pier
A Harbor-compatible framework for evaluating coding agents such as Claude Code, Codex and Cursor CLI in sandboxed Docker or Modal environments, with support for installing agents in air-gapped tasks and a trajectory viewer.
Pier is a fork of the Harbor coding-agent evaluation framework, built to be smaller and more opinionated, that reads Harbor's task format and runs trials against it. It lets agents that must be installed inside the sandbox, such as Claude Code or Codex, declare install scripts and a network allowlist so they still work under tasks with internet access disabled, honored by both its Docker and Modal environments.
It also implements an augmented version of the ATIF trajectory schema with strict one-step-per-turn recording and no fabricated assistant text, adds a chat-style trajectory viewer and a critique command that inspects a completed trial with a fresh agent in a fresh sandbox. Supported agents include Claude Code, Codex, Cursor CLI, Gemini CLI, OpenCode and mini-swe-agent, run against local Harbor-format task directories. Apache-2.0 licensed, installed via uv or pip.