agent-transcript
Pass
Audited by Gen Agent Trust Hub on Jun 14, 2026
Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes a local Node.js script to discover session logs, process text, and generate HTML previews.
- [DATA_EXPOSURE]: The skill is designed to read potentially sensitive agent logs (e.g., from
.claude/projects,.codex/sessions) to facilitate transparency in PRs. It implements multiple security controls: - Comprehensive regex-based redaction for API keys (OpenAI, AWS), GitHub tokens, JWTs, and auth headers.
- Redaction of local system paths and personal contact information (emails, phone numbers).
- A 'fail-closed' security check that aborts rendering if unsafe patterns are detected after redaction.
- Instructions mandate user confirmation and local HTML preview before any data is appended to a PR body.
- [INDIRECT_PROMPT_INJECTION]: As the skill processes logs from prior agent sessions, it inherently interacts with an attack surface where malicious instructions could be embedded in those logs.
- Ingestion points: Reads local JSONL files from various agent session directories (defined in
scripts/agent-transcript). - Boundary markers: Uses Markdown code blocks (
text) and HTML comments (<!-- agent-transcript:start -->) to isolate the transcript. - Capability inventory: The script performs local file reads, creates temporary files for previews, and outputs to stdout.
- Sanitization: Specifically filters out system/developer prompts, raw tool outputs, and internal browser/session data to minimize the footprint of the exported log.
Audit Metadata