[BUG] API Error when using the Requesy provider

[MichaelYochpaz]

Problem (one or two sentences)

I tried using Gemini 3 Pro through Requesty. The initial message (starting the task) worked great, but I tried to send it another message as a follow-up after its response and then got an API error saying:

Requesty completion error: 400 Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call `default_api:execute_command` , position 2. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.

Then, I tried the same with Claude Opus and got another error with a similar flow with the following message:

Requesty completion error: 400 messages.1.content.0.type: Expected `thinking` or `redacted_thinking`, but found `text`. When `thinking` is enabled, a final `assistant` message must start with a thinking block (preceeding the lastmost set of `tool_use` and `tool_result` blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable `thinking`. Please consult our documentation at https://docs.claude.com/en/docs/build-with-claude/extended-thinking

Context (who is affected and when)

It happened to me specifically on 2 tasks, one with the Gemini 3 Pro and once with Claude Opus 4.5 (with thinking enabled) models.

Both errors occurred after the initial starting prompt was working and I got a response, then got a request to run a command for additional debug information (right after the initial response from the model), which I ran manually because the issue is on a remote server, and then when I try to send it back in the conversation, I get errors with both models.

Reproduction steps

1. Start a new task using Request as a provider, and google/gemini-3-pro-preview as a model.
2. After the initial response from the model, write a response in the thread.

Expected result

The task work regularly as expected

Actual result

Getting an error

Variations tried (optional)

No response

App Version

v3.38.0

API Provider (optional)

Requesty

Model Used (optional)

google/gemini-3-pro-preview, anthropic/claude-opus-4-5

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

Requesty completion error: 400 Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call `default_api:execute_command` , position 2. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.

Requesty completion error: 400 messages.1.content.0.type: Expected `thinking` or `redacted_thinking`, but found `text`. When `thinking` is enabled, a final `assistant` message must start with a thinking block (preceeding the lastmost set of `tool_use` and `tool_result` blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable `thinking`. Please consult our documentation at https://docs.claude.com/en/docs/build-with-claude/extended-thinking

[roomote]

I've analyzed this issue and here's my proposed implementation plan:

Problem Analysis

The errors occur when using Requesty as a router to access models with reasoning/thinking capabilities:

1. Gemini 3 Pro error: "Function call is missing a thought_signature in functionCall parts"

   • Gemini thinking models require thoughtSignature on function call parts when reasoning is enabled
   • The current Requesty handler doesn't inject these signatures for tool calls in conversation history

2. Claude Opus 4.5 with thinking error: "Expected thinking or redacted_thinking, but found text"

   • Claude with thinking enabled requires the last assistant message to start with a thinking block
   • Thinking blocks from previous turns need to be preserved and passed back

Proposed Solution

The fix mirrors how OpenRouter handles these scenarios (see openrouter.ts):

For Gemini models with native tools:

Inject reasoning_details with skip_thought_signature_validator for assistant messages that have tool calls but no existing encrypted reasoning. This tells the Gemini API to skip thought signature validation when the history doesn't have real signatures (e.g., when switching models or on follow-up messages).

Changes to requesty.ts:

1. Detect if the model is a Gemini model and using native tool protocol
2. Post-process openAiMessages to inject fake reasoning.encrypted blocks for tool calls that lack them
3. Add getReasoningDetails() method to capture and return reasoning details from streaming responses (similar to OpenRouter)
4. Accumulate reasoning_details during streaming to preserve them for conversation history

For Claude models with thinking:

This is more complex as it requires preserving Anthropic-native thinking blocks. However, since Requesty uses OpenAI-compatible format, this may be a limitation of the router itself. I'll investigate further during implementation whether Requesty supports passing thinking blocks in their API format.

Files to modify:

• src/api/providers/requesty.ts - Main handler changes

This approach will solve the Gemini error definitively. For the Claude thinking error, I'll verify during implementation what's supported by Requesty's API.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.