fix: preserve extra_content for Gemini 3 thought_signature support

[maywzh]

Context

This PR adds support for preserving extra_content (including Gemini 3's thought_signature) throughout the tool call lifecycle.

Gemini 3 models include extra_content with thought_signature in tool call responses. This metadata was being lost during multi-turn conversations, potentially affecting model coherence and performance.

Implementation

• Added extra_content field to ToolUse, McpToolUse, and streaming interfaces
• Preserved extra_content in OpenAI and OpenRouter providers during streaming
• Passed extra_content through message format conversion
• Tracked and inherited extra_content during tool call streaming lifecycle in Task.ts

Changes:

• src/api/providers/openai.ts - Extract and preserve extra_content
• src/api/providers/openrouter.ts - Extract and preserve extra_content
• src/api/transform/openai-format.ts - Include extra_content in conversion
• src/api/transform/stream.ts - Add extra_content to streaming interfaces
• src/core/assistant-message/NativeToolCallParser.ts - Parse and track extra_content
• src/core/task/Task.ts - Preserve extra_content in tool call event handling
• src/shared/tools.ts - Add extra_content to ToolUse and McpToolUse interfaces

Backward Compatibility:
All changes are backward compatible:

• extra_content is optional (?)
• Existing providers without extra_content continue to work unchanged

Screenshots

before	after
N/A	N/A

How to Test

• Manual testing with Gemini 3 models via OpenRouter
• Verified multi-turn conversations preserve thought signatures

Get in Touch

[changeset-bot]

🦋 Changeset detected

Latest commit: 543050c

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
kilo-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[kilo-code-bot]

[WARNING]: Unvalidated as any cast on untrusted API response data

(toolCall as any).extra_content extracts an arbitrary property from the API stream response without any validation. This value flows through the entire pipeline and is eventually sent back to the API in subsequent requests.

Consider adding minimal validation (e.g., checking it's a plain object) to prevent unexpected data types from propagating:

const rawExtra = (toolCall as Record<string, unknown>).extra_content
extra_content: rawExtra != null && typeof rawExtra === 'object' && !Array.isArray(rawExtra)
  ? rawExtra as Record<string, unknown>
  : undefined,

[kilo-code-bot]

[WARNING]: Type safety regression — const toolUseBlock: any loses type checking

The previous code pushed a typed object literal directly into assistantContent (typed as Array<Anthropic.ContentBlockParam>). Changing to any disables all type checking on this object.

Consider using an intersection type instead to preserve type safety while allowing the extra property:

const toolUseBlock: Anthropic.ToolUseBlockParam & { extra_content?: Record<string, unknown> } = {
  type: "tool_use" as const,
  id: sanitizedId,
  name: mcpBlock.name,
  input: mcpBlock.arguments,
}

The same applies to the second toolUseBlock: any on line 3828.

[kilo-code-bot]

Code Review Summary

Status: 3 Issues Found | Recommendation: Address before merge

Fix these issues in Kilo Cloud

Overview

Severity	Count
CRITICAL	0
WARNING	3
SUGGESTION	0

Issue Details (click to expand)

WARNING

File	Line	Issue
src/api/providers/openai.ts	528	Unvalidated as any cast on untrusted API response to access extra_content
src/api/transform/openai-format.ts	521	Unvalidated as any cast on toolMessage (typed as Anthropic.ToolUseBlockParam) to access extra_content
src/core/task/Task.ts	3793	Type safety regression — toolUseBlock typed as any instead of a proper intersection type

Other Observations (not in diff)

Issues found in unchanged code or cross-cutting concerns that cannot receive inline comments:

File	Line	Issue
src/core/assistant-message/__tests__/NativeToolCallParser.spec.ts	—	Missing test coverage for extra_content propagation. No tests verify that extra_content is stored in rawChunkTracker, emitted in tool_call_start events, preserved through startStreamingToolCall → finalizeStreamingToolCall, or included in the final ToolUse/McpToolUse from parseToolCall/parseDynamicMcpTool. Per project guidelines, code changes should have test coverage.
src/api/providers/gemini.ts	325	The native Gemini provider does not pass extra_content through tool_call_partial chunks. If Gemini 3 thought_signature is also exposed through the native Gemini SDK (not just OpenAI-compatible endpoints), this would be a gap.
src/core/task/Task.ts	3160, 3189	Additional as any casts on existingToolUse in the delta and end event handlers follow the same pattern flagged at line 3793. Consider defining a shared type for content blocks that includes extra_content.

Files Reviewed (7 files)

• src/api/providers/openai.ts - 1 issue (existing)
• src/api/transform/openai-format.ts - 1 issue
• src/api/transform/stream.ts - 0 issues
• src/core/assistant-message/NativeToolCallParser.ts - 0 issues (test gap noted in observations)
• src/core/task/Task.ts - 1 issue (existing) + observations
• src/shared/tools.ts - 0 issues
• .changeset/gemini-extra-content.md - 0 issues

[kilo-code-bot]

[WARNING]: Unvalidated as any cast on toolMessage to access extra_content

Same pattern as in openai.ts:528. The toolMessage is typed as Anthropic.ToolUseBlockParam which doesn't include extra_content. Consider adding a type guard or extending the type with an intersection to avoid silently accepting any shape:

interface ToolUseBlockWithExtra extends Anthropic.ToolUseBlockParam {
  extra_content?: Record<string, unknown>
}
const toolMessageWithExtra = toolMessage as ToolUseBlockWithExtra

This would provide compile-time safety while still allowing the extra field.

[kevinvandijk]

Thanks!