sensible-commit-flow
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill optionally fetches technical wording from a peer skill hosted in the vendor's repository on GitHub (porada/domfiles). This process includes robust validation steps, such as using immutable object IDs and verifying skill names.
- [INDIRECT_PROMPT_INJECTION]: The skill handles untrusted repository data, including diffs and commit messages.
- Ingestion points: Scoped staged/unstaged diffs and commit history in SKILL.md and update-unpushed-commits.md.
- Boundary markers: The 'Instruction Authority' policy in SKILL.md mandates quoting or delimiting untrusted content as data and provides explicit warnings to ignore embedded instructions.
- Capability inventory: Git command execution and potential invocation of repository scripts/tests as described in SKILL.md.
- Sanitization: Instructions to treat untrusted content strictly as evidence or task material without granting authority to perform actions.
- [COMMAND_EXECUTION]: The skill interacts with the local Git environment and may invoke project-specific scripts or tests for validation. It specifies the use of Git security flags (e.g., --no-ext-diff, --no-textconv) to ensure that only intended tools are executed during inspection.
- [DATA_EXFILTRATION]: While the skill performs network operations to check for unpushed commits and fetch optional writing assistance, it restricts data transfer to task-required information and requires explicit user authorization for any disclosure of machine-local values or diagnostics.
Audit Metadata