The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary function involves executing the opencode command-line tool via the scripts/start-voice-session.mjs script. It uses this to process voice turns and generate agent responses. The script uses spawn for safe argument handling and includes logic to verify safety flags when running in hosted mode, failing closed if configured boundaries are unsupported by the installed CLI.
[DATA_EXFILTRATION]: In its default hosted mode, the skill establishes an outbound WebSocket connection to the vendor's service domain voice.usepipa.com to relay text between the user's browser and the local bridge. This behavior is the intended primary purpose of the skill and is documented with associated privacy and retention policies in references/privacy-and-retention.md. The relay implementation explicitly redacts tokens and message bodies from its own logs to prevent accidental data exposure.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) because it ingests untrusted speech data from a browser and interpolates it into prompts for the agent.
Ingestion points: scripts/start-voice-session.mjs receives user speech text via a local API (/api/turn) or a remote WebSocket relay.
Boundary markers: The skill uses clear delimiters in its prompt templates (e.g., User said: ${message}) and provides explicit instructions to the agent to ignore launch, setup, or session-management details if they appear in the conversation context.
Capability inventory: The bridge script can execute opencode run and initiate network connections to the hosted relay.
Sanitization: The skill performs basic string trimming on user input before processing.

pipa-huddle-beta