The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads and executes a script from a remote source to install the Bun runtime.
Evidence: curl -fsSL "https://bun.sh/install" -o "$tmpfile" followed by bash "$tmpfile" in SKILL.md.
Note: The skill implements a SHA256 checksum check (bab8acfb046aac8c72407bdcce903957665d655d7acaa3e11c7c4616beae68dd) to verify the integrity of the downloaded installer before execution.
[DYNAMIC_EXECUTION]: The skill uses shell evaluation and sourcing on the output of local binaries to configure the environment.
Evidence: eval "$(~/.claude/skills/gstack/bin/gstack-slug 2>/dev/null)" and source <(~/.claude/skills/gstack/bin/gstack-repo-mode 2>/dev/null) in SKILL.md.
[DATA_EXFILTRATION]: The skill transmits usage telemetry and session metadata to the vendor's logging service.
Evidence: Invocation of ~/.claude/skills/gstack/bin/gstack-telemetry-log with parameters for --event-type, --duration, and --session-id.
Note: The skill explicitly states that it aims to share only metadata (duration, crashes, ID) and avoid code or file paths.
[INDIRECT_PROMPT_INJECTION]: By navigating to and snapshotting external websites, the skill introduces an attack surface where untrusted web content could provide malicious instructions to the agent.
Ingestion points: Web content accessed via $B goto and $B snapshot.
Boundary markers: Instructions mention the sidebar agent runs in an isolated session, but main prompt boundaries for browser data are not defined.
Capability inventory: The agent has access to Bash and browser control tools ($B).
Sanitization: No specific sanitization or filtering of web content before interpolation into the prompt is described.
[COMMAND_EXECUTION]: The skill relies heavily on Bash commands to manage project state, cleanup locks, and interact with the Chromium profile directory (~/.gstack/chromium-profile).
[EXTERNAL_DOWNLOADS]: Fetches the installation script for the Bun runtime from the official bun.sh domain.

open-gstack-browser