Imported from sergey-gusev94/agentic-assessment-toolkit (
AGENTS.md). Install upstream withnpx skills add sergey-gusev94/agentic-assessment-toolkit. Copyright stays with the author.
AGENTS.md
Project purpose
Agentic Assessment Toolkit is a Python toolkit for benchmarking coding
agents on university coursework and for LLM-assisted grading of student
submissions, built as a thin layer over Harbor.
The distribution is named agentic-assessment-toolkit; the Python import
package is agentic_assessment_toolkit. Documentation and examples may use
aat as a local import alias.
Decisions and contracts live under docs/ (start with docs/design.md).
Documentation is maintained in place, not as a changelog: superseded
decisions are rewritten, not preserved as history.
docs/design.md and the other contract documents describe only what is
implemented, in present tense; docs/roadmap.md holds only planned
work, each item with the condition that triggers it. When work lands,
delete its roadmap entry and document the built behavior in the
appropriate document, rewritten as a description of what exists rather than a
relocation of the planning text.
Development
The project is in its initial development stage. Backward compatibility is not required unless explicitly requested.
Writing
Everything here is read by a person: documentation, comments, prompts, commit messages, PR bodies, and answers to questions. Write so a competent engineer who has not followed this work understands it on the first read.
- Use the everyday word. Reach for a specialized term only when no common word is exact, and say what it means where you first use it.
- Coin a new term only when the concept recurs and no plain phrasing stays readable.
- One concept, one name. When the repository already has a word for something, use that word everywhere; do not introduce synonyms.
- Plain wording never drops substance. Keep every decision, caveat, and exact identifier; simplify the sentences, not the information.
- Do not use em dashes or en dashes in replies, comments, documentation, prompts, code, or any other file. This includes the LaTeX double-hyphen and triple-hyphen forms. Rephrase the sentence or use commas, parentheses, colons, or semicolons instead.
Validation
Run make check after changes. It performs formatting verification, linting,
strict type checking, and tests. Run make format to apply formatting.
Normal tests must be deterministic, local, offline, and credential-free.
Live execution, including Harbor runs, Docker, subscription-authenticated agent or model calls, and network access to model providers, is out of scope for repository work. Do not attempt it or make any deliverable depend on it; the maintainer performs all live validation outside the repository.