**Design only — no implementation.** Proposes `GSharp.Formatting` (library) + `gsfmt` (dotnet tool CLI), consumed **in-process** by the language server and as a fail-soft post-pass by cs2gs. ## Why this is worth doing regardless of the metric G# has **three disagreeing formatters**: | authority | indent | wraps? | |---|---|---| | `GSharpPrinter.cs` (2416 lines) | 4-space | yes, ad-hoc, 120 cols | | `FormattingEngine.cs` (265 lines) — *the one advertised to editors* | **2-space** | no | | hand-written `.gs` | 4-space | n/a | `textDocument/rangeFormatting` and `onTypeFormatting` are registered as **no-ops and deliberately not advertised** (`ServerCapabilitiesFactory.cs:57-61`, #1660) — not because of LSP plumbing, but because a token-stream engine has no syntax tree and cannot produce a correct partial result. ## Honest scope — a formatter fixes about two thirds of the long lines, not all Measured over run 33943018295 (3806 `.gs` files, 630 long lines): | category | count | share | fixable by a formatter | |---|---:|---:|---| | wrappable code | 537 | 85% | **yes — but this is an upper bound**, realistically 400–480 | | one string literal wider than any budget | 61 | 10% | **no** | | `///` doc comments | 32 | 5% | **no**, and it shouldn't | The 93 are **cs2gs defects, not formatter work**: `RenderLiteral` already emits backtick raw strings but bails when the value contains a backtick (one fixture becomes a 5122-char one-liner), and cs2gs collapses an 8-line `<remarks>` into a 424-char `///`. **Those two PRs are cheaper than the whole formatter** — Phase 9 has the best effort-to-metric ratio in the plan and could go first. Shape breakdown of the 537: member chains 201, argument lists 116, object/collection literals 105, `+` chains 68, `&&`/`||` **5** — low precisely *because* `RenderWrapped` already handles that one family. Evidence the rest are unimplemented, not intractable. ## The hardest thing the investigation found **G# is newline-insensitive except at nine parser sites** (`Expressions.cs:292/457/473/830`, `Creation.cs:728/842/1566`, `Literals.cs:1049`, `Patterns.cs:323/427`, `Statements.cs:1429/1493/1511`). That knowledge exists **only in prose today**. A wrapper that breaks at one of them *silently changes program meaning*. Phase 1 turns it into `SyntaxFacts.IsBreakLegalBetween` with a test asserting every parser call site is covered — the single most valuable artifact in the design. ## Key decisions - **Wadler `Doc` algebra, not Roslyn-style rules.** Roslyn's formatter *preserves author line breaks and never wraps* — which is why `dotnet format` has never shortened a long line. Adopting it would leave the 537 exactly where they are. - **The LS references the library in-process, not the tool.** This contradicts the "standalone tool the LS consumes" framing deliberately: the LS already formats in-process and hosts all of `GSharp.Core`. `gopls` links `go/format`; it does not exec `gofmt`. `gsfmt` stays standalone for CI, hooks, and cs2gs. - **Zero configuration.** The in-repo 2-space/4-space split is the proof of what options cost — `FormattingOptions.tabSize` was threaded through and produced two canonical forms. - **Lint stays out**, in ADR-0169's analyzer framework. Lints need semantics (#892's `var`→`let` needs data flow) and *want* to change tokens; a formatter must never change tokens. #892 retargeted. - **C# first, migrated later** — but in `guard_apps` from Phase 1, because gsfmt enters the cs2gs corpus **by existing**, the hazard that has taken the gate down four times. ## Verified against the tree, not assumed `SyntaxToken` carries **no trivia** and `Parser.cs:150-160` discards whitespace/comments — but `SyntaxTree.ParseTokens` returns a **lossless** token stream with verbatim text and absolute positions. So trivia is recoverable by position-joining **without touching 156 node classes**. Adding real trivia to `SyntaxToken` is the architecturally right shape and is documented as the escalation path if that proves fragile. ~12 PRs across 9 phases; phases 1–3 have no consumer and cannot regress anything. Phase 7 carries the churn (454 print call sites, ~4756 assertions) and 7a measures it before 7b commits. Closes #916 by adoption. Refs #892, #1660, #3501, #3931. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

Uh oh!

Linter/formatter should autoconvert var to let when not mutating within scope #892

Description

Activity

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

Linter/formatter should autoconvert `var` to `let` when not mutating within scope #892