The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages a background daemon process using platform-specific service managers to ensure persistence.
On macOS, it creates a launchd configuration file at ~/Library/LaunchAgents/com.claude-to-im.bridge.plist to manage the lifecycle of the bridge process.
On Windows, the supervisor-windows.ps1 script provides functionality to install the daemon as a Windows Service using WinSW or NSSM.
In src/codex-provider.ts, the skill uses a Function constructor with a dynamic import statement to lazily load the optional @openai/codex-sdk dependency at runtime.
[EXTERNAL_DOWNLOADS]: The scripts/doctor.sh diagnostic tool and the various platform adapters in src/adapters/ perform outbound network requests to official messaging platform APIs (including Telegram, Feishu, QQ, and DingTalk) to validate credentials and poll for incoming messages.
[PROMPT_INJECTION]: The bridge architecture creates a surface for indirect prompt injection as it retrieves untrusted data from messaging platforms and forwards it to the agent context.
Ingestion points: Incoming messages from messaging channels enter the agent's context through the polling logic defined in src/adapters/.
Boundary markers: There are no explicit delimiters or instructions observed in the code that would warn the model to ignore potential injection content within the IM messages.
Capability inventory: The skill's configuration in SKILL.md allows the agent to use highly privileged tools including Bash, Read, Write, and Edit.
Sanitization: While the skill implements log redaction for its own credentials, raw message text from IM platforms is passed to the AI model's query interface.

claude-to-im