fix(core): correctly consolidate multi-part text content

This PR fixes an issue where multi-part text content in streaming responses was incorrectly consolidated by ensuring new text is appended to the last existing text part and improving the identification of pure text content.

What was done

• Refactored the content consolidation logic in GeminiChat to correctly handle multi-part streaming responses.
• Changed text appending from the first part of existing content to the last text part.
• Introduced isFirstPartText and isLastPartText helper methods to precisely check if the first or last content part is pure text.
• Enhanced isLastPartText to validate only model-generated content and explicitly exclude non-text content types like functionCall, inlineData, or thought.
• Updated isFirstPartText to also explicitly exclude non-text content types for accurate identification.

Impact

• Resolves incorrect concatenation of text within multi-part streaming outputs from the Gemini model.
• Improves the reliability and accuracy of text merging, especially when responses contain mixed content types (e.g., text interwoven with images or tool calls).
• Prevents text from being erroneously merged into non-text content parts.
• No breaking changes are introduced.

Technical details

• Modifications were made in packages/core/src/core/geminiChat.ts.
• The core logic update involves changing how consolidatedOutputContents and lastHistoryEntry are modified to append text to parts[parts.length - 1] instead of parts[0].
• New type predicate functions isFirstPartText and isLastPartText were added and refined to ensure strict validation of pure text parts, filtering out other content types.

Related issues

• #4053

Metadata

• Source: github
• Repo: google-gemini/gemini-cli
• URL: https://github.com/google-gemini/gemini-cli/pull/6235