TDD Skill Evaluation Exercise

This repo tests one hypothesis: good skills and context engineering can help lower models perform closer to stronger models.

Before You Run

• Read .codex/skills/tdd/SKILL.md to understand the TDD behavior being tested.
• Read .codex/skills/eval-result/SKILL.md to understand how coder evidence is packaged.
• Read .codex/agents/*.yaml to see which model each coder agent uses.
• Read .codex/skills/skill-eval/references/rubric.md so you know how results are judged.

Prerequisites

• Codex with agent/subagent support.
• Git with a clean baseline commit.
• Basic tooling for your chosen stack, for example JS/npm, Python, or Go.

Run

1. Ask Codex: Use $env-setup to prepare a TDD benchmark for <stack/problem>.
2. Answer the setup questions for stack, problem, commands, constraints, and acceptance criteria.
3. Confirm prompt and .codex/tdd-setup.json were created.
4. Commit or snapshot the clean baseline.
5. Ask Codex: Use $skill-eval to start eval.
6. Ask Codex: Use $skill-eval to show report.

$skill-eval runs junior, mid, and senior coder agents in isolated worktrees. Each coder uses $tdd and $eval-result; the evaluator scores outputs against the senior baseline.

To iterate, tune $tdd from report feedback, commit the skill change, throw away generated run state, and run the flow again.

Note: this is experimental and built for Codex. It has not been verified on other agent systems.