Fix unnecessary parentheses in indexed assignment RHS

[ParamChordiya]

Summary

Fixes #4349.

When an assignment target contains brackets (e.g. indexed access like x[key] = expr), Black would wrap the right-hand side expression in unnecessary parentheses when the line was too long.

Before:

dictionary_of_arrays["long_key_name_for_the_example"][
    very_long_index_name, index_zero
] = (10 - 5)

After:

dictionary_of_arrays["long_key_name_for_the_example"][
    very_long_index_name, index_zero
] = 10 - 5

Details

The root cause is in can_omit_invisible_parens() in src/black/lines.py. This function decides whether optional parentheses around the RHS can be safely removed. It returned False for simple expressions with operators (like 10 - 5) because they don't start or end with brackets — falling through to the default return False.

The fix adds a targeted check: when the RHS head ends with = and the LHS contains brackets (indicating an indexed/subscript assignment), and the body expression is short enough, allow omitting the optional parens. The downstream _prefer_split_rhs_oop_over_rhs() still makes the final decision on whether the resulting layout is better.

The line_length // 2 threshold ensures we only omit parens for genuinely short RHS expressions — long expressions that need their own line splitting still get wrapped as before.

Test plan

• Added test cases in tests/data/cases/prefer_rhs_split.py (idempotent formatting)
• Added input→output test case in tests/data/cases/prefer_rhs_split_reformatted.py
• All 209 existing formatting tests pass
• Verified fix against the exact reproducer from Unnecessary parentheses added to expression in indexed assignment #4349

[github-actions]

diff-shades results comparing this PR (0ce601c) to main (e34bb1b):

--preview style: no changes

--stable style: no changes

---

What is this? | Workflow run | diff-shades documentation

Didn't realize the changes to Black and in shades

[cobaltt7]

Almost missed this:
There's a few side-effects here that don't look like an improvement to me

Black would previously prefer to wrap long lines in parentheses to get it under the line length:

# -l 80
middle.children[1].prefix = (
    middle.children[0].prefix + middle.children[1].prefix
)

With this change, Black always prefers to split lines instead of wrap them, which isn't an improvement

# -l 80
middle.children[1].prefix = middle.children[0].prefix + middle.children[
    1
].prefix

[ParamChordiya]

Hey @cobaltt7 , thanks for catching that, really appreciate the thorough review. You were right, and it took a fair bit of digging to get to the bottom of it.

What was happening

The regression you spotted (middle.children[1].prefix = ... splitting mid-chain at the subscript instead of paren-wrapping the whole RHS) was happening because _prefer_split_rhs_oop_over_rhs in linegen.py was choosing the OOP (omit-optional-parens) split path even when the result was ugly. The OOP split would break at middle.children[ and put the index 1 on its own line, which is clearly wrong for a chained attribute access.
To fix it, I added a guard in _prefer_split_rhs_oop_over_rhs that says: if the OOP tail starts with ] followed by a . (meaning we'd be splitting inside a subscript access chain like obj[idx].attr), and the = is in the head (so the split is on the RHS, not the LHS), prefer paren-wrapping instead.

The first round of issues
After gating this guard behind Preview.fix_unnecessary_parens_in_indexed_assignment (to avoid affecting stable mode), the diff-shades corpus check was clean, but "Format ourselves" started failing on lint and the uvloop platforms. Turned out the guard was slightly too broad the condition len(rhs_oop.tail.leaves) > 1 was firing for list literal assignments like expected: list[Path] = [items] because the tail had ] plus a NEWLINE token (count=2, not just ] alone). That caused those list assignments to get incorrectly paren-wrapped. Fixed by tightening to rhs_oop.tail.leaves[1].type == token.DOT only suppress OOP when the character immediately after ] is actually a dot (attribute access).

The second bug, the trickier one
Even after that, "Format ourselves" was still failing because init.py kept getting reformatted. Traced it down to a separate bug in can_omit_invisible_parens in lines.py. The indexed-assignment early-return block checks whether there are brackets in the LHS: any(leaf.type in BRACKETS for leaf in rhs.head.leaves[:-2])
The problem: Python's AST wraps multi-target assignments like root, method = expr with invisible parentheses — these show up as empty-value LPAR/RPAR tokens in the leaf list. Those invisible tokens were matching the bracket condition and causing the indexed-assignment OOP path to fire for completely unrelated tuple assignments. That led to the outer parens being removed from the ternary in init.py, which is stable-idempotent but not unstable-idempotent, creating a formatting loop.

The fix was a one-character addition: and leaf.value, only count a bracket if it has an actual visible value, which correctly excludes the implicit/invisible tokens that the parser inserts.

How we verified

• All 211 format tests pass, including the prefer_rhs_split regression cases added in the original PR
• test_piping passes (reads init.py in stable mode and expects idempotency)
• python -m black --check src/ tests/ with the project's own unstable = true config reports 57 files unchanged
• init.py is idempotent under both stable and unstable mode now

The net change across both files is two lines, but it took some real archaeology to get there, the invisible-bracket issue in particular was subtle since nothing in the usual formatting output hinted that those tokens even existed.

[cobaltt7]

Please check the shades changes - it's still splitting within function calls.

[cobaltt7]

Hey @ParamChordiya, are you still updating this PR?

[ParamChordiya]

Hey @ParamChordiya, are you still updating this PR?

Hey @cobaltt7 sorry i got caught up with something else and lost track of this PR. let me circle back soon.

[ParamChordiya]

Hey @cobaltt7, back on this — thanks for your patience. You were right that the earlier approach was still too aggressive, so I reworked the heuristic instead of patching around it.

What was causing the bad shades changes

The check in can_omit_invisible_parens had an escape hatch that allowed omitting the parens whenever the RHS contained any bracket, even when the RHS didn't fit on one line. The splitter would then break at the last bracket inside the expression — which is exactly the "splitting within function calls" you flagged (ternaries like kwargs["empty_label"] = kwargs.get(...) if ... else None, and method chains like _get_totp(...).generate(...).decode()). It was also matching things that aren't indexed assignments at all: parenthesized tuple targets ((self.next,) = ...) and annotated assignments (self.args: list[str] = ...), because it only looked for "any bracket before the =".

The new approach

The paren-omission path now only fires when all of these hold:

• the target genuinely ends in a subscript — a visible ] immediately before the =
• it isn't an annotated assignment (so the ] of list[str] doesn't count)
• the LHS is too long to fit on one line, so Black is forced to split at the subscript brackets regardless
• the RHS fits on the resulting ] = expr tail line

In other words, it only removes the parens in the exact #4349 situation: the line has to split at the LHS anyway, and the parens on the tail are pure noise. When the LHS fits, paren-wrapping the RHS stays the preferred style, untouched. With the bad candidates eliminated at the source, the subscript-chain guard I'd added to _prefer_split_rhs_oop_over_rhs became dead code, so it's removed — the diff is smaller than before.

How I verified

• I reconstructed every hunk from the last diff-shades report (all 88 hunks / 758 changes across the 9 projects) and ran the new build over them: zero changes remain, so the only diff-shades delta left should be the intended Unnecessary parentheses added to expression in indexed assignment #4349 shape.
• All format tests pass, and I added regression cases for each failure mode you surfaced (ternary RHS, method chains, tuple targets, annotated assignments, and a too-long RHS that should still wrap).
• Black's own sources stay clean and idempotent under stable, preview, and unstable modes.
• A generated sweep of ~1000 assignment variants (slices, augmented/chained/annotated targets, walrus, lambdas, comments, wide chars) shows no stable-vs-preview differences outside the intended improvement, and no idempotency issues.

Also merged main and resolved the conflicts, so the branch is mergeable again. Would appreciate another look when you get a chance!

[cobaltt7]

Thanks!