omh-harness-session-inventory
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill accesses local filesystem paths to retrieve session metadata and transcripts from various agent platforms.
- Evidence: References to
~/.codex/sessions/and~/.claude/projects/<project-slug>/for metadata collection. - Note: The skill includes instructions to only read transcript bodies when explicitly requested by the user and to produce 'secret-redacted' reports.
- [COMMAND_EXECUTION]: The skill executes commands using a project-specific CLI tool (
omh). - Evidence: Usage of
omh runtime recordandomh harness mcp-tool-name-compatibility --snapshot <path>. - [INDIRECT_PROMPT_INJECTION]: The skill processes external data from multiple sources, creating a potential surface for indirect instructions.
- Ingestion points: Processes external MCP tool naming compatibility snapshots, session metadata from local directories, and session transcripts.
- Boundary markers: Employs 'prepared-vs-observed' and 'not-evidence' boundaries to distinguish between system instructions and external data.
- Capability inventory: Executes CLI commands via the
omhtool and reads local files viaskill_view. - Sanitization: Explicitly mandates 'secret-redacted' outputs for inventory drift reports and compatibility snapshots.
Audit Metadata