The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The npm/scripts/postinstall.js script downloads a platform-specific binary (e.g., agentmail_0.4.0_macos_arm64.zip) from the official vendor repository at https://github.com/agentmail-to/agentmail-cli/releases. This is a standard distribution method for cross-platform Go binaries via NPM.
[COMMAND_EXECUTION]: The skill uses child_process.execFileSync in npm/bin/agentmail to execute the downloaded binary. Additionally, pkg/cmd/cmdutil.go may execute a local pager program (like less) to display long outputs, which is standard CLI behavior.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection because it ingests untrusted data from external sources.
Ingestion points: Commands like inboxes:messages list and inboxes:messages retrieve in pkg/cmd/inboxmessage.go fetch email content from the AgentMail API.
Boundary markers: Explicit boundary markers or warnings to ignore embedded instructions are absent in the CLI output.
Capability inventory: The skill has the capability to write files (downloading attachments in pkg/cmd/cmdutil.go), execute subprocesses (launching a pager in pkg/cmd/cmdutil.go), and perform network operations (API requests).
Sanitization: The skill performs standard formatting (JSON, YAML, or Pretty-print) but does not specifically sanitize email bodies for LLM instruction markers.

agentmail