[BUG] LlmResponse reports a candidate that produced nothing as a success

[svetanis]

Please make sure you read the contribution guide and file the issues in the right place.
Contribution guide.

🔴 Required Information

Describe the Bug

LlmResponse.Builder.response(GenerateContentResponse) decides whether a model response is a
success by testing whether the candidate's content object is present, never whether it holds
anything:

// core/src/main/java/com/google/adk/models/LlmResponse.java
this.finishReason(candidate.finishReason().orElse(null));
if (candidate.content().isPresent()) {
  this.content(candidate.content().get());
  this.groundingMetadata(candidate.groundingMetadata().orElse(null));
} else {
  candidate.finishReason().ifPresent(this::errorCode);
  candidate.finishMessage().ifPresent(this::errorMessage);
}

A candidate that was cut off mid-generation comes back with content present and part-less —
"content": {} on the wire. content().isPresent() is true, so the success branch runs:
errorCode and errorMessage are never set, and candidate.finishMessage(), the model's own
explanation, is discarded. The turn ends with an empty answer that is indistinguishable from a model
that simply had nothing to say.

errorCode is documented as "Error code if the response is an error", yet the result carries
finishReason=MAX_TOKENS with errorCode empty.

Steps to Reproduce

Run any LlmAgent against a thinking-capable model with room to think and no room to answer, so the
turn ends before a visible token is emitted:

GenerateContentConfig.builder()
    .maxOutputTokens(24)
    .thinkingConfig(ThinkingConfig.builder().thinkingBudget(2048).build())
    .build();

On gemini-3.1-flash-lite this returns:

{
  "candidates": [{ "content": {}, "finishReason": "MAX_TOKENS", "index": 0 }],
  "usageMetadata": { "thoughtsTokenCount": 37, "totalTokenCount": 70 }
}

Expected Behavior

An error response: errorCode set from the finish reason, errorMessage from finishMessage. A
candidate that produced nothing and did not stop normally is not a successful turn.

Observed Behavior

A success response. content is set to the empty candidate content, errorCode and errorMessage
are absent, finishMessage is dropped. Observed end-to-end through a real agent run:

--- Starved - maxOutputTokens=24, thinkingBudget=2048 ---
  finish reasons        : [MAX_TOKENS]
  LlmResponse errorCodes: [(absent)]
  parts per response    : [0]
  text reaching caller  : []

--- Control - maxOutputTokens=200, thinking off ---
  finish reasons        : [STOP]
  LlmResponse errorCodes: [(absent)]
  parts per response    : [1]
  text reaching caller  : [READY]

The caller gets an empty answer and no diagnostic anywhere.

Environment Details

• ADK Library Version: 1.7.2-SNAPSHOT
• OS: Windows 11 — not OS-specific

Model Information

gemini-3.1-flash-lite; any thinking-capable Gemini model reaches the same shape.

---

🟡 Optional Information

Regression

No — the branch has had this shape since the method was introduced.

Willingness to contribute

Yes — a PR follows immediately. One predicate, in one method: decide on whether the candidate
produced parts, treating a STOP with nothing to say as a legitimate empty turn.

boolean hasParts =
    candidate.content().flatMap(Content::parts).map(parts -> !parts.isEmpty()).orElse(false);
boolean stopped =
    candidate.finishReason().map(reason -> reason.knownEnum() == FinishReason.Known.STOP).orElse(false);
if (hasParts || stopped) {
  this.content(candidate.content().orElse(null));
  this.groundingMetadata(candidate.groundingMetadata().orElse(null));
} else {
  candidate.finishReason().ifPresent(this::errorCode);
  candidate.finishMessage().ifPresent(this::errorMessage);
}

The same predicate also corrects the mirror-image case in this branch — a candidate with no
content and finishReason=STOP currently comes out as errorCode=STOP, marking a turn that ended
normally as an error. It becomes a plain successful response.

The prompt-feedback and no-candidates branches are left exactly as they are.

How often has this issue occurred?

Always (100%), whenever the model returns a part-less candidate.