[BUG] AgentFlow V2: "Unknown content type thinking" on any agent using Gemini models with thinking enabled

[azamiftikhar1000]

fix(FlowiseChatGoogleGenerativeAI): handle 'thinking' content type round-trip

Summary

Closes the round-trip serialization bug in FlowiseChatGoogleGenerativeAI.ts that breaks every agent loop running on a Gemini model with thinking enabled. Users see this in production as:

Error in Agent node: Unknown content type thinking

What was wrong

FlowiseChatGoogleGenerativeAI.ts has two sites that produce LangChain content blocks of shape { type: 'thinking', thinking: '...', signature: '...' } when parsing Gemini's response parts (the parts where Google sets thought: true). But _convertLangChainContentToPart has no case for content.type === 'thinking' — it handles text, executableCode, codeExecutionResult, image_url, media, tool_use, tool_call, mimetype patterns and functionCall, then falls through to the throwing else.

So the conversion is asymmetric: response ➜ LangChain works, LangChain ➜ request throws. On the FIRST agent turn this is invisible. On the SECOND iteration (or any tool-call loop / structured-output step) the assistant message's previous thinking block hits _convertLangChainContentToPart and crashes with Unknown content type thinking at line 307.

How the patch fixes it

Adds a branch that translates the LangChain thinking block back into Google's native shape per the Gemini thinking docs: a text Part with thought: true and an optional thoughtSignature (preserving the Gemini-3 signature continuity that #5715 added for functionCall parts).

} else if (content.type === 'thinking') {
    const text = (content as any).thinking ?? (content as any).text ?? ''
    const signature = (content as any).signature ?? (content as any).thoughtSignature
    return {
        text: typeof text === 'string' ? text : String(text ?? ''),
        thought: true,
        ...(signature ? { thoughtSignature: signature } : {})
    } as Part
}

Why two LangChain keys are accepted

The existing response-parser sites in this file emit signature (the LangChain v0 convention). thoughtSignature is also accepted because newer @langchain/google-genai versions put the signature there. Defensive on both for forward compatibility.

Verification

Added FlowiseChatGoogleGenerativeAI.test.ts with 5 cases. Red-green proven:

	Without patch	With patch
round-trips a thinking content block back into a Gemini text Part with thought=true	❌ throws at line 307	✅ pass
preserves the thoughtSignature for Gemini-3 multi-turn tool-call continuity	❌ throws	✅ pass
also accepts thoughtSignature on the LangChain block (alternate key name)	❌ throws	✅ pass
coerces non-string thinking payload to a string instead of throwing	❌ throws	✅ pass
still throws "Unknown content type" for truly unrecognized types	✅ pass	✅ pass

PASS nodes/chatmodels/ChatGoogleGenerativeAI/FlowiseChatGoogleGenerativeAI.test.ts
Tests:       5 passed, 5 total

The throwing path is preserved — only the thinking type is intercepted; truly unknown types still surface as before.

Reproduction

1. Build any AgentFlow V2 chatflow with at least one agentAgentflow node
2. Set its model to gemini-3-flash-preview (or gemini-2.5-flash with a thinkingConfig)
3. Set thinkingLevel: LOW (or higher)
4. Send any user message that requires at least one internal iteration of the agent loop (e.g. with tools attached, or structured output enabled)
5. Without this patch: the flow errors at ~9s with Error in Agent node: Unknown content type thinking
6. With this patch: the flow completes normally; thought summaries are sent back to Gemini per the official API contract

Related

• fix: preserve thoughtSignature in Gemini function calls for thinking models #5715 — preserved thoughtSignature on functionCall parts (related but doesn't address the thinking content type round-trip)
• Feat/add thinking budget to gemini #5395 — added thinkingBudget/thinkingLevel to the Gemini UI (made it possible to opt in to thinking but the parser wasn't updated for the response shape)
• Google docs: Thinking, Thought signatures

Test plan

• Unit tests pass (5/5)
• Pre-commit hooks pass (lint-staged + prettify)
• No new TypeScript errors introduced (verified against base — same 12 pre-existing errors in src/storage/* and src/utils.ts, none in touched file)

[azamiftikhar1000]

#6449