agent-transcript
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFE
Full Analysis
- [DATA_EXPOSURE]: The skill accesses local agent session logs stored in user-specific directories (e.g.,
~/.claude,~/.openclaw,~/.codex). To mitigate data leakage, it utilizes a redaction engine targeting secrets, tokens, and local identifiers, and includes a "fail-closed" mechanism that prevents rendering if sensitive patterns are detected in the final output. - [COMMAND_EXECUTION]: The
agent-transcriptscript executes a localcodexbinary viaspawnto interface with a local application server through standard I/O for reading session thread data. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted session data from local agent logs and incorporates it into PR bodies.
- Ingestion points: Reads JSONL session logs from various agent platform directories in the user's home directory.
- Boundary markers: Transcript content is delimited using HTML comments (
<!-- agent-transcript:start -->) and collapsed<details>blocks. - Capability inventory: The script performs local file operations and executes the local
codexcommand; it does not perform network operations. - Sanitization: The skill implements regex-based redaction, explicitly omits system and developer prompts, and applies HTML escaping for local previews.
Audit Metadata