The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script scripts/agent-transcript executes the local codex binary via child_process.spawn to retrieve thread data. This is a legitimate interaction with the host environment to access session history.
[CREDENTIALS_UNSAFE]: The skill is designed to read and process sensitive session logs from local directories such as ~/.codex and ~/.claude. It implements significant mitigations to prevent accidental exposure:
A redact function uses regular expressions to mask private keys, API tokens (OpenAI, GitHub, AWS), JWTs, auth headers, emails, and local system paths.
An unsafe check prevents the script from returning content if certain sensitive patterns are still detected after redaction.
Instructions mandate user authorization before any transcript is published to an external platform like GitHub.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted session logs which may contain malicious instructions from previous interactions.
Ingestion points: The render and find commands read .jsonl files from local session directories as seen in scripts/agent-transcript.
Boundary markers: Transcripts are enclosed in <details> tags and unique HTML comments () as defined in SKILL.md.
Capability inventory: The skill uses the fs module for file operations and child_process to call local binaries within scripts/agent-transcript.
Sanitization: Includes multi-layered redaction, byte/line-based truncation to prevent context window flooding, and explicit omission of system prompts and raw tool outputs.

agent-transcript