diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 000000000..264124de2
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,212 @@
+# AGENTS.md
+
+Guidance for AI agents working in the `substrait-java` repository. This complements
+`CONTRIBUTING.md` (commit conventions, style guide) with practical, codebase-specific
+knowledge.
+
+## What this project is
+
+`substrait-java` is the Java implementation of [Substrait](https://substrait.io/) —
+a cross-language specification for relational query plans. It provides an immutable
+POJO model for plans/relations/expressions/types and bidirectional conversion to and
+from the Substrait protobuf wire format, plus integrations (Isthmus → Apache Calcite,
+Spark).
+
+## Module layout
+
+| Module | Path | Purpose |
+| --- | --- | --- |
+| `:core` | `core/` | POJO model, proto converters, function extension handling. The heart of the project. |
+| `:isthmus` | `isthmus/` | Calcite SQL ⇄ Substrait conversion. |
+| `:spark` | `spark/` + variants `spark-3.4_2.12`, `spark-3.5_2.12`, `spark-4.0_2.13` | Spark plan ⇄ Substrait (Scala). Source shared in `spark/src`; each variant compiles it against a different Spark/Scala version — see Spark notes. |
+| `:examples` | `examples/isthmus-api`, `examples/substrait-spark` | Runnable examples. |
+| `:isthmus-cli` | `isthmus-cli/` | CLI over `:isthmus`, compiled to a GraalVM **native image** (`nativeCompile`) — the `isthmus` binary + smoke tests. |
+| `build-logic` | `build-logic/` | Gradle **included build**: Kotlin-DSL convention plugins (`substrait.java-conventions` → shared Java config + PMD). Has its own `gradle.properties`; does not inherit the root's. |
+
+The `substrait/` directory is a **git submodule** of the upstream spec, and several
+`:core` build inputs are read from it — none are vendored in this repo, so don't look
+for them under `core/src`:
+
+- **Proto**: `substrait/proto/substrait/*.proto` → generated Java in package `io.substrait.proto`.
+- **ANTLR grammars**: `substrait/grammar/*.g4` (notably `SubstraitType.g4`, the type /
+  function-signature grammar) → parser generated under `io.substrait.type`.
+- **Standard extension YAMLs**: `substrait/extensions/*.yaml` (e.g. `functions_arithmetic.yaml`)
+  → packaged into `:core` resources at `substrait/extensions/`.
+- **Validation schemas**: `substrait/text/simple_extensions_schema.yaml` and
+  `dialect_schema.yaml` → validate extensions/dialects (mainly in tests).
+
+`:core:submodulesUpdate` (submodule init/update) therefore gates proto and grammar
+generation. These files are owned by the upstream spec — to change them, update
+`substrait` upstream and bump the submodule pin, don't edit them in-tree.
+
+## Core architecture (the pattern most changes follow)
+
+The POJO model uses **Immutables** (`org.immutables:value`):
+
+- Interfaces/abstract classes are annotated `@Value.Immutable`; the enclosing type
+  (e.g. `Expression`) is `@Value.Enclosing`. The build generates
+  `ImmutableExpression.Foo` etc. **These generated classes do not exist until you
+  compile**, so your IDE/diagnostics will show "cannot be resolved" errors for
+  `ImmutableExpression.*` and new `builder()` methods until `:core:compileJava` runs.
+  This is expected — compile to confirm.
+- Each POJO exposes a static `builder()` delegating to the generated immutable.
+- Immutables copies an accessor's Javadoc **verbatim** into the generated
+  `ImmutableXxx` (which lives in the same package as the abstract type). A
+  `{@link}`/`@see` targeting a type in a *different* package resolves in the source
+  file (which imports it) but NOT in the generated file, so `:core:javadoc` fails with
+  `reference not found`. **Fully-qualify** cross-package `{@link}` targets in the
+  Javadoc of any `@Value.Immutable` accessor.
+
+Expressions are visited via a double-dispatch **visitor**:
+
+- `ExpressionVisitor` (interface) — one `visit(...)` overload per concrete expression
+  type. Direct implementors **must** implement every method.
+- `AbstractExpressionVisitor` — provides `visitFallback`-based defaults for every
+  method, so subclasses only override what they need. Implementors that extend this
+  (e.g. isthmus `ExpressionRexConverter`, spark `DefaultExpressionVisitor`) do **not**
+  break when a new expression type is added.
+- Adding a new expression type therefore means: add the POJO, add a `visit` method to
+  `ExpressionVisitor`, add a default to `AbstractExpressionVisitor`, and update the
+  **direct** implementors (`ExpressionProtoConverter`, `ExpressionCopyOnWriteVisitor`,
+  examples `ExpressionStringify`).
+
+The same double-dispatch pattern recurs across the model, so the "add a case → update the
+direct implementors" rule applies well beyond expressions:
+
+- **Relations**: `RelVisitor` / `AbstractRelVisitor` — same `*Visitor` + `Abstract*Visitor`
+  shape as expressions — plus copy-on-write transformers `RelCopyOnWriteVisitor` and
+  `ExpressionCopyOnWriteVisitor`.
+- **Types**: `TypeVisitor`, extended by `ParameterizedTypeVisitor` / `TypeExpressionVisitor`
+  for function-signature and derived-type expressions. Interface-only (no `Abstract*Visitor`
+  fallback), so implementors must handle every type kind.
+- **Function arguments**: a `FunctionArg` is an `Expression`, `Type`, or `EnumArg`, dispatched
+  via `FunctionArg.accept(fnDef, argIdx, FuncArgVisitor, ctx)`; the nested `FuncArgVisitor`
+  exposes `visitExpr` / `visitType` / `visitEnumArg` instead of per-POJO overloads. (Function
+  *invocations* are `Expression` subtypes, so `ExpressionVisitor` already covers them.)
+- Expression, relation, and function-argument visitors thread a `VisitationContext` (`C`) type
+  parameter; the type visitors do not.
+
+Proto conversion is split into two directions, and the class name tells you which:
+**`<Thing>ProtoConverter` = POJO → proto**, **`Proto<Thing>Converter` = proto → POJO**.
+
+- **POJO → proto** (`<Thing>ProtoConverter`) is a **visitor** over the POJO model —
+  `ExpressionProtoConverter` (an `ExpressionVisitor`), `RelProtoConverter` (a `RelVisitor`),
+  `TypeProtoConverter` (a `TypeExpressionVisitor`). Reuse `toProto(io.substrait.type.Type)`
+  to get a wrapped proto `Type`, then extract the needed sub-message.
+- **proto → POJO** (`Proto<Thing>Converter`) is a `switch` on the proto `oneof` case enum —
+  `ProtoExpressionConverter`, `ProtoRelConverter`, `ProtoTypeConverter`.
+- The same pair exists for each model layer: expressions, types, relations, `Plan`
+  (`PlanProtoConverter` / `ProtoPlanConverter`, which delegate to the rel converters),
+  function extensions (`ExtensionProtoConverter`), extended expressions, and field masks.
+  Changing or adding a proto message means updating **both** converters of the affected
+  layer, plus a round-trip test.
+- **Enum ↔ proto-enum** mapping lives in the POJO enum itself, not in these converters:
+  each constant is built with its generated proto counterpart and exposes `toProto()` +
+  static `fromProto(proto)` (e.g. `Set.SetOp`, `Join.JoinType`, `Expression.WindowBoundsType`,
+  `Plan.VariableEvaluationMode`); the message converters just call those. `fromProto` scans
+  `values()` and throws on an unrecognized proto value — so when several POJO constants map
+  to one proto value, order the deprecated ones **last** so the reverse lookup returns the
+  canonical one (see `Join.JoinType`). Adding an enum value is an edit to the enum, not the
+  converters.
+- POJO types are created with `TypeCreator.REQUIRED` / `TypeCreator.NULLABLE`.
+
+## Building, testing, formatting
+
+- Build/run a module's tests: `./gradlew :core:test`
+- Single test class: `./gradlew :core:test --tests "io.substrait.<pkg>.<Class>"`
+- Format code: `./gradlew spotlessApply` (Google Java Format); scope to a module with
+  `./gradlew :core:spotlessApply` when iterating. `spotlessCheck` runs in CI.
+- Static analysis: `substrait.java-conventions` applies **PMD** (custom ruleset at
+  `build-logic/src/main/resources/substrait-pmd.xml`) to the Java modules, and `check`/
+  `build` **fail** on violations. Easy ones to trip: missing `@Override`, unused private
+  fields/methods/locals, `var` (rule `UseExplicitTypes` — use explicit types), and
+  `public` JUnit 5 test classes/methods (they must be package-private).
+- Spark subprojects are nested: compile with
+  `./gradlew :spark:spark-3.5_2.12:compileScala`.
+- Requires running Gradle with JDK 17 and the `--add-exports` flags in
+  `~/.gradle/gradle.properties` (see `CONTRIBUTING.md`), or `spotlessApply` may fail.
+  Keep the daemon on JDK 17 *consistently*: `build-logic`'s convention plugins are
+  compiled to bytecode matching the daemon's JDK, so switching JDKs between builds can
+  leave cached plugins a later daemon can't load (`UnsupportedClassVersionError`) —
+  `./gradlew --stop` then rebuild clears the stale daemon/cache.
+- The `isthmus-cli` **native image** is the exception to JDK 17: `nativeCompile` uses
+  whatever JDK runs the Gradle daemon (`graalvmNative { toolchainDetection = false }` in
+  `isthmus-cli/build.gradle.kts`), so it needs the daemon on a **GraalVM** JDK with
+  `native-image` — CI uses GraalVM **25**, while everything else declares a JDK 17
+  toolchain. Switching the daemon between the two is a common cause of the daemon/cache
+  churn noted above.
+- `build-logic/` is an **included build**, not a normal subproject: it does **not**
+  inherit the root `gradle.properties`, so its Kotlin-compile daemon heap is set in
+  `build-logic/gradle.properties` (raising the root heap does nothing for it). Avoid
+  `--no-build-cache` casually — it forces `build-logic`'s Kotlin plugins to recompile
+  every run.
+- CI (`.github/workflows/pr.yml`) runs the **full** `./gradlew build --rerun-tasks`,
+  plus `yamllint`, `editorconfig-checker`, and commitlint (all also wired as local
+  pre-commit hooks). Narrower local tasks pass while the PR fails — build the whole
+  thing before pushing.
+- `javadoc` runs doclint that **fails the build**, but only via `build`/`javadocJar` —
+  not via `compileJava` / `test` / `spotlessCheck`. So Javadoc errors surface only on
+  the PR; run `./gradlew :core:javadoc` before pushing.
+- The `substrait` proto submodule needs its **tags** fetched: the generated
+  `SubstraitVersion` derives from `git describe --tags` on the submodule, so a
+  shallow/tagless checkout yields a malformed version that throws at *runtime* (not
+  build time). CI fetches them explicitly because `actions/checkout` doesn't.
+
+When you add to the public expression/visitor API, verify the dependent modules still
+compile — they have their own visitor implementors:
+`./gradlew :core:spotlessCheck :isthmus:compileJava :spark:spark-3.5_2.12:compileScala :examples:substrait-spark:compileJava`
+
+## Isthmus (Calcite conversion) notes
+
+- Build Calcite `RelBuilder`s for Substrait→Calcite conversion with
+  `.typeSystem(SubstraitTypeSystem.TYPE_SYSTEM)` (see `ConverterProvider.getRelBuilder`).
+  Calcite's default type system caps decimal precision at 19 while converted
+  expressions carry precision-38 types; the mismatch makes `RexSimplify` (run by
+  `RelBuilder.project`) re-derive decimal arithmetic at precision 19 and wrap the result
+  in a truncating `CAST(… AS DECIMAL(19,0))`. `SubstraitTypeSystem` keeps a public
+  no-arg constructor because Frameworks/Avatica reinstantiates it from its class name
+  via reflection — don't remove it.
+
+## Spark (multi-version) notes
+
+- The Scala source is **shared, not duplicated**: it lives in `spark/src/main/scala`
+  (+ `spark/src/test/scala`). The three subprojects (`spark-3.4_2.12`, `spark-3.5_2.12`,
+  `spark-4.0_2.13`) have **no source of their own** — each sets its source set to the
+  shared dir and compiles it against a different version: Spark 3.4.4 / 3.5.4 on Scala
+  **2.12**, Spark 4.0.2 on Scala **2.13**. Edit the shared source once; don't copy
+  changes across variants. `spark/` itself is just an orchestrator (`buildAllVariants`).
+- Version-specific code goes in overlay source dirs `spark/src/main/spark-<major.minor>`
+  (`spark-3.4`, `spark-3.5`, `spark-4.0`) — that's how per-version Spark API differences
+  are handled without `#if`-style branching.
+- Because Spark 4.0 compiles on Scala **2.13** and the others on 2.12, a shared-source
+  change that compiles on one variant can still break another. Compiling only
+  `:spark:spark-3.5_2.12` misses 2.13 breaks — also run
+  `./gradlew :spark:spark-4.0_2.13:compileScala`, or build all three with `:spark:build`.
+- Formatting is done by the **parent** `:spark` project (`./gradlew :spark:spotlessApply`);
+  spotless is disabled in the variant subprojects.
+
+## Testing conventions
+
+- Proto round-trip tests extend `io.substrait.TestBase` and call
+  `verifyRoundTrip(Expression)` / `verifyRoundTrip(Rel)` to assert POJO → proto → POJO
+  fidelity. See `core/src/test/java/io/substrait/type/proto/DynamicParameterRoundtripTest.java`
+  for the canonical pattern.
+
+## Conventions & workflow
+
+- **Conventional commits** are required (CI lints them, and PR title + body must form a
+  valid commit message). Scope tags seen in history: `feat(core)`, `feat(pojo)`,
+  `feat(isthmus)`, `feat(extensions)`, `build(deps)`, `chore(release)`. A `!` marks a
+  breaking change.
+- **No GitHub issue/PR references in source** (comments or Javadoc) — they belong in
+  commit messages and PR descriptions. `Closes #NNN` in the commit/PR body is fine;
+  in the code, describe behavior and spec version (e.g. `spec v0.88.0`) instead.
+- Many features track upstream Substrait spec releases (see epic-style issues); a new
+  proto message usually needs: POJO + visitor wiring + both proto converters + a
+  round-trip test, and often `ExpressionCreator` factories and `dsl/SubstraitBuilder`
+  helpers for ergonomics.
+- When monitoring PR checks, budget for a long tail: the **macOS `Build Isthmus Native
+  Image`** job is the long pole — it `needs:` the `java` + `integration` jobs (so it
+  starts late), then AOT-compiles for ~15–20 min on a slower macOS runner. A PR staying
+  yellow after the other checks pass usually just means that job is still running, not
+  that it's stuck or failing.
diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 000000000..43c994c2d
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1 @@
+@AGENTS.md