The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill captures session-specific feedback and persists it in a shared "work graph" (realm=<граф работы>) that is accessible across different sessions and potentially different agents.
Ingestion points: The skill ingests data from the current agent's session memory, including tool outputs, user complaints, and agent observations. It also retrieves data from the shared graph using iskron_semantic_search to check for existing entries.
Boundary markers: No delimiters, markers, or instructions to ignore embedded instructions are specified for the feedback content being written to or read from the shared graph.
Capability inventory: The skill uses vendor-provided tools (iskron_semantic_search, iskron_search, iskron_orient) to interact with the external graph infrastructure.
Sanitization: There is no mechanism described to sanitize or validate session data before it is stored in the shared graph. This allows a malicious user or a compromised session to inject instructions into the graph that could be executed by other agents during retrieval.
[DATA_EXPOSURE]: The skill collects and stores environment-specific metadata, including the "contour" (environment ID), harness details, and timestamps (attrs.posed_by). These operations are facilitated by official vendor tools (iskron_*) and are documented as part of the tool's core functionality for tracking and addressing tool defects.

feedback