CASC Configuration Format (CASC)

Copy Markdown View Source

1. Overview

CASC is a hierarchical configuration language that combines:

  • Assignments (key = value, or key value — the operator is optional, §5.3) and blocks ({ ... }), which double as CASC's only map type (§5.4)
  • Dotted paths (a.b.c = 1) as sugar for nested blocks; key { ... } (no operator) is sugar for key = { ... } — all forms desugar to the same tree before merge (§5.4)
  • Imports (import "path/**/*.casc"), including consumer-defined import sources via URI scheme (import "vault://...", §9.3)
  • Variables (public @name, private @*name, §5.2), including block-scoped loop variables (§5.5)
  • Three brace-delimited sigils: variables (@{...}, §7.1), environment expansion (${...}, bash-like, §7.2), config references (%{...}, §7.3)
  • Extensible resolution (!{resolver:arg}, §7.4) and tagged values (!Name(...), §7.5) for consumer-defined vocabulary — unregistered use of either always fails loudly (§9.4)
  • Lists ([...]) and tuples ((...)) — tuples are fixed-arity and never merged, only replaced wholesale (§6.11, §8.3)
  • Native duration (5ms, 1h30m, §6.8) and byte-size (512MiB, 10GB, §6.9) literals — decimal (SI, powers of 10) vs. binary (IEC, powers of 2)
  • Secret keys (*password = ...) for consumer-side redaction (§4.3)
  • Merge control sigils (~key, +key, -key) overriding default merge behavior per key (§5.7, §8.4)
  • A loop construct (for ... [from ...] as ... { ... }) generating repeated structure from one or more parallel-iterated lists (§5.5)
  • One comment marker (# ...), deliberately shared with the version header and with disabling a statement (§3.2, §5.6)

This document defines CASC at the language/spec level; implementations may add conveniences but must not break the syntax described here.

Result values in this document are shown as native Elixir terms: maps use string keys (%{"key" => value}), CASC atoms become Elixir atoms (:info), CASC tuples become real Elixir tuples ({...}). Where a result has many uniform entries, it's shown in TOON instead, for density.

2. Mandatory header

Every CASC file must begin with a version header, as the first non-blank, non-comment statement:

#@version = <version>

The operator is optional here like everywhere else (§5.3); #@version = <version> is the recommended spelling. <version> should follow SemVer 2.0, e.g. 1, 1.0, 1.0.0, 2.1.3.

3. Lexical structure

3.1 Whitespace

Spaces, tabs, and newlines separate tokens. Newlines end statements, except where a value spans multiple lines (triple-quoted strings, backslash continuation, §6.5).

3.2 Comments

# a line comment, runs to end of line

No multi-line comment syntax exists. # serves three related purposes, sharing one character deliberately — all three are "treat what follows differently, based on what comes right after #":

  1. A plain comment — # followed by whitespace or anything other than a letter, @, or *.
  2. The version header (§2) — #@version ..., only as the file's first non-blank statement.
  3. A disabled statement (§5.6) — # followed directly by a letter or *, anywhere else (#key = 1, #*key { ... }, #@name = 1).

This is a hard lexical rule: #TODO (no space) is a disabled-statement attempt, not a comment, and fails to parse unless TODO happens to be valid statement syntax. # TODO (space) is unambiguously a comment.

4. Identifiers and naming rules

4.1 Key names, atoms, and variable names

/^[a-zA-Z][a-zA-Z0-9_+-]*[?!]?$/ — mixed case, underscores, hyphens, digits, +, optionally ending ? or !.

Examples: logger, handleOtpReports, handle-otp-reports, enabled?, ready!, MyApp.

for, in, from, as (§5.5) are contextual keywords — special only inside a for statement, ordinary identifiers everywhere else. This differs from the six reserved words in §6.4 (nil, true, false, inf, +inf, -inf), which are reserved everywhere.

4.2 Quoted key segments

A key segment may be written as a string, for text that doesn't fit the identifier rule — spaces, punctuation, arbitrary content:

headers = {
  "Content-Type" = "application/json"
}

foo."bar baz".dronf = "fnord"

Double-quoted key segments support interpolation and escapes, same as double-quoted strings (§6.5). Single-quoted key segments are literal, same as single-quoted strings. Bare identifier keys cannot be interpolated (§4.1's rule has no room for @{}), so an interpolated key must be double-quoted.

4.3 Secret keys

Any key segment may be prefixed * to mark it secret: *password = "...", db.*password = "...", *db.password = "...". The stored value is unaffected; consumers should redact secret values on display (e.g. "[~~REDACTED~~]").

5. Statements

A CASC file is a sequence of statements: imports, variable declarations, assignments, blocks, loops, blank lines. Statements may be disabled (§5.6) and may carry a merge control sigil (§5.7) — both are prefix modifiers, not separate statement kinds.

5.1 Imports

import "path/to/file.casc"
import "config/{base,dev}/**.casc"
import "vault://secret/base"

  1. No scheme:// prefix — resolves relative to the current file. Brace ({a,b}) and glob (**.casc) patterns expand against the filesystem; matches load in lexicographic order.

  2. scheme:// prefix — dispatched entirely to the loader registered for that scheme (§9.3); expansion, ordering, and path meaning are that loader's to define. An unregistered scheme is a load-time error naming it (§9.4).

  3. Merge: later imports override earlier ones for the same path, per §8.

  4. A path may interpolate ${NAME} or ${NAME:default} (§7.2), which is how one document selects among several:

    import "env/${MIX_ENV:dev}.casc"

    Only ${...} works. An import is resolved while the document is parsed — the imported file's statements are spliced into the importer — so a reference needing the finished tree (%{...}, §7.3) cannot be available yet and is a load-time error. The environment is available, which is the same thing ${?NAME} reads (§7.2).

    Unset and empty are treated alike, as everywhere else. An unset variable without a default is an error rather than an empty segment, since a path that silently became env/.casc would import the wrong file or none.

5.2 Variable declarations

@name = "value"       # public
@*private = "value"   # private, file-local

Operator optional, per §5.3.

Visibility. A public @name belongs to the whole load, not to the file that wrote it. It is visible to the files that import that file (transitively), and to the files that file imports — so an entry document can declare a value before its imports and have every imported file see it:

# app.casc
@service = "checkout"
import "inc/telemetry.casc"   # sees @{service}

# inc/telemetry.casc
telemetry.service_name = "@{service}"

A private @*name is usable anywhere inside the file that declares it and nowhere else. It never reaches an importer, and it is never visible to a file it imports. Two files may each declare an unrelated @*name without colliding, and inside its own file a private declaration shadows a public one of the same name.

Because visibility is decided per file, and the last file to assign a public name wins, prefer @*name for anything a file only needs for itself — a public name is effectively part of the load's shared surface.

5.3 Assignments

= is the assignment operator, and optional:

foo = "bar"
foo "bar"

Both are fully equivalent; an implementation must accept both. = is recommended for readability, never required. : has no assignment role — it's reserved solely for the atom sigil (§6.4).

This is unambiguous because a key path only continues via a literal . (§5.4): once whitespace appears with no . following, the key is finished, and whatever comes next — =, {, or a value — can only be read one way.

5.4 Key paths and blocks

logger.console.colors.enabled = true

logger {
  level = info
  console { colors.enabled = true }
}

Dotted paths, nested blocks, and explicit = { ... } are fully equivalent — the same tree, not just similar. key { ... } is sugar for key = { ... }; combined with the optional operator, these all produce the identical structure:

foo { bar { baz = "dronf" } }
foo { bar.baz "dronf" }
foo = { bar = { baz = "dronf" } }
foo.bar.baz "dronf"

Result (identical for all four):

%{"foo" => %{"bar" => %{"baz" => "dronf"}}}

Desugaring to the canonical tree always happens before merge (§8) — a correctness requirement, not convenience: foo.bar.baz = 1 in one file and foo = { bar = { qux = 2 } } in another must merge into one foo.bar with both keys, regardless of surface form.

Allowing three equivalent forms — rather than picking one, as TOML does — is a deliberate readability tradeoff: a deep config often reads better as dotted paths, a config with many siblings often reads better as a block. Consistency across a codebase is left to tooling (a formatter), not enforced by the grammar.

Block keys may be identifier-shaped (§4.1) or quoted strings (§4.2) — so a block with quoted keys already covers what a separate "map" type would. There is no separate map syntax in CASC; a block in value position is a map, however its keys are written.

5.5 Loops

for <binding> [, <binding>]* [from <template-path>] as <destination-path> {
  <body>
}

Bindings, comma-separated, each one of:

  • @name — index binding: the current iteration number (0, 1, 2, ...). At most one per loop; not itself iterable, takes no in.
  • @name in <iterable> — element binding: the current element of <iterable> (typically @{...} or %{...}, resolving to a list).

At least one element binding is required — an index-only loop is a load-time error.

With more than one element binding, iteration is parallel (zipped), not a nested cartesian product: iteration n uses index n of every bound list. This favors the common "N parallel lists describe N generated things" shape over an explosion of every combination. All element-bound lists must share the same length — a mismatch is a load-time error naming each list and its actual length, never a silent truncation.

<destination-path> is an ordinary key path (§5.4), evaluated per iteration, and may use interpolated segments (§4.2) built from the bindings — any segment may be dynamic, not just a fixed trailing one.

from <template-path> (optional) — each generated destination starts as a copy of the value at <template-path>, with the loop body's statements applied as overrides on top (§8). <template-path> is a bare key path, resolved lazily against the fully-merged final tree, same timing as %{...} (§7.3).

Scope — loop-bound variables are ordinary @name variables (§7.1), scoped to the body: they shadow any outer variable of the same name for that scope only, and cease to exist once the loop ends.

Example:

@domains = ["us-east.example.com", "eu-west.example.com"]
@ports = [8443, 8444]

for @idx, @domain in @{domains}, @port in @{ports} as endpoints."domain-@{idx}" {
  url = "https://@{domain}:@{port}"
}

Result:

endpoints[2]{name,url}:
  domain-0,https://us-east.example.com:8443
  domain-1,https://eu-west.example.com:8444

(Mismatched @domains/@ports lengths would be a load-time error naming both lists, not two endpoints and a dropped port.)

Example, with from:

defaults.replica {
  cpu = 1
  memory_mb = 512
}

@instances = ["a", "b", "c"]
for @instance in @{instances} from defaults.replica as replicas."@{instance}" {
  cpu = 2
}

Result:

defaults.replica: {cpu: 1, memory_mb: 512}

replicas[3]{name,cpu,memory_mb}:
  a,2,512
  b,2,512
  c,2,512

5.6 Disabled statements

Prefixing a statement with # directly before its key (or before another leading sigil, §5.7) disables it entirely — syntactically valid, but produces nothing:

section {
  #ignored = 1
  #sub { remove = yes }
  kept = true
}

Result:

%{"section" => %{"kept" => true}}

ignored and sub are absent, not present-but-null. Distinct from a # comment (§3.2) because the rest of the line can still look like a valid statement.

5.7 Merge control sigils and sigil ordering

Blocks/maps deep-merge and lists replace by default (§8 has full rationale). Three sigils override this per key, in the same leading position as * and #:

  • ~key { ... } — force replace instead of merge, for a block/map.
  • +key = [...] — append to an existing list. Behaves like a plain assignment if key doesn't exist yet.
  • -key = [...] — remove matching elements from an existing list.
  • -key.path (bare, no value) — delete the path entirely, any type.

+/- (list form) apply only to lists — using either against a tuple is a load-time error (§8.3), since it would silently change the tuple's fixed arity.

Sigil ordering — #, then merge-control (~/+/-), then *, directly before the key, no separators:

~*database { host = "db.internal", password = "hunter2" }
#~*database { host = "db.internal", password = "hunter2" }

The first replaces (not merges) database wholesale and marks it secret; the second is the same statement, disabled. # always comes first when present — a disabled statement produces nothing, so nothing downstream has anything left to act on.

6. Values

6.1 Nil

nil

6.2 Booleans

true, false

6.3 Numbers

  • Integers: -17, 0, +99
  • Hex/octal/binary: 0xDEAD_BEEF, 0o755, 0b11010110
  • Floats: 3.1415, -0.01
  • Exponent: 5e+22, -2E-2
  • Digit separators: _ between digits
  • Infinity: inf, +inf, -inf

6.4 Atoms

Bare identifiers, same naming rule as keys:

level = info
remove = yes

Reserved words take precedence over the atom rule — nil, true, false, inf, +inf, -inf are always their literal meaning (§6.1–6.3), never bare atoms, even though they're valid identifiers under §4.1.

An atom may optionally carry a leading : (Ruby/Lisp/Elixir convention) — required to reach an atom named one of the six reserved words, optional everywhere else:

level = :info    ; identical to `level = info`
enabled = :true   ; the atom :true -- NOT the boolean
enabled = true    ; the boolean -- NOT an atom

: has no other meaning in CASC — not an assignment operator (§5.3), and appears nowhere else in the grammar.

6.5 Strings

  • Double-quoted — escapes (\n, \r, \t, \", \\, \uXXXX) and interpolation (§7).
  • Single-quoted — literal, no escapes, no interpolation.
  • Triple-quoted (""" ... """) — multi-line; the smallest common leading whitespace across non-empty lines is stripped from each line.
  • Backslash-continued — a bare value starting with \ at end-of-line joins onto the next line, dropping the line break.

Example:

motd = """
    Welcome to the service.
    Status: operational
    """

Result:

%{"motd" => "Welcome to the service.\nStatus: operational\n"}

6.6 Dates and times

ISO-8601-like: offset datetime (1979-05-27T07:32:00Z), local datetime (1979-05-27T07:32:00), local date (1979-05-27), local time (07:32:00), all with optional fractional seconds.

6.7 IP addresses

IPv4/IPv6, optionally with CIDR: 127.0.0.1/32, ::1/128.

6.8 Durations

A number immediately followed by a unit, no whitespace between. Units, lowercase: ns, us/µs, ms, s, m, h, d.

timeout = 500ms
cache_ttl = 1h30m

A single-unit literal may be fractional (1.5h); a compound literal (multiple units concatenated) requires each component to be a plain integer, strictly descending, no unit repeated — 1h30m is valid, 30m1h is not. Digit separators (_) are allowed per §6.3.

Result (timeout above, illustrating the resolved shape — the exact tag/base-unit is implementation-defined):

%{"timeout" => {:duration, 500_000_000}}

Also reachable as a tagged value (§7.5) — !duration("5m") — for computed/interpolated values; the bare literal is recommended when static.

6.9 Byte sizes

A number immediately followed by a unit, no whitespace. Matching is case-insensitive; an i right after the magnitude letter (before B) signals binary, its absence decimal.

  • Decimal (SI, ×1000): B, kB, MB, GB, TB, PB
  • Binary (IEC, ×1024): KiB, MiB, GiB, TiB, PiB
memory_limit = 512MiB
disk_quota = 10GB

The numeral may be fractional (1.5GiB).

Result (memory_limit above):

%{"memory_limit" => {:bytes, 536_870_912}}

Also reachable as !bytes("512MiB") (§7.5), for computed/interpolated values.

6.10 Lists

list1 = ["a", "b", "c"]
list2 = [
  "a"
  "b"
]

Comma, newline, or whitespace all separate elements.

6.11 Tuples

Parentheses — chosen to avoid colliding with blocks, which already own { ... } (§5.4) as both structure and map type:

office_location = (52.5200, 13.4050)

Result:

%{"office_location" => {52.52, 13.405}}

A genuine, fixed-arity Elixir tuple — not a list. Arity is fixed at parse time and is part of the value: a tuple is never merged element-wise or appended to, only replaced wholesale by a later statement (§8.3). An implementation must carry a real tuple type through parsing, merging, and the final returned value.

Comma, newline, or whitespace separate elements, same as lists. ; and | are reserved for splitting environment-variable strings into lists (§7.2), not for tuples.

7. Interpolation and references

Three brace-delimited sigils, plus two extensibility mechanisms sharing the same shape (§9).

7.1 Variables

@name = value (§5.2) or a loop binding (§5.5) defines a variable; referencing one — standalone, in a string, or inside another sigil — always uses @{name}. There is no bare @name reference form.

  • @{name}
  • @{name:default} — default if undefined
  • @{name:+alt} — substitute alt if defined (any value), else empty string
  • @{name:?"message"} — fail loading with message if undefined
  • @{name[i]}, @{name[i]:default} — index into a list-valued variable

  • @{name | filter} — normalize the resolved value (§7.2)

Private variables (@*var) are visible only in the declaring file. Loop-bound variables are visible only in their loop's body, shadowing any outer variable of the same name for that scope.

7.2 Environment expansion

Deliberately bash-like. Names match /^[a-zA-Z_][a-zA-Z0-9_]*$/.

  • ${NAME}
  • ${NAME:default} — default if unset/empty
  • ${NAME:+alt} — substitute alt if set, else empty string
  • ${NAME:?"message"} — fail loading if unset
  • ${?NAME} — skip the whole statement if unset/empty
  • ${NAME[]:[...]} — parse as a list (split on ,/;), with a default
  • ${NAME[i]:default} — index into the split list

${...} always resolves to a string. For a number or boolean, wrap it in the matching tagged value (§7.5) rather than relying on CASC to guess:

port = !int(${PORT:8080})
debug = !bool(${DEBUG:false})

This is deliberate, not an oversight: guessing a type from what a string looks like is exactly the kind of implicit behavior that produces surprises, so CASC never does it silently.

Note: real bash uses ${NAME:-default} (with a dash); CASC drops it so the suffix grammar (:default, :+alt, :?"msg") is identical across @{...}, ${...}, and %{...} — one rule, not three near-identical ones.

Filters

A reference may be followed by one or more | filter clauses, which normalize the resolved string. Like the suffix grammar, this is one rule shared by all three reference forms — @{...}, ${...}, and %{...} — not three near-identical ones.

  • ${NAME | trim} — strip leading and trailing whitespace

  • ${NAME | downcase}, ${NAME | upcase} — change case

  • ${NAME | trim_prefix: "..."}, ${NAME | trim_suffix: "..."} — strip an affix if present

Filters apply after the suffix has settled, left to right, so a filter always sees the value that will actually be used — whether that came from the reference or from its default:

scheme = ${SCHEME | trim_suffix: "://" | downcase}
region = ${REGION:"  us-east-1  " | trim}

An argument may be double- or single-quoted. Single quotes exist so a filter remains usable inside an interpolated string, where a double-quoted argument has no way to nest:

endpoint = "${SCHEME | trim_suffix: '://'}://%{host}"

Filters normalize; they do not convert. A filter applied to a non-string is a load-time error rather than a silent coercion, an unknown filter name is a load-time error naming it, and so is an argument given to a filter that takes none (or omitted from one that requires it). To change a value's type, use a tagged value (§7.5).

Example, REGION=eu-west set, others unset:

region = ${REGION}
max_retries = !int(${MAX_RETRIES:3})
${?FEATURE_FLAG}
enabled = true
allowed_hosts = ${ALLOWED_HOSTS[]:["localhost"]}

Result (enabled is skipped entirely, guarded by ${?FEATURE_FLAG}):

%{"region" => "eu-west", "max_retries" => 3, "allowed_hosts" => ["localhost"]}

Built names

A reference's name may be built by interpolation instead of written out, by giving it as a double-quoted string:

@which = "HOST"
host = ${"APP_@{which}"}     # reads APP_HOST

This exists for one thing a document otherwise cannot do at all: follow a deployment convention of one variable per tenant. CASC reads named variables and has no way to enumerate the environment, so without it a set like TOKEN_1, TOKEN_2, … can only be written out one line at a time. With a loop (§5.5) it becomes a list of ids:

@supervisor_ids = ["1", "2", "3"]

for @id in @{supervisor_ids} as tokens {
  "@{id}" = ${"TOKEN_@{id}"}
}
%{"tokens" => %{"1" => "tok-one", "2" => "tok-two", "3" => "tok-three"}}

The same applies to @{"..."} and to a %{...} path segment (%{tokens."supervisor-@{id}"}).

This is interpolation — the concatenation the language already has, pointed at the name — not a new expression form. There is deliberately no concatenation operator, for the same reason the filter set is closed: that road ends at an expression language, and this is a config format.

Three rules:

  • A built ${...}/@{...} name must resolve to an identifier ([A-Za-z_][A-Za-z0-9_]*), so a built name and a written-out one are interchangeable. A %{...} key needn't — keys are quoted and may contain anything but a ., which would silently split the path.
  • A name or key may not be built from a secret. Names appear in error messages, which do not redact.
  • Only the bare form may be built. A reference nested inside a larger string keeps a plain name: its token ends at the first }, so a nested @{...} would truncate it — the same scope trim the reduced :default grammar already has in that position.

7.3 Config references

  • %{path}
  • %{path:default}
  • %{path[i]}, %{path[i]:default}

  • %{path | filter} — normalize the resolved value (§7.2)

path follows the dotted key-path rules (§5.4).

Resolves lazily, against the fully-merged final tree — after every import, merge, and override sigil (§8), not eagerly per-file. A reference tracks a value "wherever it ends up," independent of import order: if a later import replaces the whole tls block via ~tls { ... }, %{tls.min_version} resolves to the replacement's value even if written in an earlier-processed file.

Cycle detection is required. %{a} referencing something that transitively contains %{a} again must fail at load time with the cycle path in the error — never loop forever or silently return a partial value.

Example:

server.host = "api.internal"
server.port = 8080
health_check.url = "http://%{server.host}:%{server.port}/health"
admin_email = %{contact.admin:"ops@example.com"}

Result (no contact.admin defined):

%{
  "server" => %{"host" => "api.internal", "port" => 8080},
  "health_check" => %{"url" => "http://api.internal:8080/health"},
  "admin_email" => "ops@example.com"
}

7.4 Extensible resolution

!{resolver:payload} dispatches to a consumer-registered resolver named resolver, passing payload through. The first : splits name from payload; everything to the matching closing } is the payload verbatim, brace-balanced.

*password = !{vault:secret/db/password}

Giving vault meaning is entirely the consumer's job. An unregistered resolver name is a load-time error naming it (§9.4, §9.2).

7.5 Tagged values

!Name(argument) constructs a value of type Name from one argument (typically a string). Parsing only needs to recognize "a tag plus one parenthesized argument" — giving it meaning is the registered handler's job.

Built in: !int, !float, !bool (coercion, mainly for ${...}, §7.2), !duration, !bytes (constructors for §6.8/§6.9), !trim, !downcase, !upcase (normalization), and !module (below). The normalizing tags do for a whole value what the matching filter (§7.2) does for one reference, and share its rule that a non-string argument is an error rather than a coercion. Anything else — e.g. !uuid("...") — is consumer-defined. An unregistered tag is a load-time error naming it (§9.4, §9.1).

!module("Name") names a module of the host language:

client_module = !module("ASCO.Redis.TestClient")
formatter = !module("${LOG_FORMATTER}")

It exists because §6.4's atoms are bare identifiers, so a dotted module name cannot be written as a literal — and because a module is often deployment-selected, which means it arrives through ${...} as a string.

The tag is the same in every implementation; the shape it accepts is not. What counts as a module name belongs to the language a given implementation targets, so a document that names a module stays readable across ports even where the naming convention differs. This implementation accepts dot-separated identifiers, mapping an upper-case initial to an Elixir module (Foo.Bar → Elixir.Foo.Bar) and anything else to an Erlang module (crypto → :crypto).

Like a bare atom literal, this creates an atom, with the same caveat: fine for a fixed, trusted set of configuration files, not for untrusted input.

8. Merge model

Merge operates on the fully-desugared tree (§5.4). Imports behave like "early writes," overridden by later imports and then the current file's own statements (§5.1); later statements override earlier ones for the same path. Each type's default is whichever behavior is safe without a schema:

8.1 Blocks and maps

Deep-merge by default. a { b = 1 } then a { c = 2 } produces both a.b and a.c, at every depth — safe, since merging two maps by key has one obvious meaning.

8.2 Lists

Replace wholesale by default. Merging positionally is well-defined but rarely intended; merging by identity requires knowing which field to key by, which CASC has no schema to supply. Replacing avoids guessing.

8.3 Tuples

Always replace, never merge. A tuple's fixed arity (§6.11) is part of its meaning; element-level append/remove would silently change it. None of ~/+/- apply to a tuple-valued key — using any of them is a load-time error, not a silent no-op.

8.4 Overriding the default (sigils)

  • ~key { ... } — replace instead of merge.
  • +key = [...] — append instead of replace (plain assignment if key doesn't exist yet).
  • -key = [...] — remove matching elements instead of replace.
  • -key.path (bare) — delete the path entirely, any type.

Ordering with #/* is fixed (§5.7).

Example — an earlier import established:

server { host = "0.0.0.0", port = 8080 }
tags = ["a", "b", "c"]
feature.legacy_mode = true

A later file writes:

~server { port = 9090 }
+tags = ["d"]
-tags = ["b"]
-feature.legacy_mode

Result:

%{"server" => %{"port" => 9090}, "tags" => ["a", "c", "d"], "feature" => %{}}

server lost host (replaced, not merged); tags gained "d", lost "b"; feature.legacy_mode is gone entirely, not false.

8.5 Deferred: merge-by-key

An explicit, opt-in merge-by-key for lists — e.g. ~items[id] = [...], matching elements by an id field, a CASC-native analogue of Kustomize's strategic merge patch. Deliberately out of scope for this version.

9. Extensibility

CASC's core grammar is intentionally small and frozen. Three points let a consuming application add its own vocabulary:

9.1 Tagged values

!Name(...) (§7.5) — the parser recognizes shape only, never meaning. !int/!float/!bool/!duration/!bytes are built in; anything else is consumer-registered.

9.2 Extensible resolvers

!{resolver:payload} (§7.4) — a named callback the consumer registers, receiving the payload verbatim. Mirrors Ichor's own Actions mechanism, rather than introducing a new architectural idea.

9.3 Extensible import sources

import "scheme://..." (§5.1) — dispatched to a consumer-registered loader for that scheme. A bare path always keeps filesystem-plus-glob/brace behavior; anything past the scheme, including whether expansion applies at all, is the loader's to define.

9.4 Failure semantics

The rule underlying all three: an unregistered tag, resolver, or import scheme is always a load-time error naming the offender — never silently ignored, never passed through as an opaque value. This keeps the core language small and predictable while staying genuinely open: consumers add vocabulary explicitly, and CASC never guesses at what an unrecognized extension point meant.

10. Minimal valid file

#@version = 1.0