The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/relay.mjs script executes external binaries including vibe and git. When the --full-access flag is utilized, it enables the auto-approve profile for the vibe agent, granting it the capability to execute arbitrary shell commands and tools on the host system. This high-privilege execution environment is a core feature of the skill but requires careful management to prevent the execution of malicious commands.
[EXTERNAL_DOWNLOADS]: The skill setup involves downloading and installing the mistral-vibe package from an external source using the uv tool. mistral-vibe is a well-known CLI tool from Mistral AI, and this dependency is documented as a prerequisite for the skill's operation.
[DATA_EXFILTRATION]: As noted in scripts/relay.mjs and references/writing-the-brief.md, the task 'brief' is passed to the CLI via a command-line argument (--prompt=). This makes the content of the prompt, which may contain sensitive context or proprietary logic, visible in the system's process list, potentially exposing it to other users or monitoring tools on the same host.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it processes 'briefs' that may incorporate untrusted data from the project workspace or external sources.
Ingestion points: Untrusted data enters the agent context through the brief text processed in SKILL.md and scripts/relay.mjs.
Boundary markers: The skill suggests using XML delimiters (e.g., <task>) in references/writing-the-brief.md to structure the input, though these do not provide programmatic isolation.
Capability inventory: Across its scripts and the vibe tool it wraps, the skill can modify the local filesystem and execute arbitrary shell commands.
Sanitization: The skill lacks automated sanitization of the input brief, relying instead on a manual human review of the generated diffs ('Review and Land') to identify and block malicious or unintended changes.

vibe-delegate