'codex exec "$(cat /tmp/prompt.txt)" < /dev/null > /tmp/out.md 2>&1' to interact with another model tier.
[INDIRECT_PROMPT_INJECTION]: The skill involves a significant attack surface for indirect prompt injection by consuming untrusted content from the open web to influence its validation and 'kill' logic.
Ingestion points: The skill searches and parses content from Google, GitHub, npm, PyPI, ProductHunt, Stack Overflow, and Reddit as part of its 'kill chain' validation (documented in VALIDATION.md).
Boundary markers: Absent. No specific boundary markers or instructions to ignore embedded commands within the ingested content are mentioned.
Capability inventory: The agent has the ability to write to the local file system (e.g., gap-finder-state.md, /tmp/out.md) and execute shell commands via the codex interface.
Sanitization: Absent. There is no evidence of sanitization, escaping, or filtering of the external data before it is processed by the agent.
[DYNAMIC_EXECUTION]: The skill requires the runtime execution of an external node-based script (codex) that is not distributed with the skill itself, relying on local environment configurations such as nvm and custom shell aliases.
Evidence: Referenced in DEEP-DFS.md under the Codex section as a required component for the 'Deep Mode' adversarial cross-attack.

gap-finder