Full TrainForge interface for the coding agent. Capture conversations as
regression scenarios, run them against the user's agent (no LLM key
needed by default), label pending checks using the 20-NLP-consistency
rubric, score deterministically, diff against previous runs, debug
failures, drive A/B prompt+model swaps, render HTML reports, set up
pytest / CI. The skill is the test driver; it NEVER modifies the
user's agent (see HARD RULE below). Invoke when the user says "test
my agent", "turn this into a regression test", "I changed the prompt,
did anything break?", "is my agent reliable?", "compare these two
models", or pastes a chat with their agent and asks how to test it.
2026-05-28