The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes session transcripts, which are considered untrusted data as they contain outputs from AI agents and user inputs.
Ingestion points: The skill reads .jsonl files from local directories such as $HOME/.claude/projects, $HOME/.codex/sessions, and $HOME/.openclaw.
Boundary markers: The skill parses the JSONL records individually but does not implement specific 'ignore' markers within the data; however, the data is strictly rendered in a viewer context.
Capability inventory: The skill has the capability to write files to the local system (fs.writeFile) and spawn a browser process (spawn) to display the result.
Sanitization: The tool implements a comprehensive sanitization strategy. In scripts/html.ts, it uses htmlEscape for text content and scriptJsonEscape for embedded JSON data. The client-side rendering engine uses strict regular expressions for link parsing (restricting protocols to http, https, and file) and prevents active scheme URLs like javascript: to mitigate Cross-Site Scripting (XSS) risks.

openclaw-session-viewer