captain
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses several local binaries located in
~/.claude/skills/gstack/bin/to manage its lifecycle and internal state. These includegstack-skill-startfor initialization,gstack-ship-queuefor merge train operations, andgstack-question-logfor recording user decisions. - [DATA_EXFILTRATION]: Telemetry data, including the session ID, workflow outcome, and execution duration, is logged to a local vendor binary (
gstack-skill-end) after completion. - [DYNAMIC_EXECUTION]: The agent is instructed to process and follow optional instruction blocks (
GSTACK_INSTRUCTION_BEGIN) that may be returned by thegstack-skill-startinitialization tool. This mechanism is secured by session ID validation and restriction to direct tool output. - [INDIRECT_PROMPT_INJECTION]: As a controller for other agents, the skill ingests potentially untrusted data from 'lanes' (other worker agents) via PR bodies, READY blocks, and reports. It mitigates risks by requiring strict field validation against expected templates and implementing a 'user-origin gate' that prevents tool output from modifying user preferences unless specifically prompted in a user message.
Audit Metadata