From e660a3935e1d91b27347ece794cd120fe8b06330 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Anselm=20Sch=C3=BCler?= <mail@anselmschueler.com>
Date: Thu, 26 May 2022 17:50:24 +0200
Subject: [PATCH] [RFC 0126] Standardize magic properties

---
 rfcs/0126-standardize-magic-properties.md | 81 +++++++++++++++++++++++
 1 file changed, 81 insertions(+)
 create mode 100644 rfcs/0126-standardize-magic-properties.md

diff --git a/rfcs/0126-standardize-magic-properties.md b/rfcs/0126-standardize-magic-properties.md
new file mode 100644
index 000000000..e4288bb83
--- /dev/null
+++ b/rfcs/0126-standardize-magic-properties.md
@@ -0,0 +1,81 @@
+---
+feature: standardize-magic-properties
+start-date: 2022-05-26
+author: Anselm Schüler
+co-authors: None
+shepherd-team: None
+shepherd-leader: None
+related-issues: None
+---
+
+# Summary
+[summary]: #summary
+
+Rename attrset attribute names treated specially by Nix to start with a double underscore.
+Add a builtin function `guardMagic` that returns `null` for double underscore prefixed strings.
+
+# Motivation
+[motivation]: #motivation
+
+Nix treats many attributes specially with unique behaviours. Some of these are prefixed with double underscores to disambiguate them from everything else. But some aren’t: `recurseForDerivations` isn’t, `type`, when set to `"derivation"`, changes visual output.
+
+It would also be a good idea to avoid accidentally using reserved special attribute names. Therefore I suggest `guardMagic` that can be used in dynamic attribute names to avoid this.
+
+# Detailed design
+[design]: #detailed-design
+
+Properties treated specially by the Nix tools and Nix language when dealing with arbitrary are renamed to start with a double underscore. Instances where this is not necessary are `__functor` and `__toString`. `recurseForDerivations` is renamed `__recurseForDerivations` and `type` is renamed `__type`.
+
+A new builtin function is introduced, `builtins.guardMagic`.
+If called on a string, it returns the string if it does not start with a double underscore, and `null` otherwise. This can be used by semantically charged non-generic functions that update attrs dynamically to avoid unwanted interactions.
+
+Nixpkgs & Nix need to be updated to reflect this. Therefore, it may be advisable to have a grace period during which unprefixed magic properties are still supported, but warn.
+
+# Examples and Interactions
+[examples-and-interactions]: #examples-and-interactions
+
+```nix
+rec {
+  set-val-to-null = k: attrs: attrs // {
+    ${guardMagic k} = null;
+  };
+  x = set-val-to-null "__functor" { __functor = _: _: { }; };
+  y = set-val-to-null "y" { __functor = _: _: { }; };
+}
+```
+evaluates to
+```nix
+{
+  set-val-to-null = «lambda»;
+  x = { __functor = «lambda»; };
+  y = { y = null; __functor = «lambda»; };
+}
+```
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+- This change is backwards-incompatible
+- This requires judging if certain properties count as magic attributes
+
+# Alternatives
+[alternatives]: #alternatives
+
+- `guardMagic` could be introduced on its own, thereby complicating its implementation
+- These could be left alone
+
+# Unresolved questions
+[unresolved]: #unresolved-questions
+
+I honestly don’t have a full list of these magic names, I only listed the ones I know about.
+
+It’s also unclear what exactly counts. I think attrsets passed to functions as de-facto named arguments shouldn’t count (e.g. for `derivation`). You could even argue the outputs of `listToAttrs` should be prefixed!
+
+The name of `guardMagic` might need some work.
+
+# Future work
+[future]: #future-work
+
+This eases introduction of any future magic names.
+
+It can also be assumed the implementation of `guardMagic` won’t need to be updated if this standard naming convention is adhered to.