The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process large, untrusted datasets from external sources such as email archives and API dumps.\n
Ingestion points: Data is pulled from remote APIs and local file exports (e.g., email takeouts, audio libraries) into the sources/ directory as described in SKILL.md.\n
Boundary markers: The skill explicitly refers to conventions/untrusted-content.md and mandates the use of the untrusted-quoted fence to isolate third-party text, as seen in the Phase 3 (TRIAL) instructions.\n
Capability inventory: The skill possesses the ability to write brain pages to the filesystem and execute shell commands through the gbrain jobs submission system.\n
Sanitization: The agent is instructed to flag imperatives found within datasets and treat all fetched content strictly as data, never as instructions.\n- [COMMAND_EXECUTION]: The skill orchestrates bulk processing via shell commands submitted to a background worker system.\n
Evidence: SKILL.md contains examples of shell command submission: gbrain jobs submit shell --params '{"cmd": "<your pipeline command> ..."}'.\n
Context: The documentation includes safety warnings regarding the risks of shell execution, noting it as an 'RCE-class operator authorization' and explicitly stating that the agent should not enable this capability itself, relying on the user's environment configuration.\n- [DYNAMIC_EXECUTION]: The skill generates and executes shell commands dynamically based on the specific pipeline configuration created at runtime.\n
Evidence: The orchestration logic in Phase 9 (BULK) and Phase 10 (MONITOR) involves constructing command strings to be executed by minions.\n
Context: This is used for paginated fetching and periodic state checkpoints to ensure the ingestion process survives restarts and crashes.

bulk-ingestion