auto

What problem does it solve? Running a full mechanistic interpretability research workflow — literature review, hypothesis formulation, experiment execution, robustness verification, and iterative refinement — requires manually chaining many stages and agents. This Skill automates that entire pipeline end-to-end, delegating each stage to an isolated agent with its own context window and configurable model. ## Core Features & Use Cases - Multi-Stage Orchestration: Chains claim, experiment, verify, and iteration stages, each running in a dedicated agent with per-stage model selection (opus/sonnet/haiku) and AUTO_PROCEED-governed gates. - Flexible Behavior & Mechanism Modes: Supports given, given-validation, and discovery behavior sources, plus user-specified or auto-routed mechanism families, including a phenomenon-validation M0 gate. - Resume & Safety Guards: RESUME mode skips completed stages based on artifact checks, and a multi-round guard prevents silently overwriting prior-round outputs. - Use Case: A researcher provides a research direction about an LLM behavior, and the pipeline autonomously generates claims, runs mechanism experiments on local GPUs, stress-tests claims with swap variants, and iterates until a target review score is reached. ## Quick Start Run the auto pipeline with a research direction such as asking it to investigate how a specific model behavior emerges internally, optionally overriding flags like auto-proceed or per-stage models.