learning-capture
Pass
Audited by Gen Agent Trust Hub on Apr 23, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill installs a bash script at ~/.claude/hooks/learning-capture.sh and configures it as a PostToolUse hook in the agent's settings. This configuration causes the script to execute automatically after every bash command tool call to filter and process results.
- [DATA_EXFILTRATION]: The skill captures stdout and stderr from failed shell commands and persists them in ~/.local/share/terraphim/learnings/. Users should be aware that this process may inadvertently store sensitive information, such as API keys or access tokens, if they appear in command arguments or failure messages.
- [PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by recording untrusted error output from various bash commands into the agent's knowledge base. * Ingestion points: tool_result (stdout/stderr) from failed bash commands. * Boundary markers: Captured data is wrapped in YAML frontmatter and Markdown sections. * Capability inventory: Uses vendor-specific CLI tools (terraphim-agent) to write and query stored learnings. * Sanitization: The documentation does not specify sanitization or filtering of captured error content.
Audit Metadata