The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses a bootstrap script (references/support/runtime-bootstrap.mjs) to fetch a versioned manifest and component archives from the official GitHub repository of the author (github.com/time-attack/gstack). This process involves downloading tar.gz files containing the GStack runtime from trusted GitHub release hosts.- [REMOTE_CODE_EXECUTION]: The bootstrap process installs code and binary components (such as Chromium and Bun) to the local system. While these are executed locally, the source is an external download. The skill implements security measures including SHA-256 integrity checks and optional Cosign verification to ensure the authenticity of the downloaded code before extraction and deployment.- [COMMAND_EXECUTION]: The skill instructions and internal modules frequently invoke shell commands and local binaries (e.g., git, grep, xcodebuild, node) to perform debugging tasks, search for code patterns, and manage investigation state. These operations are within the scope of a software debugging tool.- [DATA_EXFILTRATION]: The skill includes an optional feature called 'GBrain' (references/CODE-INTELLIGENCE.md) that can send repository content to an external semantic search service. Additionally, it supports Context.dev (references/WEB-CONTEXT.md) for web search capabilities. Both features are clearly described as opt-in and require explicit user consent before any data is transmitted externally.- [PROMPT_INJECTION]: The skill processes potentially untrusted data from repository code, logs, and error messages (references/legacy/investigate.md). It explicitly handles this risk by instructing the agent to treat such content as untrusted data rather than instructions.
Ingestion points: Reads repository files, logs, and tool outputs via standard agent tools.
Boundary markers: Explicit policies in AUTHORITY-POLICY.md define repository content and tool output as untrusted data.
Capability inventory: Includes file system access and shell command execution.
Sanitization: Relies on the agent following explicit instructions to treat inputs as untrusted data.

debug