The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of local shell commands for repository auditing and management, including git (for integrity verification and branch management), ps (for monitoring worker PIDs), and the codex CLI for subagent execution.
[REMOTE_CODE_EXECUTION]: It utilizes external execution environments such as crabbox and Testbox for running heavy test suites, Docker containers, and live provider proofing, which involves executing code in remote sandboxes.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core function of ingesting and acting upon untrusted external data.
Ingestion points: Reads and processes the OpenClaw source code in implementations worktrees, user-submitted GitHub PR descriptions, scenario YAML files, and Control UI screenshots.
Boundary markers: Employs read-only sandboxes for reviewer workers and immutable Git-object verification to ensure source integrity, though it lacks explicit natural language delimiters (like XML tags) to wrap untrusted data during interpolation.
Capability inventory: Includes CLI execution (codex exec), broad file system access via Git, repository automation (scripts/pr), and remote infrastructure leases (Testbox).
Sanitization: Includes strong explicit instructions to classify source trust and avoid executing unreviewed contributor scripts or package installations on credential-hydrated machines.
[CREDENTIALS_UNSAFE]: While the skill manages credentials, it follows safe practices by instructing the agent to use secret-backed workflows and explicitly forbids logging, printing, or exporting credentials into untrusted workers.

auto-qa