The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a relay script (relay.mjs) to execute the vibe CLI tool. It also invokes git to track changes in the working tree and taskkill on Windows to ensure clean process termination.
[EXTERNAL_DOWNLOADS]: The documentation guides users to install the mistral-vibe tool via uv. These are well-known, legitimate tools from Mistral AI and Astral respectively.
[DYNAMIC_EXECUTION]: The relay script includes functionality to dynamically resolve and execute a lane.mjs script from a sibling skill (delegate-setup) if specific configuration flags are used. This is used for cross-skill integration.
[DATA_EXFILTRATION]: The skill passes the task 'brief' (instructions) as a command-line argument to the vibe process. The documentation explicitly notes that this exposure is a property of the OS process list and advises against placing secrets directly in the brief on shared machines.
[INDIRECT_PROMPT_INJECTION]: Because the skill processes 'briefs' which are natural language task descriptions, there is an inherent surface for indirect prompt injection. The skill mitigates this by enforcing a 'Review and Land' boundary where the orchestrator must manually verify all diffs and re-run project tests before committing work.
Ingestion points: Briefs provided via file or stdin to relay.mjs (processed in SKILL.md and references).
Boundary markers: The skill suggests using XML-style delimiters (e.g., <task>, <verification_loop>) within briefs to structure the agent's context.
Capability inventory: File system writes (via Vibe), arbitrary shell execution (only if --full-access is authorized), and git operations.
Sanitization: The skill relies on human review and the execution of project-specific 'gates' (tests, linters) to sanitize the implementer's output before it is accepted.

vibe-delegate