creating-letta-code-channels
Pass
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill includes a TypeScript utility (
scripts/scaffold-user-channel-plugin.ts) designed to be executed vianpx tsx. This script automates the creation of directory structures and boilerplate files (JSON and MJS) within the user's home directory under~/.letta/channels. - [INDIRECT_PROMPT_INJECTION]: The skill establishes an infrastructure for message channel adapters that ingest untrusted data from external sources (e.g., Telegram, Slack) and deliver it to the agent.
- Ingestion points: Data enters the agent's context through the
adapter.onMessagehandler, which processes external platform messages as defined inreferences/user-plugins.mdand the scaffoldedplugin.mjs. - Boundary markers: The skill instructions in
SKILL.mdandreferences/first-party-channels.mdexplicitly advise developers to suppress tool approval prompts on public channels to mitigate the risk of unauthorized command execution via spoofed replies. - Capability inventory: Plugins use the
MessageChanneltool to enable agent communication. The development environment relies onbunfor package management and testing, andnpxfor scaffolding. - Sanitization: The provided plugin skeleton implements a
formatTextfunction within its action handler to process outgoing message text, providing a layer of content control. - [DYNAMIC_EXECUTION]: The skill facilitates the development of dynamic user plugins that are loaded as external ECMAScript modules (
.mjs) at runtime by the platform's channel service.
Audit Metadata