devex-review
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONDYNAMIC_CONTEXT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill uses several utility scripts located in
~/.claude/skills/gstack/bin/(e.g.,gstack-skill-start,gstack-question-preference,gstack-skill-end). These are vendor-owned tools belonging to the 'gstack' framework used for telemetry and state management. - [COMMAND_EXECUTION]: The skill executes shell commands to detect git platforms (
git remote get-url), check CLI status (gh auth status,glab auth status), and run the target product's CLI for error capture (gstack-error-capture). These are necessary for auditing the developer experience. - [DATA_EXFILTRATION]: Telemetry data (outcome, session ID, TTHW) is sent to the vendor's local analytics directory and via
gstack-skill-end. This is a standard feature of the 'gstack' ecosystem for tracking audit progress and does not involve sensitive user credentials. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted web content (snapshot trees, page text, console output) through the browser. It includes explicit boundary markers (
--- BEGIN UNTRUSTED EXTERNAL CONTENT ---) and instructions to treat this data as content, not instructions, to mitigate injection risks. - [DYNAMIC_EXECUTION]: The skill generates and executes JavaScript snippets within the Aside browser (
aside repl) to perform audits. This execution is confined to the browser environment and follows a strict 'consent to look, not to act' policy for non-local targets. - [DYNAMIC_CONTEXT_INJECTION]: The skill uses the dynamic context injection pattern (e.g.,
!git remote get-url origin) to populate the skill instructions with live project data at load time. This is used for platform detection and is benign in this context.
Audit Metadata