The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructions include commands to read from a local launch.env file in the worktree using grep. While environment files are classified as sensitive, the skill performs a targeted extraction of a non-sensitive port number (HIVE_DESKTOP_HTTP_PORT) for local loopback communication.
[COMMAND_EXECUTION]: The skill uses local shell commands including curl, jq, grep, and cut to interact with the desktop application's loopback API and configuration files. These operations are restricted to the local environment and are intended for developer observation and testing.
[INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes data from the local desktop application's inbox and event logs, which may contain content from external sources such as GitHub pull requests.
Ingestion points: Tools such as list_inbox, list_inbox_item_events, and get_flow (referenced in SKILL.md) ingest external payload data into the agent's context.
Boundary markers: The skill uses structured JSON responses and jq for filtering, but does not explicitly document natural language boundary markers or instructions to ignore embedded commands.
Capability inventory: The skill possesses capabilities for network operations via curl to 127.0.0.1 and local file reading via grep on launch.env.
Sanitization: External content is processed through jq and intended for developer review or simulation within the application's context.
[DYNAMIC_EXECUTION]: The execute_flow tool allows for the runtime execution of flow logic provided as YAML or JSON objects (flowYaml or flow). The skill documentation specifies that this tool operates as a stateless dry-run that does not spawn external processes or modify the underlying data store, serving as a simulator for development purposes.

desktop-api