Say where /api/generate's load answer deviates from Ollama

The isLoadOnly Javadoc attributed the "and without a suffix" condition
to Ollama, and toOllamaLoadResponse called its answer "as Ollama gives
it". Ollama's GenerateHandler reads the prompt alone and answers a
keep_alive of 0 with done_reason "unload". Both are deliberate choices
here: an empty prompt with a suffix stays fill-in-the-middle at the
start of a file (also for an empty suffix), and keep_alive 0 keeps the
load answer because nothing is unloaded. The Javadoc, the README row of
the Ollama routes, TODO.md and the F205 entry now say so. Documentation
only, no behaviour change.

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

Author: claude

BUGHUNT-FOUND.md

@@ -1705,6 +1705,7 @@ Repos plus eins. Geprüfte Klassen ohne Fund stehen in [`BUGHUNT-CHECKED.md`](BU
 - **Rot-Grün-Nachweis:** Ohne den Fix der HTTP-Test rot (das Backend generierte); mit ihm grün (64 + 37).
 - **Status:** gefixt
 - **Commit:** java-llama.cpp `200cc90`
+- **Nachtrag (Branch-Review 2026-10-20, Funde J1-4, J3-3, J16-2):** Beide Regeln sind bewusste Abweichungen von Ollama, nicht Ollamas Verhalten: Ollamas `GenerateHandler` prüft nur `req.Prompt == ""` (der `suffix` fügt dort nur `CapabilityInsert` hinzu), beantwortet also einen leeren Prompt mit `suffix` mit der Ladeantwort, und antwortet bei `keep_alive: 0` ohne Prompt mit `done_reason "unload"`. Hier bleibt der leere Prompt mit `suffix` FIM am Dateianfang (auch bei `"suffix":""`, `hasSuffix` fragt nur nach einem String), und `keep_alive: 0` bekommt die Ladeantwort, weil nichts entladen wird — ein „unload“ wäre eine Meldung über eine Aktion, die nicht stattfand. Die Javadoc von `isLoadOnly`/`toOllamaLoadResponse`, die README-Zeile der Ollama-Routen und `TODO.md` sagen das jetzt; vorher schrieb die Javadoc die `suffix`-Bedingung Ollama zu. Reine Doku, kein Test.
 
 ### F213 – java-llama.cpp: Ollama `/api/chat` mit einem `messages`-Element, das kein Objekt ist, schließt die Verbindung ohne Antwort
 - **Zeit:** 2026-10-20 19:55 UTC

README.md

@@ -998,7 +998,7 @@ path), all supporting tools and streaming:
 
 | Surface | Routes | For |
 |---|---|---|
-| **Ollama-native** | `GET /api/version`, `GET /api/tags`, `POST /api/show`, `POST /api/chat` (NDJSON streaming), `POST /api/generate` (through the chat template with `system`, as Ollama does; the bare prompt with `raw: true`; FIM with `suffix`; without a prompt Ollama's "model loaded" answer) | Copilot's built-in **Ollama** provider; Ollama-hardcoded tools |
+| **Ollama-native** | `GET /api/version`, `GET /api/tags`, `POST /api/show`, `POST /api/chat` (NDJSON streaming), `POST /api/generate` (through the chat template with `system`, as Ollama does; the bare prompt with `raw: true`; FIM with `suffix`; without a prompt and without a suffix Ollama's "model loaded" answer; unlike Ollama, a request with only a suffix stays fill-in-the-middle at the start of the file, and `keep_alive: 0` does not unload and is answered "load") | Copilot's built-in **Ollama** provider; Ollama-hardcoded tools |
 | **Anthropic Messages** | `POST /v1/messages` (SSE event stream) | Claude-shaped clients (Claude Code); Copilot `messages` apiType |
 | **OpenAI Responses** | `POST /v1/responses` (SSE event stream) | Copilot `responses` apiType; Responses-API clients |
 

TODO.md

@@ -520,6 +520,8 @@ they are not rediscovered:
   until a client needs the exact count.
 - **`OpenAiCompatServer` cannot unload a model**: an Ollama `/api/generate` with `keep_alive: 0`
   and no prompt gets the "load" answer (F205), the model stays resident (one model per server).
+  Ollama itself unloads there and answers `done_reason "unload"`; this server does not claim an
+  unload it did not do.
 
 ### The release: 6.0.0, with an "Upgrading from 5.1" section (decided 2026-10-15; roadmap item 9)
 

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

@@ -542,8 +542,16 @@ static boolean hasSuffix(JsonNode request) {
 
     /**
      * Whether the {@code /api/generate} request only asks for the model to be loaded: Ollama answers a
-     * request without a prompt (and without a {@code suffix}) with {@code done} at once, which is how
-     * its clients preload a model. The model of this server is always loaded, so there is nothing to do.
+     * request without a prompt with {@code done} at once ({@code done_reason "load"}), which is how its
+     * clients preload a model. The model of this server is always loaded, so there is nothing to do.
+     *
+     * <p>Two deliberate differences from Ollama, whose check reads the prompt alone. A request with an
+     * empty prompt and a {@code suffix} is not load-only here but fill-in-the-middle at the start of a
+     * file, a request editors send; Ollama answers it with the load answer. That includes an empty
+     * {@code suffix} ({@link #hasSuffix} only asks for a string), which Ollama would not treat as
+     * fill-in-the-middle either. And {@code keep_alive} is not read: Ollama unloads the model on a
+     * {@code keep_alive} of {@code 0} without a prompt and answers {@code done_reason "unload"}, while
+     * this server keeps its one model loaded and gives the load answer.
      *
      * @param request the parsed Ollama generate request
      * @return {@code true} when the request carries no prompt and no suffix
@@ -553,9 +561,11 @@ static boolean isLoadOnly(JsonNode request) {
     }
 
     /**
-     * The answer to a load-only {@code /api/generate} request, as Ollama gives it: one object, empty
-     * {@code response}, {@code done} and {@code done_reason "load"}, whether streaming was asked for
-     * or not.
+     * The answer to a load-only {@code /api/generate} request, in the form of Ollama's load answer: one
+     * object, empty {@code response}, {@code done} and {@code done_reason "load"}, whether streaming was
+     * asked for or not. It is also the answer to a {@code keep_alive} of {@code 0}, which Ollama answers
+     * with {@code done_reason "unload"}: nothing is unloaded here, so the server does not claim it (see
+     * {@link #isLoadOnly}).
      *
      * @param model the model id to echo
      * @return the answer