CAMEL-24345: camel-google-vertexai - apply the streamOutputMode and jsonMode options

[oscerd]

GoogleVertexAIConfiguration declares two options that no code ever read:

@UriParam(label = "producer", description = "Streaming output mode: complete (default) or chunks",
          defaultValue = "complete", enums = "complete,chunks")
private String streamOutputMode = "complete";

@UriParam(label = "producer", description = "Whether to use JSON request/response format", defaultValue = "false")
private boolean jsonMode;

They are documented and exposed in the catalog, so streamOutputMode=chunks looked configurable
while generateChatStreaming always concatenated the whole response, and jsonMode did nothing at
all. The CamelGoogleVertexAIStreamOutputMode header constant was equally unread.

• streamOutputMode=chunks now produces a List<String> with one element per streamed chunk, ready
  for a split(). The default complete keeps returning the concatenated String, so routes that
  do not set the option are unaffected. The header now works too and takes precedence over the
  endpoint option.
• jsonMode=true sets the request's response MIME type to application/json.

Both are covered by an upgrade-guide entry, since an option that starts being honoured is a
behaviour change for anyone who had set it.

Two smaller items from the same audit:

• generateChat and generateCode delegate verbatim to generateText while being advertised as
  distinct operations. They are now documented as aliases in the enum javadoc — no behaviour change.
• buildRawPredictRequestBody called body.getClass() on its failure path, so a message with no
  body failed with a NullPointerException instead of the intended "must be a JSON String, Map, or
  plain text prompt" message.

Claude Code on behalf of oscerd

🤖 Generated with Claude Code

[davsclaus]

Nice catch wiring up the two declared-but-never-read options. The PR is clean, well-tested, and follows project conventions.

One minor observation (not blocking): the CHUNK_COUNT header now counts only chunks with actual text content (chunks.size()) whereas the old code counted all stream response objects including those with null text. The new semantics are arguably more useful, but it's a subtle behavioral change that isn't mentioned in the upgrade guide entry. Worth adding a note if you agree.

Everything else looks good — config resolution follows the standard header-overrides-config pattern, tests use AssertJ, upgrade guide entry is clear, and the NPE fix in buildRawPredictRequestBody is a nice bonus.

This review was generated by an AI agent and may contain inaccuracies. Please verify all suggestions before applying.

Claude Code on behalf of davsclaus