The Agent Skills Directory

[COMMAND_EXECUTION]: The skill defines an 'Empirical Arbiter' role that utilizes the codex exec tool to perform experiments. This involves the generation and execution of shell commands and scripts to resolve factual disagreements between solvers.
[REMOTE_CODE_EXECUTION]: The workflow integrates the mcp__codex__codex tool to provide an independent solver role (Solver B). The skill specifies that execution occurs within a sandbox with 'read-only' defaults, and any operations requiring 'workspace-write' permissions are gated by explicit user confirmation.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes user-defined inputs across multiple autonomous agent phases:
Ingestion points: Untrusted data enters the system through the initial question and the 'Decision Contract' established in Phase 0 (SKILL.md).
Boundary markers: The skill employs template structures in the references/ directory and 'structural sealing' protocols to isolate the context of individual agents.
Capability inventory: The system has access to the codex tool for execution and the Task tool for spawning subagents (SKILL.md).
Sanitization: The orchestrator anonymizes solver outputs as 'X' and 'Y' before passing them to mapping and synthesis roles, which serves to limit the propagation of instructions between different agent stages.

Contradiction Lift