The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes output from Observe CLI commands which could contain untrusted data from the user's tenant.
Ingestion points: SKILL.md (Phase 1 and 2) reads output from observe auth status, observe content host view, and observe datastream list.
Boundary markers: The skill requires the use of a wrap helper function (defined in references/untrusted-output.md) that uses unique nonces and XML-like tags to delimit untrusted data. The agent is explicitly instructed to ignore any directives found between these tags.
Capability inventory: The skill performs resource creation and installation via observe datastream create, observe content host install, and observe ingest-token create.
Sanitization: The wrap helper uses sed and tr to sanitize ANSI escape sequences and control characters, preventing terminal manipulation attacks (e.g., cursor repositioning) that could hide content from the tags.
[COMMAND_EXECUTION]: The skill uses the observe CLI to provision datastreams, content packages, and ingest tokens. These operations are consistent with the vendor's (observeinc) primary purpose for this skill.
[SAFE]: The skill implements proactive secret management by instructing the user to export the ingest token to a shell environment variable (OBSERVE_TOKEN) and using jq 'del(.secret)' to ensure the secret value is never passed back to the assistant in plain text.

setup-linux-host-backend