The Agent Skills Directory

PROMPT_INJECTION (LOW): The skill is susceptible to indirect prompt injection during the hierarchical summarization process.
Ingestion points: tool/src/parser.ts reads conversation history from .jsonl files stored in ~/.config/superpowers/conversation-archive/.
Boundary markers: The summarizeConversation function in tool/src/summarizer.ts interpolates raw conversation text directly into the LLM prompt without using robust delimiters (e.g., XML tags or randomized separators) or escaping, which could allow a malicious conversation history to influence the summary output.
Capability inventory: The summarization process has network access to the Anthropic API via the @anthropic-ai/claude-agent-sdk and file write access to save the resulting summaries.
Sanitization: No sanitization or filtering is performed on the conversation content before it is interpolated into the prompt.
DATA_EXFILTRATION (SAFE): The skill transmits conversation data to the Anthropic API for summarization. This is the primary intended behavior of the skill and uses the official Claude Agent SDK. Per [TRUST-SCOPE-RULE], calls to established vendor APIs for the skill's core purpose are considered safe.
COMMAND_EXECUTION (SAFE): The skill includes shell scripts (tool/migrate-to-config.sh, tool/test-install-hook.sh) for maintenance and deployment. These scripts use standard system commands (cp, mkdir, sqlite3, chmod) for their stated purposes and do not exhibit malicious intent or privilege escalation.

Remembering Conversations