The repositories

Find out which one you are in with one cheap call: ls /testbed. The source is either a top-level
package directory or under src/. Then read the matching section and nothing else.

Two questions are worth answering before you touch anything: where does the module in the issue live,
and which test file covers it. The second is what protects you from scoring zero, because the tests
kept for grading are drawn mostly from the file that covers the code you changed.

sqlglot

• Source: sqlglot/ at the top level. Tests: tests/.
• The parts: tokens.py and the tokenizer, parser.py, generator.py, expressions (the AST node
  classes — a module in some versions, a package in others), optimizer/ with one module per rule
  (qualify_columns, unnest_subqueries, pushdown_predicates, merge_subqueries, simplify,
  eliminate_*, annotate_types, scope), dialects/ with one module per SQL dialect, transforms.py,
  planner.py, executor/.
• Idioms that make mutations visible: an expression class declares arg_types; a dialect nests its own
  Tokenizer, Parser and Generator; parsers key off dicts like FUNCTIONS and STATEMENT_PARSERS;
  generators off a TRANSFORMS dict and <node>_sql methods. Dialects are near-copies of one another,
  so the same method in a sibling dialect is the reference for what the mutated one should say.
• Trap: tests/fixtures/ holds .sql golden files (identity.sql, optimizer/*.sql) that the tests
  read and compare against. They are test data. Never edit them — a failing fixture means your fix is
  wrong, not that the fixture is.
• Tests: python -m pytest tests/test_optimizer.py -q. The optimizer file includes TPC-H and TPC-DS
  cases and is slow; select a single test with -k while iterating and run the file once at the end.

python-docx

• Source: src/docx/. Tests: tests/, mirroring the package (tests/oxml/, tests/text/,
  tests/parts/). features/ is a separate acceptance suite, not pytest — ignore it.
• The parts: the public objects in document.py, table.py, text/paragraph.py, text/run.py,
  section.py, shape.py; the XML layer in oxml/ (custom lxml element classes, declared with the
  descriptors in oxml/xmlchemy.py, value types in oxml/simpletypes.py); packaging in opc/ and
  parts/.
• Where behaviour lives: a public method on a docx/ object usually delegates straight to a CT_*
  element class in oxml/. If the public method looks right, the bug is one level down.
• Trap, and it is the main way to score zero here: the pytest configuration sets filterwarnings to
  error. Any warning your change newly raises — a deprecation, a resource warning — is a test
  failure across unrelated tests. Do not introduce a new warning, and do not silence one.
• Tests: python -m pytest tests/test_table.py -q, or the mirrored path for the module you touched.

python-pptx

• Source: src/pptx/. Tests: tests/, mirroring the package. features/ is behave, not pytest.
• The parts: presentation.py, package.py, shapes/, chart/, text/, table.py, util.py,
  document properties under opc/, and the same oxml/ element-class layer as python-docx, with the
  same xmlchemy descriptors.
• Same trap: filterwarnings is error. A new warning fails unrelated tests.
• Tests: python -m pytest tests/test_<module>.py -q at the mirrored path.

cantools

• Source: src/cantools/. Tests: tests/, with data files in tests/files/.
• The parts: database/can/ (database.py, message.py, signal.py, node.py, bus.py) and
  database/can/formats/ (dbc.py, kcd.py, sym.py, arxml/); the command line in subparsers/
  (list.py, dump/, decode.py, convert.py, plot.py, monitor.py, generate_c_source.py).
• Where behaviour lives: the subcommand modules print. Their tests compare exact stdout, so
  whitespace, column order and line breaks are the contract — matching the expected output character for
  character is the fix, not an approximation of it.
• Note: the pytest configuration lives in tox.ini and sets addopts = -v, so runs are verbose by
  default. Pass -q yourself to keep the output small.
• Tests: python -m pytest tests/test_list.py -q, tests/test_database.py (large, slow — use -k),
  tests/test_command_line.py, tests/test_convert.py.

pygments

• Source: pygments/ at the top level. Tests: tests/.
• The parts: lexer.py (Lexer, RegexLexer, DelegatingLexer, ExtendedRegexLexer, and the helpers
  bygroups, using, include, default), lexers/ with one module per language family and the
  generated lexers/_mapping.py, plus formatters/, filters/, styles/, token.py, util.py.
• Where behaviour lives: a lexer is mostly a tokens dict of regex-to-token rules, and the mutation is
  usually inside one rule or in a constructor. A DelegatingLexer subclass passes its two lexers up to
  super().__init__ in a fixed order — root lexer, then language lexer — and swapping them is a classic
  injected edit.
• Trap, and it is the main way to score zero here: most of the suite is golden-file snapshot tests
  generated by tests/conftest.py from tests/snippets/ and tests/examplefiles/. A change to a shared
  lexer or to lexer.py can fail hundreds of them at once. The suite also offers a flag that rewrites
  those golden files — never use it. It edits the tests and scores zero.
• Tests: python -m pytest tests/test_basic_api.py -q for the core, and the snippet directory named
  after the lexer alias you touched. python -m pytest tests/ -q is large; reach for it only with budget
  left.

astroid

• Source: astroid/ at the top level. Tests: tests/, mostly flat (test_inference.py, test_nodes.py,
  test_builder.py, test_scoped_nodes.py, test_protocols.py, test_manager.py, test_modutils.py),
  with tests/brain/ for the brain modules and tests/testdata/ for fixtures.
• What it is: the AST and static-inference engine behind pylint. The parts: nodes/ (node_classes.py,
  scoped_nodes/, node_ng.py, as_string.py), brain/ with one brain_<library>.py per third-party
  package it models, plus builder.py and rebuilder.py (source to AST), bases.py, protocols.py,
  inference_tip.py, manager.py, modutils.py, helpers.py, objects.py, arguments.py.
• Idioms that make a mutation visible: inference is generator-based — infer() and _infer() yield
  results and yield the Uninferable sentinel when they cannot decide. A mutation that turns a yield
  into a return, drops the Uninferable branch, or reverses a generator's order looks harmless and
  breaks many tests. Node classes declare their children in _astroid_fields / _other_fields;
  compare a node against its siblings in the same file.
• Two traps here, and both cost the whole task. The pytest configuration sets filterwarnings to
  error, so any new warning fails unrelated tests. It also sets xfail_strict, which means a test
  marked expected-to-fail that starts passing is a failure — so fixing more than the issue asks
  for can break tests by making them succeed. Stay minimal here more than anywhere else.
• Tests: python -m pytest tests/test_inference.py -q, or the file matching the area you touched.

oauthlib

• Source: oauthlib/ at the top level. Tests: tests/, mirroring it exactly.
• The parts: common.py, uri_validate.py, signals.py, and three protocol trees — oauth1/,
  oauth2/ (rfc6749/, rfc8628/) and openid/. Inside oauth2/rfc6749/: grant_types/
  (authorization_code, implicit, client_credentials, refresh_token,
  resource_owner_password_credentials, all on base.py), clients/, endpoints/, tokens.py,
  parameters.py, errors.py, request_validator.py, utils.py.
• Where behaviour lives: this is a protocol library, so the contract is exact strings — parameter names,
  URL encoding, header spelling, error codes. The grant types are near-copies of each other on a shared
  base, which makes base.py and the sibling grant the reference for a mutated one. errors.py
  defines the error classes with their exact error slugs and status codes; a changed constant there
  surfaces as a wrong error string far away.
• Note: there is no pytest configuration section, so runs are plain. Mirror the source path into
  tests/ — oauthlib/oauth2/rfc6749/grant_types/x.py is covered by
  tests/oauth2/rfc6749/grant_types/test_x.py.
• Tests: python -m pytest tests/oauth2 -q, or the mirrored file for the module you changed.

If it is none of these

The shape still holds: source at the top level or under src/, tests mirroring it under tests/, data
and golden files inside the test tree and never to be edited. Find the test file whose name matches your
module and run that first.