fix: Gemini thought signature validation errors (#10694)

Co-authored-by: Roo Code <roomote@roocode.com>
This commit is contained in:
Daniel 2026-01-15 23:25:42 -05:00 committed by GitHub
parent e34d93e2cb
commit ddac338fdd
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 618 additions and 11 deletions

View file

@ -17,7 +17,11 @@ import { NativeToolCallParser } from "../../core/assistant-message/NativeToolCal
import type { ApiHandlerOptions } from "../../shared/api"
import { convertToOpenAiMessages } from "../transform/openai-format"
import {
convertToOpenAiMessages,
sanitizeGeminiMessages,
consolidateReasoningDetails,
} from "../transform/openai-format"
import { normalizeMistralToolCallId } from "../transform/mistral-format"
import { resolveToolProtocol } from "../../utils/resolveToolProtocol"
import { TOOL_PROTOCOL } from "@roo-code/types"
@ -251,14 +255,23 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
const isNativeProtocol = toolProtocol === TOOL_PROTOCOL.NATIVE
const isGemini = modelId.startsWith("google/gemini")
// For Gemini with native protocol: inject fake reasoning.encrypted block for tool calls
// This is required when switching from other models to Gemini to satisfy API validation.
// Per OpenRouter documentation (conversation with Toven, Nov 2025):
// - Create ONE reasoning_details entry per assistant message with tool calls
// - Set `id` to the FIRST tool call's ID from the tool_calls array
// - Set `data` to "skip_thought_signature_validator" to bypass signature validation
// - Set `index` to 0
// For Gemini models with native protocol:
// 1. Sanitize messages to handle thought signature validation issues.
// This must happen BEFORE fake encrypted block injection to avoid injecting for
// tool calls that will be dropped due to missing/mismatched reasoning_details.
// 2. Inject fake reasoning.encrypted block for tool calls without existing encrypted reasoning.
// This is required when switching from other models to Gemini to satisfy API validation.
// Per OpenRouter documentation (conversation with Toven, Nov 2025):
// - Create ONE reasoning_details entry per assistant message with tool calls
// - Set `id` to the FIRST tool call's ID from the tool_calls array
// - Set `data` to "skip_thought_signature_validator" to bypass signature validation
// - Set `index` to 0
// See: https://github.com/cline/cline/issues/8214
if (isNativeProtocol && isGemini) {
// Step 1: Sanitize messages - filter out tool calls with missing/mismatched reasoning_details
openAiMessages = sanitizeGeminiMessages(openAiMessages, modelId)
// Step 2: Inject fake reasoning.encrypted block for tool calls that survived sanitization
openAiMessages = openAiMessages.map((msg) => {
if (msg.role === "assistant") {
const toolCalls = (msg as any).tool_calls as any[] | undefined
@ -506,9 +519,11 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
}
}
// After streaming completes, store ONLY the reasoning_details we received from the API.
// After streaming completes, consolidate and store reasoning_details from the API.
// This filters out corrupted encrypted blocks (missing `data`) and consolidates by index.
if (reasoningDetailsAccumulator.size > 0) {
this.currentReasoningDetails = Array.from(reasoningDetailsAccumulator.values())
const rawDetails = Array.from(reasoningDetailsAccumulator.values())
this.currentReasoningDetails = consolidateReasoningDetails(rawDetails)
}
if (lastUsage) {

View file

@ -3,7 +3,12 @@
import { Anthropic } from "@anthropic-ai/sdk"
import OpenAI from "openai"
import { convertToOpenAiMessages } from "../openai-format"
import {
convertToOpenAiMessages,
consolidateReasoningDetails,
sanitizeGeminiMessages,
ReasoningDetail,
} from "../openai-format"
import { normalizeMistralToolCallId } from "../mistral-format"
describe("convertToOpenAiMessages", () => {
@ -963,3 +968,338 @@ describe("convertToOpenAiMessages", () => {
})
})
})
describe("consolidateReasoningDetails", () => {
it("should return empty array for empty input", () => {
expect(consolidateReasoningDetails([])).toEqual([])
})
it("should return empty array for undefined input", () => {
expect(consolidateReasoningDetails(undefined as any)).toEqual([])
})
it("should filter out corrupted encrypted blocks (missing data field)", () => {
const details: ReasoningDetail[] = [
{
type: "reasoning.encrypted",
// Missing data field - this should be filtered out
id: "rs_corrupted",
format: "google-gemini-v1",
index: 0,
},
{
type: "reasoning.text",
text: "Valid reasoning",
id: "rs_valid",
format: "google-gemini-v1",
index: 0,
},
]
const result = consolidateReasoningDetails(details)
// Should only have the text block, not the corrupted encrypted block
expect(result).toHaveLength(1)
expect(result[0].type).toBe("reasoning.text")
expect(result[0].text).toBe("Valid reasoning")
})
it("should concatenate text from multiple entries with same index", () => {
const details: ReasoningDetail[] = [
{
type: "reasoning.text",
text: "First part. ",
format: "google-gemini-v1",
index: 0,
},
{
type: "reasoning.text",
text: "Second part.",
format: "google-gemini-v1",
index: 0,
},
]
const result = consolidateReasoningDetails(details)
expect(result).toHaveLength(1)
expect(result[0].text).toBe("First part. Second part.")
})
it("should keep only the last encrypted block per index", () => {
const details: ReasoningDetail[] = [
{
type: "reasoning.encrypted",
data: "first_encrypted_data",
id: "rs_1",
format: "google-gemini-v1",
index: 0,
},
{
type: "reasoning.encrypted",
data: "second_encrypted_data",
id: "rs_2",
format: "google-gemini-v1",
index: 0,
},
]
const result = consolidateReasoningDetails(details)
// Should only have one encrypted block - the last one
expect(result).toHaveLength(1)
expect(result[0].type).toBe("reasoning.encrypted")
expect(result[0].data).toBe("second_encrypted_data")
expect(result[0].id).toBe("rs_2")
})
it("should keep last signature and id from multiple entries", () => {
const details: ReasoningDetail[] = [
{
type: "reasoning.text",
text: "Part 1",
signature: "sig_1",
id: "id_1",
format: "google-gemini-v1",
index: 0,
},
{
type: "reasoning.text",
text: "Part 2",
signature: "sig_2",
id: "id_2",
format: "google-gemini-v1",
index: 0,
},
]
const result = consolidateReasoningDetails(details)
expect(result).toHaveLength(1)
expect(result[0].signature).toBe("sig_2")
expect(result[0].id).toBe("id_2")
})
it("should group by index correctly", () => {
const details: ReasoningDetail[] = [
{
type: "reasoning.text",
text: "Index 0 text",
format: "google-gemini-v1",
index: 0,
},
{
type: "reasoning.text",
text: "Index 1 text",
format: "google-gemini-v1",
index: 1,
},
]
const result = consolidateReasoningDetails(details)
expect(result).toHaveLength(2)
expect(result.find((r) => r.index === 0)?.text).toBe("Index 0 text")
expect(result.find((r) => r.index === 1)?.text).toBe("Index 1 text")
})
it("should handle summary blocks", () => {
const details: ReasoningDetail[] = [
{
type: "reasoning.summary",
summary: "Summary part 1",
format: "google-gemini-v1",
index: 0,
},
{
type: "reasoning.summary",
summary: "Summary part 2",
format: "google-gemini-v1",
index: 0,
},
]
const result = consolidateReasoningDetails(details)
// Summary should be concatenated when there's no text
expect(result).toHaveLength(1)
expect(result[0].summary).toBe("Summary part 1Summary part 2")
})
})
describe("sanitizeGeminiMessages", () => {
it("should return messages unchanged for non-Gemini models", () => {
const messages: OpenAI.Chat.ChatCompletionMessageParam[] = [
{ role: "system", content: "You are helpful" },
{ role: "user", content: "Hello" },
]
const result = sanitizeGeminiMessages(messages, "anthropic/claude-3-5-sonnet")
expect(result).toEqual(messages)
})
it("should drop tool calls without reasoning_details for Gemini models", () => {
const messages = [
{ role: "system", content: "You are helpful" },
{
role: "assistant",
content: "Let me read the file",
tool_calls: [
{
id: "call_123",
type: "function",
function: { name: "read_file", arguments: '{"path":"test.ts"}' },
},
],
// No reasoning_details
},
{ role: "tool", tool_call_id: "call_123", content: "file contents" },
] as OpenAI.Chat.ChatCompletionMessageParam[]
const result = sanitizeGeminiMessages(messages, "google/gemini-3-flash-preview")
// Should have 2 messages: system and assistant (with content but no tool_calls)
// Tool message should be dropped
expect(result).toHaveLength(2)
expect(result[0].role).toBe("system")
expect(result[1].role).toBe("assistant")
expect((result[1] as any).tool_calls).toBeUndefined()
})
it("should filter reasoning_details to only include entries matching tool call IDs", () => {
const messages = [
{
role: "assistant",
content: "",
tool_calls: [
{
id: "call_abc",
type: "function",
function: { name: "read_file", arguments: "{}" },
},
],
reasoning_details: [
{
type: "reasoning.encrypted",
data: "valid_data",
id: "call_abc", // Matches tool call
format: "google-gemini-v1",
index: 0,
},
{
type: "reasoning.encrypted",
data: "mismatched_data",
id: "call_xyz", // Does NOT match any tool call
format: "google-gemini-v1",
index: 1,
},
],
},
] as any
const result = sanitizeGeminiMessages(messages, "google/gemini-3-flash-preview")
expect(result).toHaveLength(1)
const assistantMsg = result[0] as any
expect(assistantMsg.tool_calls).toHaveLength(1)
expect(assistantMsg.reasoning_details).toHaveLength(1)
expect(assistantMsg.reasoning_details[0].id).toBe("call_abc")
})
it("should drop tool calls without matching reasoning_details", () => {
const messages = [
{
role: "assistant",
content: "Some text",
tool_calls: [
{
id: "call_abc",
type: "function",
function: { name: "tool_a", arguments: "{}" },
},
{
id: "call_def",
type: "function",
function: { name: "tool_b", arguments: "{}" },
},
],
reasoning_details: [
{
type: "reasoning.encrypted",
data: "data_for_abc",
id: "call_abc", // Only matches first tool call
format: "google-gemini-v1",
index: 0,
},
],
},
{ role: "tool", tool_call_id: "call_abc", content: "result a" },
{ role: "tool", tool_call_id: "call_def", content: "result b" },
] as any
const result = sanitizeGeminiMessages(messages, "google/gemini-3-flash-preview")
// Should have: assistant with 1 tool_call, 1 tool message
expect(result).toHaveLength(2)
const assistantMsg = result[0] as any
expect(assistantMsg.tool_calls).toHaveLength(1)
expect(assistantMsg.tool_calls[0].id).toBe("call_abc")
// Only the tool result for call_abc should remain
expect(result[1].role).toBe("tool")
expect((result[1] as any).tool_call_id).toBe("call_abc")
})
it("should include reasoning_details without id (legacy format)", () => {
const messages = [
{
role: "assistant",
content: "",
tool_calls: [
{
id: "call_abc",
type: "function",
function: { name: "read_file", arguments: "{}" },
},
],
reasoning_details: [
{
type: "reasoning.text",
text: "Some reasoning without id",
format: "google-gemini-v1",
index: 0,
// No id field
},
{
type: "reasoning.encrypted",
data: "encrypted_data",
id: "call_abc",
format: "google-gemini-v1",
index: 0,
},
],
},
] as any
const result = sanitizeGeminiMessages(messages, "google/gemini-3-flash-preview")
expect(result).toHaveLength(1)
const assistantMsg = result[0] as any
// Both details should be included (one by matching id, one by having no id)
expect(assistantMsg.reasoning_details.length).toBeGreaterThanOrEqual(1)
})
it("should preserve messages without tool_calls", () => {
const messages = [
{ role: "system", content: "You are helpful" },
{ role: "user", content: "Hello" },
{ role: "assistant", content: "Hi there!" },
] as OpenAI.Chat.ChatCompletionMessageParam[]
const result = sanitizeGeminiMessages(messages, "google/gemini-3-flash-preview")
expect(result).toEqual(messages)
})
})

View file

@ -1,6 +1,258 @@
import { Anthropic } from "@anthropic-ai/sdk"
import OpenAI from "openai"
/**
* Type for OpenRouter's reasoning detail elements.
* @see https://openrouter.ai/docs/use-cases/reasoning-tokens#streaming-response
*/
export type ReasoningDetail = {
/**
* Type of reasoning detail.
* @see https://openrouter.ai/docs/use-cases/reasoning-tokens#reasoning-detail-types
*/
type: string // "reasoning.summary" | "reasoning.encrypted" | "reasoning.text"
text?: string
summary?: string
data?: string // Encrypted reasoning data
signature?: string | null
id?: string | null // Unique identifier for the reasoning detail
/**
* Format of the reasoning detail:
* - "unknown" - Format is not specified
* - "openai-responses-v1" - OpenAI responses format version 1
* - "anthropic-claude-v1" - Anthropic Claude format version 1 (default)
* - "google-gemini-v1" - Google Gemini format version 1
* - "xai-responses-v1" - xAI responses format version 1
*/
format?: string
index?: number // Sequential index of the reasoning detail
}
/**
* Consolidates reasoning_details by grouping by index and type.
* - Filters out corrupted encrypted blocks (missing `data` field)
* - For text blocks: concatenates text, keeps last signature/id/format
* - For encrypted blocks: keeps only the last one per index
*
* @param reasoningDetails - Array of reasoning detail objects
* @returns Consolidated array of reasoning details
* @see https://github.com/cline/cline/issues/8214
*/
export function consolidateReasoningDetails(reasoningDetails: ReasoningDetail[]): ReasoningDetail[] {
if (!reasoningDetails || reasoningDetails.length === 0) {
return []
}
// Group by index
const groupedByIndex = new Map<number, ReasoningDetail[]>()
for (const detail of reasoningDetails) {
// Drop corrupted encrypted reasoning blocks that would otherwise trigger:
// "Invalid input: expected string, received undefined" for reasoning_details.*.data
// See: https://github.com/cline/cline/issues/8214
if (detail.type === "reasoning.encrypted" && !detail.data) {
continue
}
const index = detail.index ?? 0
if (!groupedByIndex.has(index)) {
groupedByIndex.set(index, [])
}
groupedByIndex.get(index)!.push(detail)
}
// Consolidate each group
const consolidated: ReasoningDetail[] = []
for (const [index, details] of groupedByIndex.entries()) {
// Concatenate all text parts
let concatenatedText = ""
let concatenatedSummary = ""
let signature: string | undefined
let id: string | undefined
let format = "unknown"
let type = "reasoning.text"
for (const detail of details) {
if (detail.text) {
concatenatedText += detail.text
}
if (detail.summary) {
concatenatedSummary += detail.summary
}
// Keep the signature from the last item that has one
if (detail.signature) {
signature = detail.signature
}
// Keep the id from the last item that has one
if (detail.id) {
id = detail.id
}
// Keep format and type from any item (they should all be the same)
if (detail.format) {
format = detail.format
}
if (detail.type) {
type = detail.type
}
}
// Create consolidated entry for text
if (concatenatedText) {
const consolidatedEntry: ReasoningDetail = {
type: type,
text: concatenatedText,
signature: signature ?? undefined,
id: id ?? undefined,
format: format,
index: index,
}
consolidated.push(consolidatedEntry)
}
// Create consolidated entry for summary (used by some providers)
if (concatenatedSummary && !concatenatedText) {
const consolidatedEntry: ReasoningDetail = {
type: type,
summary: concatenatedSummary,
signature: signature ?? undefined,
id: id ?? undefined,
format: format,
index: index,
}
consolidated.push(consolidatedEntry)
}
// For encrypted chunks (data), only keep the last one
let lastDataEntry: ReasoningDetail | undefined
for (const detail of details) {
if (detail.data) {
lastDataEntry = {
type: detail.type,
data: detail.data,
signature: detail.signature ?? undefined,
id: detail.id ?? undefined,
format: detail.format,
index: index,
}
}
}
if (lastDataEntry) {
consolidated.push(lastDataEntry)
}
}
return consolidated
}
/**
* Sanitizes OpenAI messages for Gemini models by filtering reasoning_details
* to only include entries that match the tool call IDs.
*
* Gemini models require thought signatures for tool calls. When switching providers
* mid-conversation, historical tool calls may not include Gemini reasoning details,
* which can poison the next request. This function:
* 1. Filters reasoning_details to only include entries matching tool call IDs
* 2. Drops tool_calls that lack any matching reasoning_details
* 3. Removes corresponding tool result messages for dropped tool calls
*
* @param messages - Array of OpenAI chat completion messages
* @param modelId - The model ID to check if sanitization is needed
* @returns Sanitized array of messages (unchanged if not a Gemini model)
* @see https://github.com/cline/cline/issues/8214
*/
export function sanitizeGeminiMessages(
messages: OpenAI.Chat.ChatCompletionMessageParam[],
modelId: string,
): OpenAI.Chat.ChatCompletionMessageParam[] {
// Only sanitize for Gemini models
if (!modelId.includes("gemini")) {
return messages
}
const droppedToolCallIds = new Set<string>()
const sanitized: OpenAI.Chat.ChatCompletionMessageParam[] = []
for (const msg of messages) {
if (msg.role === "assistant") {
const anyMsg = msg as any
const toolCalls = anyMsg.tool_calls as OpenAI.Chat.ChatCompletionMessageToolCall[] | undefined
const reasoningDetails = anyMsg.reasoning_details as ReasoningDetail[] | undefined
if (Array.isArray(toolCalls) && toolCalls.length > 0) {
const hasReasoningDetails = Array.isArray(reasoningDetails) && reasoningDetails.length > 0
if (!hasReasoningDetails) {
// No reasoning_details at all - drop all tool calls
for (const tc of toolCalls) {
if (tc?.id) {
droppedToolCallIds.add(tc.id)
}
}
// Keep any textual content, but drop the tool_calls themselves
if (anyMsg.content) {
sanitized.push({ role: "assistant", content: anyMsg.content } as any)
}
continue
}
// Filter reasoning_details to only include entries matching tool call IDs
// This prevents mismatched reasoning details from poisoning the request
const validToolCalls: OpenAI.Chat.ChatCompletionMessageToolCall[] = []
const validReasoningDetails: ReasoningDetail[] = []
for (const tc of toolCalls) {
// Check if there's a reasoning_detail with matching id
const matchingDetails = reasoningDetails.filter((d) => d.id === tc.id)
if (matchingDetails.length > 0) {
validToolCalls.push(tc)
validReasoningDetails.push(...matchingDetails)
} else {
// No matching reasoning_detail - drop this tool call
if (tc?.id) {
droppedToolCallIds.add(tc.id)
}
}
}
// Also include reasoning_details that don't have an id (legacy format)
const detailsWithoutId = reasoningDetails.filter((d) => !d.id)
validReasoningDetails.push(...detailsWithoutId)
// Build the sanitized message
const sanitizedMsg: any = {
role: "assistant",
content: anyMsg.content ?? "",
}
if (validReasoningDetails.length > 0) {
sanitizedMsg.reasoning_details = consolidateReasoningDetails(validReasoningDetails)
}
if (validToolCalls.length > 0) {
sanitizedMsg.tool_calls = validToolCalls
}
sanitized.push(sanitizedMsg)
continue
}
}
if (msg.role === "tool") {
const anyMsg = msg as any
if (anyMsg.tool_call_id && droppedToolCallIds.has(anyMsg.tool_call_id)) {
// Skip tool result for dropped tool call
continue
}
}
sanitized.push(msg)
}
return sanitized
}
/**
* Options for converting Anthropic messages to OpenAI format.
*/