improving-codebases
Pass
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes various system commands through the
node:child_processmodule to interact with Git and run model checkers. - Evidence:
scripts/checkpoint.mjsandscripts/churn.mjsusespawnSyncandexecFileSyncto rungitcommands (e.g.,git status,git rev-parse,git ls-tree). - Evidence:
scripts/run-tlc.mjsexecutes the Choice Architecture (TLA+) Model Checker (tlc) orjavaviaspawn. - [DYNAMIC_EXECUTION]: The skill identifies and runs project-specific test commands derived from the repository's configuration files.
- Evidence:
scripts/run-bounded.mjsexecutes arbitrary commands provided in arequest.jsonfile (typicallynpm testor similar) viaspawn. This is regulated by an environment allowlist (ENV_ALLOWLIST) to prevent uncontrolled privilege or credential leakage. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository content (source code, comments, Git history) which could contain malicious instructions.
- Evidence Ingestion: The skill reads file content using
git ls-filesandreadFileSyncinSKILL.md(Step 3) andscripts/churn.mjs. - Capability Inventory: The skill has the ability to write to the file system and execute shell commands via
run-bounded.mjs. - Sanitization/Boundaries: The skill includes mandatory security protocols in
shared/external-data.mdandshared/independent-review.mdwhich explicitly instruct the agent to treat repository content as data and never as instructions. It also uses independent 'goldfish' reviewers (fresh-context subagents) to verify proposed changes, reducing the risk of a single-session injection compromise.
Audit Metadata