The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill's documentation in references/primitives.md explicitly recommends using the --dangerously-skip-permissions flag when implementing the 'Ralph' bash loop pattern. This flag bypasses the platform's native security permission checks, allowing the agent to execute potentially dangerous actions without human approval or prompts.
[DYNAMIC_EXECUTION]: The skill describes and enables several dynamic execution patterns. The 'Ultracode' feature in references/primitives.md involves the agent writing its own JavaScript orchestration scripts to manage subagents. Furthermore, the 'Ralph' loop pattern involves continuous script generation and execution in a bash loop. The Stop-hook mechanism also supports executing arbitrary scripts to verify exit conditions.
[COMMAND_EXECUTION]: The core purpose of the skill is to facilitate unattended, autonomous execution of shell commands through the Bash tool. While the skill suggests various guardrails like iteration caps and cost budgets, the primary operational mode is the continuous execution of code and commands without direct human oversight.
[PERSISTENCE]: The skill utilizes the platform's /loop scheduler to maintain long-running operations. As documented in references/primitives.md, this allows the agent to schedule tasks (such as recurring checks or addressing PR comments) to run on a cron-like interval for up to three days, even across different terminal sessions.
[INDIRECT_PROMPT_INJECTION]: The skill defines a large attack surface for indirect prompt injection.
Ingestion points: The skill reads external, potentially untrusted signals such as test exit codes, linting outputs, coverage reports, and log files to determine if a loop should terminate (documented in SKILL.md and references/safety.md).
Boundary markers: There are no specified delimiters or 'ignore embedded instructions' warnings for these external data signals.
Capability inventory: The skill utilizes powerful tools including Bash, Write, Edit, Agent, and Workflow.
Sanitization: The instructions do not specify any filtering or sanitization of external tool outputs before they are used to drive agent behavior, creating a risk that malicious instructions in a test failure or log could influence the agent's next action.

loop-controller