zog

Zog is a Zod-shaped API over Go semantics, and the two disagree in ways that still compile. Import as z "github.com/Oudwins/zog". Three biases decide most calls:

• A remembered Zod default is a defect until checked. Fields are optional by default, z.Enum does not exist, and a schema key names a Go struct field rather than an input key.
• Parse and Validate disagree about what missing means, and that disagreement decides which tests in a chain run at all.
• Zog panics on a schema its author built wrong, and never on input data. A panic is a defect in the schema definition, fixed there.

Schema Shape

• A z.Shape key names a Go struct field, never an input key. The first letter is case-corrected, so "name" and "Name" both bind to Name. A key matching no field panics with Struct Schema Definition Error ... missing expected schema key.
• Struct tags map input keys and nothing else. Resolution runs json, form, query, or env for the source in use, then zog, then the schema key as written.
• z.Struct(...).Required() and .Optional() compile and do nothing. An optional nested struct needs z.Ptr(z.Struct(...)).NotNil().
• Wrap the schema in z.Ptr wherever the destination field is a pointer. A z.Slice(...) against a *[]T field panics with a type-cast error.
• z.Ptr carries only .NotNil(). Put Required, Default, Catch, Test, and Transform on the inner schema.
• Declare every schema once, at package level. A schema rebuilt per call costs roughly twice the time and several times the memory of a reused one, and the gap widens with the number of fields.
• Pick, Omit, Extend, and Merge return shallow copies and are not type-checked. A key naming no struct field panics at execution rather than failing to compile.

Read [${CLAUDE_SKILL_DIR}/references/schema-catalog.md] when reaching for a constructor, a validator, a test option, or an issue code that is not already in the chain — it carries every schema type with its full method set and signatures, the three option families, and the issue-code and type constants.

Missing Values

Parse and Validate implement "missing" differently, and every rule below follows from that split.

• Parse treats only a nil input as missing. z.String().Required().Parse("", &dest) reports nothing, and an empty form field, an empty query parameter, and a JSON "" all arrive present. Add .Min(1) where an empty value must fail.
• Validate treats every Go zero value as missing, so a legitimate 0, "", or false fails .Required().
• A missing value skips every test on an optional schema. Under Validate that means z.Bool().True() accepts false, z.Int().GT(10) accepts 0, z.Slice(...).Min(1) accepts an empty slice, and a custom TestFunc never runs. Wrapping in z.Ptr(...).NotNil() does not change it, because the inner schema still short-circuits.
• To reject a zero value under Validate, use .Required() and set its message — z.Bool().Required(z.Message("terms must be accepted")). No other test reaches the value.
• .Required() runs before every test and transform, whatever its position in the chain. z.String().Trim().Required() accepts " " and then trims it to "".
• .Default outranks .Required and applies in both modes, so Validate replaces an explicitly set zero value with the default. Tests still run on the substituted value.
• .Catch swallows every issue — coercion failure, required, and each test — writes its value, and stops the chain. Keep it off anything whose failure must reach the caller.
• Structs support neither .Default nor .Catch; slices support .Default but not .Catch.

Parse and Validate

• Parse(data, &dest, opts...) at an IO boundary, Validate(&value, opts...) on an already-typed value. Parse coerces the input; Validate does not, and allocates nothing on the success path.
• Both destinations are pointers. A value destination panics rather than reporting an issue.
• z.WithCoercer is a SchemaOption and belongs on the constructor — z.String(z.WithCoercer(fn)). z.WithCtxValue and z.WithIssueFormatter are ExecOptions and belong on Parse or Validate.
• Read a per-execution value inside a test with ctx.Get(key), paired with z.WithCtxValue(key, val) at the call.

Tests and Transforms

• Chain position is execution order. z.String().Trim().Min(3) measures the trimmed value; .Min(3).Trim() measures the raw one.
• .TestFunc(fn, opts...) is Zod's refine; .Test(z.Test[T]{Func: ...}) is superRefine, and the latter adds its own issues with ctx.AddIssue(ctx.Issue().SetMessage(...)).
• A primitive test receives a typed pointer; a struct, slice, or map test receives any — assert it to the struct pointer inside the function.
• The package-level z.TestFunc(code, fn, opts...) builds a reusable z.Test[T] for .Test(...). It takes an issue code as its first argument; the method of the same name does not.
• .Not() negates only the next test and prefixes that test's issue code with not_. The interface it returns omits Min, Max, Test, TestFunc, and a second Not.
• A transform mutates through its pointer and returns an error to stop the chain. .Trim() is a transform, so it runs where it sits rather than before the tests.
• z.Preprocess(fn, schema) receives the raw input under Parse but a pointer to the value under Validate. A type mismatch there panics rather than reporting an issue.

Types Without a Built-In Schema

• A named type over a primitive needs the Like constructor — z.StringLike[Env](), z.IntLike[Status](). A plain z.String() against a type Env string field panics with a type-cast error.
• A type Zog does not know takes z.CustomFunc(fn, opts...), which validates through a typed pointer and performs no coercion.
• A wrapper such as sql.NullString takes z.Boxed(schema, unbox, box), added in v0.21.10.

Read [${CLAUDE_SKILL_DIR}/references/custom-schemas.md] when a value's Go type has no built-in schema, or its input shape does not match one — it carries the named-type, CustomFunc, Boxed, and fully custom z.Use forms with working code, the Preprocess contract, the coercion hooks, and the experimental map and recursive constructors.

Issues

• Every schema returns z.ZogIssueList, which is []*z.ZogIssue; test it with len(errs) > 0.
• An issue path segment is the input key, not the schema key. One schema yields first_name through zjson and first-name through Validate, because each source resolves the field through its own tag. Give a field a zog tag equal to its json tag where the error keys must stay stable across both.
• A root issue carries a nil path and flattens under zconst.ISSUE_KEY_ROOT, the string "$root".
• issue.Message is the only field safe to return to a user. issue.Err holds the underlying cause and belongs in a log.
• Never read an issue after z.Issues.Collect — it returns the struct to a pool the next parse draws from.

Read [${CLAUDE_SKILL_DIR}/references/issue-handling.md] when a validation failure has to be shaped into a response body, a template, or a log line — it carries the ZogIssue fields, the path rules per input source, and the rendered output of Flatten, GroupByFlattenedPath, Treeify, and Prettify.

Input Adapters

• zhttp.Request(r) routes on method first, then on Content-Type, and falls through to query parsing for any unrecognized content type — a JSON body sent as text/plain silently yields nothing.
• Multipart requires r.ParseMultipartForm in the handler first. Without it the parser reports invalid_multipart_form and the schema never runs.
• zhttp and zjson parse into a struct only. A JSON array, a bare primitive, and null each produce one invalid_json root issue.
• An adapter failure produces one root issue and stops — invalid_json, invalid_form, invalid_multipart_form, or invalid_query.
• zenv trims every value and treats an empty variable as absent, so PORT= triggers .Required() and takes a .Default().

Read [${CLAUDE_SKILL_DIR}/references/input-adapters.md] when the input comes from an HTTP request, a JSON body, or the environment — it carries the handler and config shapes, the content-type routing table, the query and form value rules including repeated and []-suffixed keys, and the parser override hook.

Messages

• Set a message on the test that produces it — z.Message("..."), or z.MessageFunc for one built from e.Params and e.Value.
• Reassigning conf.DefaultIssueMessageMap or conf.DefaultErrMsgMap silently does nothing. The formatter closes over the map at package initialization. Mutate entries in place, or install a new formatter with conf.IssueFormatter = conf.NewDefaultFormatter(m).

Read [${CLAUDE_SKILL_DIR}/references/messages-and-i18n.md] when an issue message must be overridden or served in more than one language — it carries the four override layers with their precedence, the message template placeholders, and the i18n setup with its language-key and fallback rules.

Panics

A panic always names a schema definition error, never bad input. Four causes account for every one:

• a z.Shape key naming no struct field
• a destination passed by value rather than by pointer
• a schema whose type does not match the destination field
• a coercer returning a value of the wrong Go type

Fix the schema. Never wrap a Parse or Validate call in recover.

Application

When writing Zog, apply these conventions silently — do not narrate a rule while following it. Declare schemas as package-level variables and check every result with len(errs) > 0. Where existing code contradicts a convention, follow the codebase and flag the divergence once.

When reviewing Zog, cite the violation and show the fix inline. Do not lecture. Treat a .Required() on a value arriving from a form, a query string, or JSON as a defect until an emptiness test sits beside it.

Bad:  "Required only fires on a nil input, so an empty query parameter passes through..."
Good: z.String().Required() -> z.String().Required().Min(1)

Integration

The golang skill governs every Go decision outside the Zog API — naming, error handling, testing conventions, and the toolchain — and wins on any question of how the Go code reads. This skill governs schema definition and the Zog runtime API, and its key-naming rule outranks Go naming inside a z.Shape. The coding skill governs workflow. All are active at once.

Zog is pre-1.0 and breaks its API across minor versions. Read the version in go.mod before writing against anything this skill anchors to one.

A schema states what the data must be; the mode decides what missing means.