@@ -1,14 +1,14 @@
 name: Local engines
 
-# Agreement with Jev and throughput of typesafe-java-local's engines, as a table in the job summary.
+# Agreement with Jev and throughput of typesafe-java-client-local's engines, as a table in the job summary.
 # Model directories are cached: Laya fp32 (+ fp16 on macOS, for the GPU) and Qwen 4-bit, all onnx-community exports.
 on:
   workflow_dispatch:
   push:
     branches: [main]
-    paths: ['local/**', '.github/workflows/local.yml', '.github/scripts/local_summary.py']
+    paths: ['client-local/**', '.github/workflows/local.yml', '.github/scripts/local_summary.py']
   pull_request:
-    paths: ['local/**', '.github/workflows/local.yml', '.github/scripts/local_summary.py']
+    paths: ['client-local/**', '.github/workflows/local.yml', '.github/scripts/local_summary.py']
 
 jobs:
   numbers:
@@ -48,23 +48,23 @@ jobs:
           uvx --from huggingface_hub hf download Cloudflare/clef-flash --revision 17f0b0ad64efb65d273590632833508766b2aae6 tokenizer.json --local-dir $M/clef-flash
 
       - name: Build
-        run: ./mvnw -B -q -pl local -am test-compile dependency:build-classpath -Dmdep.outputFile=$PWD/cp.txt -Dmdep.includeScope=test
+        run: ./mvnw -B -q -pl client-local -am test-compile dependency:build-classpath -Dmdep.outputFile=$PWD/cp.txt -Dmdep.includeScope=test
 
       - name: Real-model tests
-        run: ./mvnw -B -pl local -am test -DexcludedGroups=acceptance -Dengine=laya -Dsurefire.failIfNoSpecifiedTests=false
+        run: ./mvnw -B -pl client-local -am test -DexcludedGroups=acceptance -Dengine=laya -Dsurefire.failIfNoSpecifiedTests=false
 
       - name: Agreement with Jev
-        working-directory: local
+        working-directory: client-local
         run: |
           CP="$(cat ../cp.txt):target/classes:target/test-classes"
           ENGINES="laya qwen"
           if [ "$RUNNER_OS" = macOS ]; then ENGINES="$ENGINES laya-fp16-gpu"; fi
           for e in $ENGINES; do
-            java -cp "$CP" io.github.dfa1.typesafe.local.JevComparison $e | tee -a ../jev.txt
+            java -cp "$CP" io.github.dfa1.typesafe.client.local.JevComparison $e | tee -a ../jev.txt
           done
 
       - name: Throughput
-        working-directory: local
+        working-directory: client-local
         run: |
           CP="$(cat ../cp.txt):target/classes:target/test-classes"
           ENGINES=LAYA,QWEN
 
@@ -7,16 +7,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+- **Breaking: modules split into contract (`core`) and implementations (`client-*`), names aligned** — `core` keeps the `TypeSafeClient` interface, its decorators, the model and `JsonCodec`; `DefaultTypeSafeClient`/`ApiKey`/`HttpTransport` move to the new `client-http`, `TypeSafeClient.builder()` becomes `DefaultTypeSafeClient.builder()`, and every module's artifact and package now match its directory (`client-http-jdk`, `client-http-okhttp`, `codec-jackson2`, `codec-jackson3`, `client-local`, `client-mapping`, `client-testkit`); see ADR 0003.
 - **docs: local-model setup as three steps** — dependency, Laya download, code, then a model comparison; the optional Clef 4-bit script runs from its GitHub URL with `uv run`, and load errors link to the how-to.
-- **`local`: hardening** — special tokens in caller text stay plain text (no prompt injection via `<|im_end|>`), the tokenizer cache is bounded, and Clef checks its weights' sha256 on first load.
-- **`local`: shared request validation, `evaluateAsync` off virtual threads** — every engine rejects no questions or empty criteria with `BadRequest` (Laya used to NPE or answer a null choice), and `evaluateAsync` runs on one platform thread so native inference can't pin the JVM's virtual-thread carriers.
-- **`local`: `LocalClefTypeSafeClient.loadOnGpu(...)`** — Clef-flash on ONNX Runtime's WebGPU backend (macOS on Apple Silicon), same answers as the CPU; Laya gets no GPU option, since its WebGPU answers drift between launches. (#19)
+- **`client-local`: hardening** — special tokens in caller text stay plain text (no prompt injection via `<|im_end|>`), the tokenizer cache is bounded, and Clef checks its weights' sha256 on first load.
+- **`client-local`: shared request validation, `evaluateAsync` off virtual threads** — every engine rejects no questions or empty criteria with `BadRequest` (Laya used to NPE or answer a null choice), and `evaluateAsync` runs on one platform thread so native inference can't pin the JVM's virtual-thread carriers.
+- **`client-local`: `LocalClefTypeSafeClient.loadOnGpu(...)`** — Clef-flash on ONNX Runtime's WebGPU backend (macOS on Apple Silicon), same answers as the CPU; Laya gets no GPU option, since its WebGPU answers drift between launches. (#19)
 - **docs: run Clef-flash on a Mac with MLX** — a how-to pointing the regular client at mlx-community's local Clef-flash server: about 0.55 s per request on an M5, much closer to Jev than Laya. (#14)
-- **`local`: `LocalClefTypeSafeClient`** — Cloudflare's Clef-flash (Qwen3.5-9B + joint schema head) on Ollaya's ONNX graph over the upstream bf16 weights; one forward pass per request; `scripts/clef/quantize_q4.py` converts it to 4-bit weights (7.7 GB, 2–4 s per request on an M5 GPU). (#14)
-- **`local`: `LocalLayaTypeSafeClient`, `LocalQwenTypeSafeClient`** — new module, one client class per model, evaluating in-process on ONNX Runtime from a local model directory (Laya fp32 from onnx-community, or Qwen2.5 as a prompted-LLM baseline), with a `Local engines` workflow publishing agreement-with-Jev and throughput tables. (#14)
+- **`client-local`: `LocalClefTypeSafeClient`** — Cloudflare's Clef-flash (Qwen3.5-9B + joint schema head) on Ollaya's ONNX graph over the upstream bf16 weights; one forward pass per request; `scripts/clef/quantize_q4.py` converts it to 4-bit weights (7.7 GB, 2–4 s per request on an M5 GPU). (#14)
+- **`client-local`: `LocalLayaTypeSafeClient`, `LocalQwenTypeSafeClient`** — new module, one client class per model, evaluating in-process on ONNX Runtime from a local model directory (Laya fp32 from onnx-community, or Qwen2.5 as a prompted-LLM baseline), with a `Local engines` workflow publishing agreement-with-Jev and throughput tables. (#14)
 - **build: checkstyle engine 14.3.0** — the plugin's default 9.3 can't parse Java 21 pattern matching for switch; same rules. (#14)
 - **Breaking: optional API key, `ApiKey` without a public constructor** — `TypeSafeClient.builder()` takes no argument and `.apiKey(key)` is optional (no key, no `Authorization` header, for local servers); `ApiKey` is a final class built with `of(String)`/`fromFile`/`fromDefaultFile`/`fromEnv`, and no longer exposes its value. (#18)
-- **`jackson2`: ignore unknown fields, like `jackson3`** — a field the API adds to a response no longer fails decoding with `ResponseDecoding`; Jackson 2's default rejected it. (#18)
+- **`codec-jackson2`: ignore unknown fields, like `codec-jackson3`** — a field the API adds to a response no longer fails decoding with `ResponseDecoding`; Jackson 2's default rejected it. (#18)
 - **`core`: `Content.fields(...)` and `EvaluateRequest.Builder` keep their order** — both copied with `Map.copyOf`, which reshuffles keys per JVM run, so the same request serialized with fields/questions in a different order each run; they now keep the caller's iteration order.
 - **`core`: `TokenCounter`** — running totals of input/output tokens from `EvaluateResponse#usage()`, added via `.decorateWith(tokens::decorate)`; thread-safe, shareable across clients.
 - **`core`: `Builder.decorateWith(...)` naming and docs** — named to read as additive and to pair with the `decorate(...)` factories; `Builder` javadoc and the how-to explain decorator ordering.