AGENTS.md Generator

Generator / Templates / Ray workload

AGENTS.md for Ray workload

Distributed Python and ML on Ray. This AGENTS.md template combines 110 rules from 13 packs and runs to 150 lines.

Open this template in the AGENTS.md generator

What is inside

The generated AGENTS.md

This is the file the generator produces for the Ray workload preset, before you add your project name and commands.

# AGENTS.md

## Agent behavior

- For changes touching more than three files, outline the plan and wait for confirmation before editing.
- Ask a clarifying question when requirements are ambiguous instead of guessing.
- Keep diffs small and focused on the request; do not refactor unrelated code.
- Run the relevant tests, linter and type checker before reporting a task as done, and report failures honestly.
- Never delete, skip or loosen a failing test to make it pass; fix the cause.
- Do not add new dependencies without explicit approval; prefer the standard library and existing packages.
- Do not edit generated or vendored files; change the source and regenerate.
- Never read, print or commit secrets, .env files or credentials.
- Read the relevant code and tests before proposing a change; do not edit files you have not read.
- Fix the root cause rather than suppressing the symptom with a workaround or silenced warning.
- Do not add lint or type suppressions to get a clean run; fix the underlying issue.
- Report failing tests, skipped steps and unverified claims plainly; never claim success without evidence.
- Verify library APIs in the installed version or its docs before using them; do not invent functions.
- Do not change public behavior, file formats or APIs unless the task asks for it.
- Never hardcode values or special-case inputs to satisfy a test.
- If an approach fails twice, stop and ask instead of trying increasingly invasive workarounds.
- Before editing a directory, read the nearest AGENTS.md and follow it over the root file where they differ.
- Preserve unrelated working tree changes. Never revert or rewrite edits you did not make.

## Logging

- Emit logs as structured JSON, one object per line, with timestamp, level, message and context fields.
- Use levels consistently: error for failures needing action, warn for recoverable problems, info for lifecycle events, debug for diagnostics.
- Never log secrets, tokens, passwords or personal data; redact before logging.
- Remove debug prints and console.log calls before finishing; use the project logger.
- Log errors once, at the boundary that handles them, with enough context to reproduce.
- Log timestamps in UTC using ISO 8601.
- Keep log messages static and put variable data in structured fields.
- Do not log inside tight loops or at info level for every request detail.

## Testing

- Add or update tests for every behavior change and bug fix.
- For bug fixes, write a failing test that reproduces the bug before fixing it.
- Prefer many fast unit tests, fewer integration tests and a small number of end-to-end tests.
- Mock only external boundaries (network, clock, filesystem); use real collaborators otherwise.
- Keep tests deterministic: no real time, randomness or network without control.
- Keep tests independent; no shared mutable state or required execution order.
- Structure tests as arrange, act, assert with one behavior per test.
- Build test data with small factories instead of large shared fixtures.
- Treat a flaky test as a bug: find the cause instead of adding retries or sleeps.
- Use snapshot tests only for small stable outputs and review snapshot diffs carefully.
- Test public behavior rather than private functions or implementation details.
- Make assertions specific enough that the test fails if the behavior breaks.

## Git and pull requests

- Write commit messages in Conventional Commits format (feat:, fix:, docs:, refactor:, test:, chore:) and mark breaking changes with ! or a BREAKING CHANGE footer.
- Keep pull requests small and focused on one change.
- Write commit subjects in the imperative mood, 72 characters or fewer.
- Never force-push to shared branches and never rewrite published history.
- Run lint, type check and tests before committing.
- Never commit secrets, build output or large binaries.
- Describe motivation, approach and testing in the pull request description.
- Make each commit one logical change that builds and passes tests.
- Do not amend or rebase commits that are already pushed to a shared branch.
- Add build output, local env files and editor files to .gitignore instead of committing them.
- Open pull requests only when CI is green and the diff is self-reviewed.
- Never bypass commit hooks or signing with --no-verify unless explicitly instructed.

## Security

- Load secrets from environment variables or a secret manager; never hardcode them.
- Use parameterized queries and safe APIs; never build SQL, shell commands or HTML by string concatenation with user input.
- Enforce authentication and authorization on the server for every protected operation.
- Pin dependency versions, commit lockfiles and review new packages before adding them.
- Grant the minimum permissions needed for tokens, service accounts and containers.
- Use well-reviewed cryptographic libraries; never invent your own algorithms or store plain passwords.
- Encode output for its context and avoid injecting raw HTML; sanitize rich text with a vetted library.
- Protect state-changing browser requests with CSRF tokens or SameSite cookies.
- Configure CORS with explicit origins; never reflect arbitrary origins with credentials.
- Use HttpOnly, Secure, SameSite cookies and rotate session identifiers on login.
- Validate and allow-list URLs before fetching user-supplied addresses to prevent SSRF.
- Hash passwords with Argon2id, or scrypt or bcrypt where it is unavailable, with a unique salt; never use fast hashes such as SHA-256.
- Run secret scanning in CI and rotate any credential that has been exposed.
- Validate untrusted input against a schema. Never run eval, exec or unpickling on it.

## Workflow and tooling

- Run tools through the project scripts (npm run, make, task) instead of invoking underlying binaries directly, so flags and environments match CI.
- Run the single aggregate check command (for example pre-commit or a lint script) before finishing; fix every failure and re-run until it passes.
- Run the specific test files related to your change first, and run the full suite only before handing off.
- Before editing a directory, read the README and any nested AGENTS.md that applies to it.
- Put throwaway scripts and experiment output in a gitignored scratch directory and delete them when done.
- After switching branches or pulling dependency changes, reinstall and rebuild before running tests.
- When CI fails, find the exact failing job and command, reproduce it locally, then fix the cause instead of re-running blindly.
- Stop any dev servers or watchers you started once you are done.
- Ask before running destructive or irreversible commands such as dropping databases, deleting branches, rm -rf or force operations.
- Prefer existing scripts, generators and templates in the repo over writing new ad hoc tooling.
- When the repository documents skills or runbooks for a task (migrations, releases, dependency bumps), follow them instead of improvising.
- Never start long-running dev servers or watchers in the foreground. Use targeted one-shot commands.

## Verification and CI feedback

- Run the narrowest relevant test first, then widen to the package and workspace checks as confidence grows.
- Run the project formatter and linter on changed files before committing, using the repo scripts or hooks.
- For a bug fix, reproduce the failure first, then show the regression test failing without the fix and passing with it.
- Report a task as done only after running the checks that prove it. If you could not run them, say so.
- Before finishing, re-read the full diff as a reviewer with no knowledge of this session and fix what would confuse them.

## Python conventions

- Follow PEP 8 and use Ruff for linting and formatting.
- Add type hints to all function signatures and run mypy or pyright.
- Use pathlib for filesystem paths and context managers for resources.
- Model structured data with dataclasses or Pydantic, not loose dicts.
- Write tests with pytest, using fixtures and parametrize instead of loops.
- Catch specific exceptions, never bare except.
- Never use mutable default arguments.
- Use the logging module with a module-level logger, not print.
- Do not mix blocking calls into async code; use asyncio.to_thread for unavoidable blocking work.

## uv conventions

- Add dependencies with uv add so pyproject.toml and the lockfile stay in sync. Never use pip install in a uv project.
- Run tools with uv run --locked, which fails if the lock is stale. Use --frozen only when the lock must not be checked.
- Upgrade a single package with uv lock --upgrade-package <name> instead of upgrading everything.
- Never use the system python3 or a bare pip for project work. Go through uv and the project environment.

## Ruff conventions

- Use ruff format as the formatter and ruff check --fix for autofixable lint.
- Keep Ruff settings in pyproject.toml or ruff.toml under the lint and format tables.
- Choose rule families with select or extend-select rather than enabling everything.
- Scope exceptions with per-file-ignores for tests, scripts and __init__.py instead of global ignores.

## pytest conventions

- Write tests as plain test_ functions with fixtures. Do not add test classes to new code.
- Use pytest.mark.parametrize with pytest.param(..., id=...) for named cases.
- Share fixtures through conftest.py and keep them small and composable.
- Use tmp_path for files and monkeypatch for environment and attributes.

## PyTorch and ML conventions

- Seed Python, NumPy and PyTorch, enable torch.use_deterministic_algorithms where exact repeatability matters, and log the seed with each run.
- Write device-agnostic code and move tensors and models explicitly.
- Call model.eval() and use torch.no_grad() or inference_mode for inference.
- Drive experiments from config files and track metrics and artifacts in an experiment tracker.
- Split data before any preprocessing fit and never tune on the test set.

## Ray conventions

- Do not call ray.get inside the loop that launches tasks. Launch every task first, then call ray.get once on the list of refs.
- Pass a large shared argument by reference: call ray.put once and pass the object ref to each task.
- Avoid launching huge numbers of tiny tasks. Batch work so each task amortizes scheduling overhead.

How to use it

Related guides

More AGENTS.md templates