Hermes Agent Community Reports Vertex Streaming Parsing Gap; Gemini Thought Summaries May Enter Final Responses

Hermes Agent received a report on November 1 about a streaming integration bug: when using provider: vertex in Chat Completions mode with Gemini thinking enabled, thought summaries may appear in the final response delivered to users. The author said Telegram displayed summary text such as itinerary organization and answer planning, and that the database also stored this text in the regular content field while the reasoning field remained empty. Community report

The author traced the issue to parsing a provider-specific extension field: some stream chunks place the summary in delta.content and use extra_content.google.thought: true to indicate its purpose. The report says Hermes’s accumulation and delivery paths do not check this marker, so the text is merged into the answer. It includes the request conditions for a direct Vertex call and returned stream excerpts. The test model was Gemini 3.8 Flash, and the Hermes code was pinned to commit 818c13b. Issue details and reproduction conditions

Google’s official documentation for the compatible interface clearly distinguishes two pieces of information: the Boolean thought marks thought content, while thought_signature is a signature that validates thought information; the former also takes precedence over text-form thought labels. This supports the view that the integration layer must identify the purpose of content, but the documentation alone cannot establish that every Hermes version has the same defect. Google interface documentation

Hermes’s official documentation says the system automatically maps reasoning effort settings to Vertex’s extra_body.google.thinking_config. From an engineering perspective, this report highlights the maintenance cost of compatible interfaces: request formats can be shared, while responses may still carry provider-specific semantics. If an accumulator merges summaries and answers first, later display, history records, and content exports may lose the basis for distinguishing them. This is a technical inference from the field behavior. Hermes Vertex guide

The evidence so far rests mainly on one author’s observations. The issue remains open, with no formal fix reported. Follow-up should check whether any fix covers streaming, non-streaming, and database writes, and whether it adds regression tests that preserve summary classification. This concerns thought summaries returned by the service; the available evidence is insufficient to conclude that the model’s full internal reasoning was exposed.