The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions contain explicit directives to override standard agent safety and communication protocols.
Evidence: The "Second Law" (不问不退) commands the agent to suppress intermediate inquiries and confirmation prompts, even for risky operations.
Evidence: Specific instructions to "forbid" phrases like "Is this operation risky?" and bypass standard safety prompts to maintain autonomous momentum.
[DYNAMIC_EXECUTION]: The skill architecture is built on an "Eternal Loop" that autonomously generates and executes plans without real-time supervision.
Evidence: Pseudo-code implementation of execute(plan) based on user-provided tasks which may involve runtime command generation.
Evidence: Automated self-healing logic that attempts to fix errors in generated code or environment state without human intervention.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied task descriptions while maintaining extensive system capabilities.
Ingestion points: The task parameter in the /修仙 and /sleepless commands.
Boundary markers: Absent; there are no delimiters separating the task description from the autonomous execution logic.
Capability inventory: Includes shell command execution, file system modification, and cross-session state persistence.
Sanitization: Absent; the skill does not implement validation or filtering of the task content prior to execution.
[PERSISTENCE]: The skill implements a persistence layer designed to maintain agent state and resume execution automatically across sessions.
Evidence: Checkpoint system saving "State Snapshots" and "Decision Logs" to a persistent storage layer.
Evidence: Automated session recovery logic that resumes execution without user confirmation in new sessions.

sleepless