Part 1 of 4 · 1 min read
Overview
Agent frameworks, tool servers and evaluation libraries are mostly open source and change fast. Fixing a parser, adding a test or improving documentation in one of them is a visible demonstration of the skills in this level: reproducing a failure, localising it, writing evidence and communicating clearly.
About the example, and what is real
Real: a toy git repository built and run by the code (real git, real pytest), three real branches and their real diffs, a unified-diff parser, a diff checker, a description checker, a commit-message checker and 23 tests in code/mini-pr. Assumed: the library ("tinyagent"), the bug, the thresholds (400 changed lines, 25% formatting churn, 3 areas, 72-character subject) and the list of rules; contribution norms described here come from general knowledge, not from a project's current documentation (my attempt to fetch GitHub's guide was blocked in this session). Not real: no real repository, issue or maintainer was involved.