The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes live transcripts and webchat turns, which are untrusted external inputs provided at runtime.
Ingestion points: The transcript argument in scripts/context_fetch.py, scripts/vector_query.py, and scripts/transcript_to_llm.py receives external data.
Boundary markers: The skill implements structured prompt assembly in scripts/transcript_to_llm.py using explicit headers such as === TRANSCRIPT TURN ===, === LIVE CONTACT ===, and === HISTORICAL CONTEXT === to isolate untrusted data from system instructions.
Capability inventory: The skill utilizes Bash(psql:*) and the psycopg library for database interactions, and Bash(curl:*) and urllib.request for network requests to the Podium API.
Sanitization: It implements a retrieval-time PII redaction filter in scripts/context_fetch.py that uses regex patterns to mask sensitive information like credit cards, SSNs, emails, and addresses.
[DATA_EXFILTRATION]: The skill performs network operations via urllib.request and curl to fetch live contact information from api.podium.com. This is the intended behavior for the integration, utilizing authentication tokens supplied via environment variables.
[COMMAND_EXECUTION]: The skill executes shell commands using Bash(psql:*) to query the vector store, Bash(python3:*) to run its operator scripts, and Bash(curl:*) to interact with the Podium API.

podium-rag-context-bridge