The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill relies on the agent generating and executing ad-hoc Bash shell commands and Node.js scripts to parse local session data. This is secured via strict input validation of session identifiers (preventing path traversal and shell metacharacter injection) and the use of owner-restricted temporary directories (chmod 700) for all intermediate artifacts.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from previous agent conversations, which can contain embedded instructions.
Ingestion points: Reads JSONL and SQLite session stores from .claude, .codex, and opencode local directories.
Boundary markers: Employs explicit allow-lists to extract only human and assistant text blocks while strictly filtering out system instructions, tool outputs, and subagent noise.
Capability inventory: Utilizes bash, jq, and node:sqlite for file parsing and data extraction.
Sanitization: Implements mandatory schema fingerprinting to verify data structure before extraction and enforces a strict 'local-only' policy, ensuring no data is exfiltrated to external services.
[DATA_EXPOSURE_AND_EXFILTRATION]: While the skill targets sensitive logs, it explicitly avoids credential-bearing tables in relational stores (e.g., account, credential) and prohibits reading system environment snapshots or shell history.

extract-agent-sessions