The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data when using the web evidence type, which is a standard attack surface for indirect prompt injection.
Ingestion points: Web search results and external web page content.
Boundary markers: The skill acts as a constraint by requiring specific terminal evidence markers, though it does not define explicit delimiters for the external data itself.
Capability inventory: The agent utilizes file system access (observed), log access (tool-record), and web browsing (web).
Sanitization: The skill mitigates risks by requiring verbatim trích dẫn (quotes), direct links, and clear identification of source types, which allows for human verification of the agent's claims.
[COMMAND_EXECUTION]: The skill references the use of commands and their outputs as evidence (observed type).
Mentions providing "lệnh chạy lại được kèm output" (runnable commands with output) as a source of truth.
This directs the agent to utilize its existing command execution capabilities in a transparent and auditable manner rather than introducing new, hidden, or dangerous command execution vectors.
[SAFE]: The skill is entirely instructional and does not contain scripts, binaries, or external code dependencies.
It references an external validator script evidence_terminal.py and a policy file harness/policy.yaml, which are part of the platform's existing infrastructure rather than skill-supplied components.
The logic is designed to improve the reliability of the agent's reasoning for high-stakes tasks like auditing or bug diagnosis.

graph-mode