The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several local scripts (packs-resolve.py, peer-job-runner.py, light-webserver.js) to execute system commands like git and ps. These scripts use array-based argument passing to avoid shell injection and incorporate extensive security checks, such as ownership verification (os.fstat on POSIX and SID validation on Windows) and directory traversal prevention via realpath containment.
[EXTERNAL_DOWNLOADS]: The skill includes opt-in components for Slack research and 'Model Elevation.' The elevation process uses the Claude CLI to perform reasoning-heavy tasks. This process is strictly governed by a 'read-only' posture that limits the remote model's tools to repo-reading (Read, Glob, Grep) and web-searching (WebSearch, WebFetch), explicitly excluding write or shell access.
[DYNAMIC_EXECUTION]: The peer-job-runner.py script manages the lifecycle of detached peer jobs, including double-forking on POSIX and DETACHED_PROCESS creation on Windows. Additionally, light-webserver.js starts a local Node.js server to serve visual probes. Both systems use private scratch directories in /tmp (or %LOCALAPPDATA%) and enforce strict security boundaries, including the use of O_NOFOLLOW and atomic file operations.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest external data from Slack messages and repository files. To mitigate injection risks, the instructions include an 'Adversarial Reasoning Framework' and explicit directives (references/agents/slack-researcher.md and references/reasoning-elevation.md) to treat all external inputs as untrusted data to be interpreted rather than instructions to be followed.

ce-brainstorm