The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The expansion_analyst agent in infra/agents/analyst.ts utilizes the webSearch capability to gather information from external websites.
Ingestion points: External web content enters the agent's context through webSearch in infra/agents/analyst.ts.
Boundary markers: The agent is instructed to format its output according to a strict JSON schema and limit reasons to three sentences, which provides structural boundaries, although no specific "ignore embedded instructions" warning is included for the search results.
Capability inventory: While the agent itself is read-only, the workflow defined in infra/plays/flag-expansion.ts writes the agent's output (judgment and reason) to the gtm_accounts database model. The expansion_digest agent in infra/agents/digest.ts subsequently reads these records and posts the content to a Slack channel using slack.actions.postMessage.
Sanitization: There is no evidence of sanitization or filtering of the content retrieved from web searches before it is processed by the agent or written to the database.
[DYNAMIC_EXECUTION]: The evaluation script evals/contract.mjs uses dynamic import() statements to load skill components during validation.
Evidence: The script uses computed paths and query strings (e.g., await import("../infra/plays/flag-expansion.ts?contract=${stamp}")) to bypass module caching and verify the skill's resource registry at runtime.

expansion-signals