The Agent Skills Directory

[SAFE]: The skill operates as a legitimate driver for a development evaluation harness. Its primary function is to coordinate the execution of local tests and the capture of session data for performance grading.
[COMMAND_EXECUTION]: The skill instructs the agent to use various subcommands of the t3 CLI (e.g., t3 eval run, t3 eval capture-subagent). These are standard operational commands for the intended development workflow.
[DATA_EXPOSURE]: The skill identifies and accesses specific file paths within the agent's session directory (~/.claude/projects/) to retrieve sub-agent execution logs. This is necessary for the skill's purpose of capturing transcripts for evaluation and does not involve exfiltration of sensitive user credentials.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing untrusted data by reading execution transcripts generated by sub-agents.
Ingestion points: Reads sub-agent JSONL files from ~/.claude/projects/<slug>/<session>/subagents/.
Boundary markers: None explicitly defined in the skill instructions; markers are likely managed by the underlying t3 CLI tool.
Capability inventory: Uses shell command execution via the t3 toolset.
Sanitization: Not explicitly described in the skill body; relies on the t3 tool's internal grading logic.

running-evals