Describe the tool_choice mapping as the code does it

OpenAiRequestMapper's class Javadoc listed tool_choice and
parallel_tool_calls among the fields forwarded verbatim as raw JSON.
Only messages and tools are: tool_choice goes in its string form, a
named function becomes "required" with only that tool forwarded, any
other object form is ignored and the request runs as "auto", and
parallel_tool_calls is read as a boolean. The tool_choice comment in
ResponsesApiSupport.toOpenAiChatRequest described the string form only
and now says the same. Closes the F109 addendum that named both texts.
Documentation only.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VBFFnrhtvXiUeyxjmUc4R2

Author: claude

BUGHUNT-FOUND.md

@@ -1056,6 +1056,7 @@ Repos plus eins. Geprüfte Klassen ohne Fund stehen in [`BUGHUNT-CHECKED.md`](BU
 - **Reproduktion:** `OpenAiRequestMapperTest#toolChoiceNamingAFunctionRequiresAToolCall`, `ResponsesApiSupportTest#toolChoiceNamingAFunctionRequiresAToolCall`
 - **Rot-Grün-Nachweis:** ohne Fix rot, je Weg für sich (2 Fehlschläge, je „Expected: is "required" but: was ""“), mit Fix grün (28/28 und 51/51, skipped 0); Gegenprobe per `git stash` beider Produktionsdateien rot (1 + 1), nach `pop` grün. `OpenAiCompatServerHttpTest` 47/47, `AnthropicApiSupportTest` 22/22, `ResponsesStreamTranslatorTest` 5/5, `OllamaApiSupportTest` 32/32, `LlamaArchitectureTest` 13/13, `spotless:apply` und `spotbugs:check` grün.
 - **Nachtrag (2026-10-18, nur beschrieben):** Zwei Kommentare neben dem Fix sind nicht nachgezogen: die Klassen-Javadoc von `OpenAiRequestMapper` nennt `tool_choice` unter den Feldern, die „verbatim as raw JSON“ weitergehen (galt schon vorher nur für die Textform), und der Kommentar über dem Werkzeugblock in `ResponsesApiSupport.toOpenAiChatRequest` beschreibt nur die Textform. Reine Doku-Abweichung, nach der Nutzerregel ohne Test nicht nachträglich geändert; Vorschlag: `tool_choice` aus der Verbatim-Aufzählung nehmen und „ein benanntes Werkzeug wird zu `required`“ ergänzen.
+- **Nachtrag (Branch-Review 2026-10-20, Funde J2-6, J5-8):** Erledigt. Die Klassen-Javadoc von `OpenAiRequestMapper` nennt als wörtlich weitergereicht nur noch `messages` und `tools`; `tool_choice` geht in der Textform weiter, ein benanntes Werkzeug wird zu `required` mit nur diesem Werkzeug (F227), jede andere Objektform (`allowed_tools`, `custom`) wird ignoriert und läuft als `auto`, `parallel_tool_calls` wird als Boolean gelesen. Der Kommentar über `tool_choice` in `ResponsesApiSupport.toOpenAiChatRequest` sagt dasselbe. Reine Doku, kein Test.
 - **Nachtrag (Branch-Review 2026-10-20):** Mit mehreren Werkzeugen ließ `required` jeden Aufruf zu; seit F227 geht auf beiden Wegen nur das benannte Werkzeug weiter, ein unbekannter oder fehlender Name ist 400.
 - **Status:** gefixt
 - **Commit:** java-llama.cpp `c5823a5`

llama/src/main/java/net/ladenthin/llama/server/OpenAiRequestMapper.java

@@ -18,12 +18,15 @@
 /**
  * Pure mapping from an OpenAI {@code /v1/chat/completions} request body to {@link InferenceParameters}.
  *
- * <p>The structural fields — {@code messages}, {@code tools}, the string form of {@code tool_choice}, and
- * {@code parallel_tool_calls} — are forwarded
- * <em>verbatim</em> as raw JSON so the full OpenAI shape (assistant {@code tool_calls},
- * {@code role:"tool"} results with {@code tool_call_id}, and vision {@code image_url} content parts)
- * round-trips untouched into the native chat-template parser. A {@code tool_choice} naming one function
- * becomes {@code required} with only that tool forwarded (see {@link #namedToolChoice}). Sampling fields are translated to the
+ * <p>The structural fields {@code messages} and {@code tools} are forwarded <em>verbatim</em> as raw JSON
+ * so the full OpenAI shape (assistant {@code tool_calls}, {@code role:"tool"} results with
+ * {@code tool_call_id}, and vision {@code image_url} content parts) round-trips untouched into the native
+ * chat-template parser. {@code tool_choice} is forwarded in its string form ({@code auto}, {@code none},
+ * {@code required}, the only values llama.cpp's chat parser accepts); a {@code tool_choice} naming one
+ * function becomes {@code required} with only that tool forwarded (see {@link #namedToolChoice}), and any
+ * other object form ({@code allowed_tools}, {@code custom}) is ignored, so the request runs as
+ * {@code auto}, which is also what upstream reads from a non-string. {@code parallel_tool_calls} is read as
+ * a boolean. Sampling fields are translated to the
  * matching {@code InferenceParameters.with*} setter, as are the llama.cpp sampling extensions upstream's
  * request schema reads ({@code min_p}, {@code repeat_penalty}, {@code grammar}, ...) and the reasoning
  * controls upstream's chat

llama/src/main/java/net/ladenthin/llama/server/ResponsesApiSupport.java

@@ -159,8 +159,9 @@ static ObjectNode toOpenAiChatRequest(JsonNode request) {
                 throw new IllegalArgumentException("tool_choice names unknown tool '" + namedTool + "'");
             }
             // The Responses API uses the same tool_choice + parallel_tool_calls fields as chat; forward
-            // them so the shared chat core honors them. The mapper consumes the string form of
-            // tool_choice ("auto"/"none"/"required"), which is what we forward here.
+            // them so the shared chat core honors them. The string form (auto/none/required) goes as is,
+            // a named function becomes "required" with only that tool forwarded above, and any other
+            // object form is dropped, so the request runs as "auto".
             if (toolChoice.isTextual()) {
                 openAi.put("tool_choice", toolChoice.asText());
             } else if (namedTool != null) {