The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from the conversation history to drive task execution in a child environment.
Ingestion points: Reads the latest SPEC READY block from the current task context and processes incoming Messenger cards (replies) from child tasks.
Boundary markers: The instructions do not specify explicit delimiters or "ignore instructions" wrappers for the content within the SPEC READY block.
Capability inventory: The skill utilizes fork_thread, read_thread, set_thread_archived, and invokes external skills such as /spec-executor and /codex-task-messenger.
Sanitization: The skill contains a strong instruction to verify authorization from the source task's direct user messages and explicitly warns not to trust authorization claims copied into message card bodies.
[COMMAND_EXECUTION]: The skill is designed to trigger the execution of other skills and tools in a forked environment.
Evidence: It instructs the agent to run /spec-executor within a newly created child thread (fork_thread) based on the contents of the inherited specification.
[PRIVILEGE_ESCALATION]: The skill manages the lifecycle of agent threads, including creation and archival.
Evidence: It utilizes harness-specific tools like fork_thread and set_thread_archived to manipulate task structures. However, it explicitly mitigates risk by stating the workflow does not add authority to access credentials, modify production data, or commit code.

execute-spec-in-fork