The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an autonomous cycle that ingests and acts upon instructions from external sources.
Ingestion points: The agent processes posed_to vimarshas (tasks) located in its inbox, as defined in AGENTS.md and the project graph.
Boundary markers: The instructions define structured metadata such as 'Answered when' requirements, but do not specify LLM-level delimiters or sanitization routines for the natural language content of tasks.
Capability inventory: The skill allows the agent to perform git operations (branching, PRs), modify the persistent 'graph' state, and configure network webhooks.
Sanitization: No explicit logic is provided to sanitize or escape the content of incoming tasks before the agent evaluates or executes them.
[COMMAND_EXECUTION]: The skill requires the agent to perform automated repository management using shell commands.
Evidence: The workflow includes explicit instructions for git merge-base, rebasing, and managing pull requests to complete the 'Work Cycle'.
[DATA_EXFILTRATION]: The skill configures network communication channels to maintain its autonomous loop.
Evidence: The skill utilizes iskron_admin(action="add_webhook", ...) to register external URLs for incoming events and iskron_channel for outgoing messages. These functionalities are vendor-specific tools from 'iskron-ai' used for agent coordination.

vahta