This document contains critical information about working with this codebase. Follow these guidelines precisely.
-
Package Management
- ONLY use uv, NEVER pip
- Installation:
uv add <package> - Running tools:
uv run <tool> - Upgrading:
uv lock --upgrade-package <package> - FORBIDDEN:
uv pip install,@latestsyntax
-
Code Quality
- Type hints required for all code
- Public APIs must have docstrings
- Functions must be focused and small
- Follow existing patterns exactly
- Line length: 120 chars maximum
- FORBIDDEN: imports inside functions. THEY SHOULD BE AT THE TOP OF THE FILE.
-
Testing Requirements
- Framework:
uv run --frozen pytest - Async testing: use anyio, not asyncio
- Do not use
Testprefixed classes, use functions - Coverage: test edge cases and errors
- New features require tests
- Bug fixes require regression tests
- IMPORTANT: The
tests/client/test_client.pyis the most well designed test file. Follow its patterns. - IMPORTANT: Be minimal, and focus on E2E tests: Use the
mcp.client.Clientwhenever possible. - IMPORTANT: Before pushing, verify 100% branch coverage on changed files by running
uv run --frozen pytest -x(coverage is configured inpyproject.tomlwithfail_under = 100andbranch = true). If any branch is uncovered, add a test for it before pushing. - Avoid
anyio.sleep()with a fixed duration to wait for async operations. Instead:- Use
anyio.Event— set it in the callback/handler,await event.wait()in the test - For stream messages, use
await stream.receive()instead ofsleep()+receive_nowait() - Exception:
sleep()is appropriate when testing time-based features (e.g., timeouts)
- Use
- Wrap indefinite waits (
event.wait(),stream.receive()) inanyio.fail_after(5)to prevent hangs
- Framework:
Test files mirror the source tree: src/mcp/client/streamable_http.py → tests/client/test_streamable_http.py
Add tests to the existing file for that module.
-
For commits fixing bugs or adding features based on user reports add:
git commit --trailer "Reported-by:<name>"Where
<name>is the name of the user. -
For commits related to a Github issue, add
git commit --trailer "Github-Issue:#<number>" -
NEVER ever mention a
co-authored-byor similar aspects. In particular, never mention the tool used to create the commit message or PR.
-
Create a detailed message of what changed. Focus on the high level description of the problem it tries to solve, and how it is solved. Don't go into the specifics of the code unless it adds clarity.
-
NEVER ever mention a
co-authored-byor similar aspects. In particular, never mention the tool used to create the commit message or PR.
When making breaking changes, document them in docs/migration.md. Include:
- What changed
- Why it changed
- How to migrate existing code
Search for related sections in the migration guide and group related changes together rather than adding new standalone sections.
-
Ruff
- Format:
uv run --frozen ruff format . - Check:
uv run --frozen ruff check . - Fix:
uv run --frozen ruff check . --fix - Critical issues:
- Line length (88 chars)
- Import sorting (I001)
- Unused imports
- Line wrapping:
- Strings: use parentheses
- Function calls: multi-line with proper indent
- Imports: try to use a single line
- Format:
-
Type Checking
- Tool:
uv run --frozen pyright - Requirements:
- Type narrowing for strings
- Version warnings can be ignored if checks pass
- Tool:
-
Pre-commit
- Config:
.pre-commit-config.yaml - Runs: on git commit
- Tools: Prettier (YAML/JSON), Ruff (Python)
- Ruff updates:
- Check PyPI versions
- Update config rev
- Commit config first
- Config:
scripts/sync_snippets.py replaces the content between
<!-- snippet-source ... --> / <!-- /snippet-source --> markers with code
from the referenced source file. The source file is the source of truth —
never edit synced content directly in the target.
To sync only part of a file, append #RegionName to the path. Regions are
delimited in source files by # region Name / # endregion Name markers.
Each example lives in a named function (returning -> None) wrapping a region.
Names follow ClassName_methodName_variant for methods, functionName_variant
for standalone functions, or module_overview for module docstrings. Pick a
descriptive variant suffix (_basic, _sync/_async, _with_context, etc.):
def MyClass_do_thing_basic(obj: MyClass) -> None:
# region MyClass_do_thing_basic
result = obj.do_thing("arg")
print(result)
# endregion MyClass_do_thing_basicFunction parameters supply typed dependencies the example needs but does not create
(e.g., server: MCPServer); module-level stubs are only for truly undefined references
(e.g., async def fetch_data() -> str: ...).
NEVER put # type: ignore, # pyright: ignore, or # noqa inside a region — these
sync verbatim into the target. Restructure the code to address the errors instead.
After editing an example file, run uv run --frozen pyright to verify types, then
uv run python scripts/sync_snippets.py to sync. Use --check to verify without
modifying files.
Code examples in README.v2.md and docs/**/*.md use explicit paths relative
to the repo root:
<!-- snippet-source examples/snippets/servers/basic_tool.py -->
```python
# replaced by sync script
```
<!-- /snippet-source -->Code examples in src/ docstrings use companion files in
examples/snippets/docstrings/, mirroring the source tree
(src/mcp/foo/bar.py → examples/snippets/docstrings/mcp/foo/bar.py).
Companion files are standalone scripts (not packages) starting with
from __future__ import annotations.
Docstrings use path-less #Region markers (only supported in src/ files):
Example:
<!-- snippet-source #MyClass_do_thing_basic -->
```python
result = obj.do_thing("arg")
print(result)
```
<!-- /snippet-source -->
-
CI Failures
- Fix order:
- Formatting
- Type errors
- Linting
- Type errors:
- Get full line context
- Check Optional types
- Add type narrowing
- Verify function signatures
- Fix order:
-
Common Issues
- Line length:
- Break strings with parentheses
- Multi-line function calls
- Split imports
- Types:
- Add None checks
- Narrow string types
- Match existing patterns
- Line length:
-
Best Practices
- Check git status before commits
- Run formatters before type checks
- Keep changes minimal
- Follow existing patterns
- Document public APIs
- Test thoroughly
- Always use
logger.exception()instead oflogger.error()when catching exceptions- Don't include the exception in the message:
logger.exception("Failed")notlogger.exception(f"Failed: {e}")
- Don't include the exception in the message:
- Catch specific exceptions where possible:
- File ops:
except (OSError, PermissionError): - JSON:
except json.JSONDecodeError: - Network:
except (ConnectionError, TimeoutError):
- File ops:
- FORBIDDEN
except Exception:- unless in top-level handlers