refactor(git): use SDK protocol classes for node kinds in tasks

Merged
OG
4mo ago
+72-30

Note: The below text covers the reasoning around this, but a review of the actual code changes might be quicker :)

  • Swap out text strings for kinds -> protocols
  • Remove resolved typing violations (mypy & ty)
  • Split apart remaining ty rules for infrahub.git so that it's easier to solve specific violations in the future

Why

Fetching a node with a kind string makes the SDK return a bare InfrahubNode, so every attribute access on it goes through InfrahubNode.__getattr__ and resolves to Attribute | RelationshipManager[InfrahubNode] | RelatedNode[InfrahubNode]. Nothing useful can be narrowed from that union, which is why backend/infrahub/git/ suppresses unresolved-attribute and invalid-assignment for the whole module. 57 of the module's 133 suppressed ty diagnostics come from that one union, 29 of them on .value.

The module already knows the answer: it imports infrahub_sdk.protocols in three files and 16 call sites already pass a protocol class. This PR finishes that migration for tasks.py.

Goal: convert the remaining kind-string node fetches in tasks.py, and restructure the module's ty override so progress is visible per file instead of hidden behind one module-wide glob.

Non-goals: integrator.py's 16 equivalent sites (same edit, separate PR), and the remaining union errors that come from InfrahubNode-annotated parameters rather than from the kind argument.

What changed

  • 14 call sites in tasks.py now pass the generated protocol class instead of InfrahubKind.*: CoreGenericRepository (5), CoreProposedChange (2), CoreReadOnlyRepository (2), CoreFileCheck (2), CoreRepositoryValidator, CoreUserValidator, CoreStandardCheck. tasks.py drops from 73 to 48 suppressed ty diagnostics; the module from 133 to 108.

  • The backend/infrahub/git/** ty override is split per rule, per file, matching the convention already used for core/** and graphql/**. Only unused-type-ignore-comment stays module-wide, since that is a genuine mypy/ty clash rather than debt.

    rulefiles still listed
    invalid-argument-typebase, integrator, repository, tasks, utils
    invalid-assignmentintegrator, tasks, utils
    unresolved-attributeintegrator, tasks, utils
    no-matching-overloadbase, integrator, repository, tasks
    invalid-return-typebase

    invalid-await and possibly-missing-attribute are deleted outright: neither reports anything in the module. The split immediately enforces all five rules for sync.py, models.py, directory.py, worktree.py, graphql_queries.py, constants.py, __init__.py and the closure_builder/ and fingerprint/ subpackages, so no new violation can slip into any of them.

  • Three stale mypy disable_error_code entries removed: attr-defined and index on infrahub.git.base, return-value on infrahub.git.repository. These no longer match any reported error and were already stale before this change. infrahub.git_credential.askpass was measured too and genuinely needs all three of its entries, so it is untouched.

What stayed the same: behaviour. get, filters and create all declare kind: str | type[SchemaType], and a class is resolved through _get_schema_name (python_sdk/infrahub_sdk/schema/__init__.py:263), which returns __name__. Every converted constant's value is exactly the class name that replaced it, e.g. InfrahubKind.FILECHECK == "CoreFileCheck", so the identical string reaches the schema lookup as before. The typing gain comes purely from overload selection: kind: type[SchemaType] returns SchemaType, and the overload used when raise_when_missing is omitted returns a non-Optional node, so no new narrowing is introduced.

How to review

pyproject.toml is the part worth reading carefully; the tasks.py diff is 14 one-token substitutions plus 5 added imports, and formatting collapsed a few calls that now fit on one line.

How to test

uv run ty check .                                  # All checks passed!
uv run invoke backend.mypy                         # Success: no issues found in 1601 source files
uv run invoke backend.ruff
uv run pytest backend/tests/unit/git/ --no-cov -q  # 178 passed

Integration coverage for these code paths lives in backend/tests/integration/git/ and needs the testcontainers stack, so CI is the gate there.

Impact & rollout

  • Backward compatibility: none affected; the same kind string reaches the SDK.
  • Performance: none.
  • Config/env changes: none.
  • Deployment notes: safe to deploy.

Follow-ups

  • integrator.py: the same conversion at 16 sites (CoreTransformJinja2, CoreGraphQLQuery, CoreArtifactDefinition, CoreGroup, CoreGeneratorDefinition, CoreCheckDefinition, CoreTransformPython).
  • After that, 28 of tasks.py's remaining 48 diagnostics are still the same SDK union, but sourced from parameters annotated InfrahubNode and from .peer traversals. Those need the surrounding signatures tightened, which is judgement work rather than substitution.
  • The 10 client.schema.get(kind=...) sites can also take a protocol class, but there is no typing payoff since they return a schema rather than a node.

Checklist

  • Tests added/updated
  • Changelog entry added
  • External docs updated (if user-facing or ops-facing change)
  • Internal .md docs updated (internal knowledge and AI code tools knowledge)
  • I have reviewed AI generated content