Implementation Standards

Layer-Specific Rules

Before writing any code, identify the layer you're editing and apply its rules.

API Models (signnow_client/models/)

• ✅ Pydantic BaseModel, Field, validators, type aliases, __all__ exports.
• ❌ Imports from sn_mcp_server.* or signnow_client/client*.py. Any business logic.
• Pure data definitions only — mirror the SignNow REST API schema.

API Client (signnow_client/client_*.py)

• ✅ Use self._get(), self._post(), self._put(), self._post_multipart() from client_base.py. Import models from signnow_client/models/. Use exceptions from signnow_client/exceptions.py.
• ❌ Imports from sn_mcp_server.* (upward import). Direct httpx calls. Starlette imports. MCP logic.
• Each method takes token: str and passes Authorization: Bearer {token}.

Tool Response Models (sn_mcp_server/tools/models.py)

• ✅ Pydantic BaseModel, Field. Type references from signnow_client/models/.
• ❌ Raw API passthrough. Fields the agent won't act on. Empty lists, null metadata, internal IDs.
• Every field must justify: "Does the agent need this for its next decision?"

Tool Business Logic (sn_mcp_server/tools/<feature>.py)

• ✅ Import SignNowAPIClient, tools/models.py. Pure logic, data transformation.
• ❌ Import Starlette. Import from other tools/<feature>.py. Resolve tokens. Module-level mutable state. Generic errors.
• Functions receive (client: SignNowAPIClient, token: str, ...) → return curated DTOs.

Tool Orchestrator (sn_mcp_server/tools/signnow.py)

• ✅ Call _get_token_and_client(), delegate to tools/<feature>.py, use Annotated[..., Field(...)] for params.
• ❌ Business logic. Complex conditionals. Data transformation.
• Each tool function: resolve token → call business logic → return result. Docstrings = MCP tool descriptions.

Auth (sn_mcp_server/auth.py, token_provider.py)

• ✅ Import signnow_client, config.
• ❌ Import from tools layer. Expose tokens/secrets in logs or errors.

Transport (sn_mcp_server/app.py, cli.py)

• ✅ Import server.py, config.py, auth.py.
• ❌ Import tools or API client directly. Never touch business logic.

Coding Standards

• Language: English only — code, comments, docstrings, docs.
• Formatting: Black, line-length 200. Ruff for linting.
• Typing: Strict mypy. No Any unless wrapping untyped externals. Use str | None over Optional[str]. Use Annotated[..., Field(...)] for MCP tool parameters.
• Docstrings: Required for all public functions/classes. Tool docstrings = MCP descriptions — concise, actionable.
• Exports: __all__ in model files. One logical unit per file.
• DRY: Extract to tools/utils.py or signnow_client/utils.py at 3+ usages. Three similar lines are better than a premature abstraction.

Critical Gotcha: Transport Leakage

# ✅ Token is a parameter, no transport awareness
def _get_document(client: SignNowAPIClient, token: str, document_id: str) -> DocumentResponse:
    ...

# ❌ Importing Starlette, accessing headers — FORBIDDEN in tools
from starlette.requests import Request
def _get_document(request: Request, document_id: str) -> DocumentResponse:
    token = request.headers["authorization"]

Critical Gotcha: Secrets in Output

# ✅ Masked for diagnostics
from sn_mcp_server.config import _mask_secret_value
logger.info(f"Using token: {_mask_secret_value(token)}")

# ❌ Raw secret — FORBIDDEN
logger.info(f"Using token: {token}")
raise ValueError(f"Auth failed with token {token}")

Bug Fix Protocol

When fixing a documented bug:

1. Fix the code in source files.
2. Verify lint/type checks pass.
3. Testability analysis: Can the fix be verified with unit tests?

• ✅ Testable (logic, transformation, errors, validation): produce a prompt for the Tester agent with bug name, affected files, changes made, and what to assert.
• ❌ Not testable (transport, OAuth2 e2e, external API): state why and request manual verification with sn-mcp serve / sn-mcp http.

Pre-commit Rules (write it right the first time)

Pre-commit runs: ruff check → ruff format → black → mypy --strict. Every file you create or edit must pass all four before committing.

Imports

# 1. Always first line in every file
from __future__ import annotations

# 2. Order: stdlib → third-party → local. One blank line between groups.
# ruff (I001) enforces this automatically — just keep groups separate.
import json
from pathlib import Path
from typing import Any

import pytest
import respx

from signnow_client import SignNowAPIClient

Type annotations (ANN rules — all enforced)

# ✅ All public functions must have full annotations
def get_link(client: SignNowAPIClient, token: str, doc_id: str) -> str:
    ...

# ✅ Use X | Y syntax (UP007), not Optional or Union
def foo(value: str | None = None) -> dict[str, Any]:
    ...

# ✅ ANN401: avoid bare Any — use it only to wrap untyped externals
# If you must, add a comment explaining why
def _handle(data: dict[str, Any]) -> None:  # Any because SignNow API schema varies
    ...

# ❌ Missing return type → ruff ANN201 + mypy error
def bad(x: str):
    return x

`self` and `cls` are exempt — ANN101/ANN102 are in lint.ignore. Do NOT annotate them.

Test files specifically

# ✅ Import fixture types from their actual packages
from collections.abc import Callable, Generator
import respx
from signnow_client import SignNowAPIClient

# ✅ All test method params must be typed, return -> None
def test_something(
    self,
    client: SignNowAPIClient,
    mock_api: respx.MockRouter,
    token: str,
    load_fixture: Callable[[str], dict[str, Any]],
) -> None:
    ...

# ✅ Fixture return types
@pytest.fixture()
def mock_api() -> Generator[respx.MockRouter, None, None]:
    with respx.mock(...) as router:
        yield router

@pytest.fixture()
def load_fixture() -> Callable[[str], dict[str, Any]]:
    return _load_fixture

Hardcoded secrets in tests (S105/S106)

# ✅ Suppress with noqa — it's intentional fake test data
FAKE_TOKEN = "test-token"  # noqa: S105
cfg = SignNowConfig.model_construct(client_secret="test_secret")  # noqa: S106

json.loads return type (mypy no-any-return)

# ✅ Suppress the Any return from json.loads
return json.loads(path.read_text())  # type: ignore[no-any-return]

mypy in tests

[[tool.mypy.overrides]] disables misc for tests.* — so @pytest.fixture() decorators won't trigger mypy. No action needed; just keep the full type annotations above.

Verification Checklist

Complete ALL before declaring done:

1. Static analysis (run on changed files):

• ruff check src/ tests/ 2>&1 — must pass
• ruff format --check src/ tests/ 2>&1 — must pass
• mypy src/ tests/ 2>&1 — must pass

1. Runtime validation (if business logic or API client changed):

• Create scripts/verify-fix.py → import and call your new function with mock data
• Execute it. If it crashes, fix and retry.
• Delete the script when passing.

1. Regression tests (if related tests exist):

• pytest tests/ -v 2>&1
• If fails, fix and retry until green.

Do not ask the user to test. YOU test.

Rules

• No tests (Tester agent handles them).
• No markdown docs unless explicitly asked.
• No plan references in comments (@see .plans/).
• No orphan tools — register in tools/__init__.py and document in README.md.
• Strictly follow AGENTS.md.