[fastapi] Handle callable class dependencies with __call__ method (FAST003)

[stakeswky]

Fixes #23526

Problem

When a class with a __call__ method (but no __init__) is used as a FastAPI dependency, FAST003 emits a false positive:

class Query:
    def __call__(self, thing_id: int):
        pass

@app.get("/things/{thing_id}")
async def read_thing(query: Annotated[str, Depends(Query)]): ...
# FAST003: Parameter `thing_id` appears in route path, but not in `read_thing` signature

Root Cause

In from_dependency_name, the ClassDefinition branch checked for Pydantic base models and __init__, but returned None (not Some(Self::Unknown)) when neither was found. This caused the dependency to be silently skipped in the caller, leaving the path parameter unmatched.

Fix

Two changes:

1. Fall back to __call__ when no __init__ is found. This correctly handles the callable instance pattern from FastAPI's docs, where an instance with __call__ is passed to Depends.

2. Return Some(Self::Unknown) instead of None when neither __init__ nor __call__ exists, so we conservatively suppress the diagnostic rather than emitting a false positive.

Tests

Added four new test cases:

• Callable class with __call__(self, thing_id) → no diagnostic ✓
• Class with both __init__(self, thing_id) and __call__ → no diagnostic (uses __init__) ✓
• Callable class where path param is NOT in __call__ → FAST003 emitted ✓
• Empty class (no __init__, no __call__) → no diagnostic (Unknown) ✓

[ntBre]

Thanks! I had a couple of small questions/suggestions and then realized I had a follow up question on the issue itself.

[astral-sh-bot]

ruff-ecosystem results

Linter (stable)

✅ ecosystem check detected no linter changes.

Linter (preview)

✅ ecosystem check detected no linter changes.

[stakeswky]

Thanks for the thorough review! I've reworked the approach based on your question on the issue.

After testing, the key insight is:

• Depends(Query) — FastAPI calls the class constructor, so params come from __init__
• Depends(Query()) — FastAPI calls the instance, so params come from __call__

The original fix was wrong to conflate both cases. The new approach:

1. Splits from_depends_call to handle both Name and Call expressions
2. from_dependency_name → class reference → uses __init__
3. from_dependency_instance → class instance → uses __call__
4. Shared class_params_from_method helper for both paths

The matches! nit no longer applies since the find() with OR is gone. The None → Unknown change is kept — when a class has no matching method, returning Unknown suppresses false positives (we know there's a dependency, just can't analyze it).

[ntBre]

I'm going to wait to confirm my understanding with the issue author before reviewing this again, but I think this aligns with my current understanding.

[stakeswky]

Addressed the review feedback:

• Reverted the return None → Some(Self::Unknown) change as suggested
• Rebased onto latest main

Re the approach: the issue discussion confirmed that __init__ for Depends(Query) and __call__ for Depends(Query()) is the right split, which is what the current code does via from_dependency_name and from_dependency_instance respectively.

[stakeswky]

Thanks for checking! Confirmed — the PR already handles the two cases separately:

• Depends(Query) (class) → from_dependency_name → resolves via __init__ (existing behavior, unchanged)
• Depends(Query()) (instance) → from_dependency_instance → resolves via __call__ (new behavior added by this PR)

So __call__ is only consulted when FastAPI receives an already-constructed instance. Also just pushed a snapshot format update to fix the failing cargo tests (the insta snapshot was in legacy format).

[stakeswky]

Thanks for confirming on the issue! That aligns with the current implementation — the PR only checks __call__ for the Depends(Query()) pattern (call expression), while Depends(Query) (name reference) continues to use __init__ as before. The two paths are split into from_dependency_instance (for call expressions → __call__) and from_dependency_name (for name references → __init__). Let me know if you'd like any adjustments.

[ntBre]

Thanks! There was a .snap.new file mistakenly added, so I pushed a commit removing that and also threw in a couple of non-Annotated test cases while I was here.