mirror of
https://github.com/RooVetGit/Roo-Code.git
synced 2026-09-06 08:18:39 +00:00
fix: parse XML thinking and tool_call blocks in OpenRouter responses
- Add XML parsing for <think> and <tool_call> blocks in OpenRouter handler - Handle incomplete XML blocks across streaming chunks - Convert tool_call blocks to user-friendly format - Add comprehensive tests for XML parsing functionality Fixes #6630
This commit is contained in:
Roo Code
parent
a88238f68b
commit
aa8cb2f0ad
2 changed files with 259 additions and 1 deletions
|
|
@ -265,6 +265,199 @@ describe("OpenRouterHandler", () => {
|
|||
const generator = handler.createMessage("test", [])
|
||||
await expect(generator.next()).rejects.toThrow("OpenRouter API Error 500: API Error")
|
||||
})
|
||||
|
||||
it("parses <think> blocks correctly", async () => {
|
||||
const handler = new OpenRouterHandler(mockOptions)
|
||||
const mockStream = {
|
||||
async *[Symbol.asyncIterator]() {
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: { content: "Before <think>This is thinking content</think> After" } }],
|
||||
}
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: {} }],
|
||||
usage: { prompt_tokens: 10, completion_tokens: 20 },
|
||||
}
|
||||
},
|
||||
}
|
||||
|
||||
const mockCreate = vitest.fn().mockResolvedValue(mockStream)
|
||||
;(OpenAI as any).prototype.chat = {
|
||||
completions: { create: mockCreate },
|
||||
} as any
|
||||
|
||||
const generator = handler.createMessage("test", [])
|
||||
const chunks = []
|
||||
|
||||
for await (const chunk of generator) {
|
||||
chunks.push(chunk)
|
||||
}
|
||||
|
||||
// Should have 3 text/reasoning chunks and 1 usage chunk
|
||||
expect(chunks).toHaveLength(4)
|
||||
expect(chunks[0]).toEqual({ type: "text", text: "Before " })
|
||||
expect(chunks[1]).toEqual({ type: "reasoning", text: "This is thinking content" })
|
||||
expect(chunks[2]).toEqual({ type: "text", text: " After" })
|
||||
expect(chunks[3]).toEqual({
|
||||
type: "usage",
|
||||
inputTokens: 10,
|
||||
outputTokens: 20,
|
||||
cacheReadTokens: undefined,
|
||||
reasoningTokens: undefined,
|
||||
totalCost: 0,
|
||||
})
|
||||
})
|
||||
|
||||
it("parses <tool_call> blocks correctly", async () => {
|
||||
const handler = new OpenRouterHandler(mockOptions)
|
||||
const mockStream = {
|
||||
async *[Symbol.asyncIterator]() {
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [
|
||||
{ delta: { content: "Text before <tool_call>Tool call content</tool_call> text after" } },
|
||||
],
|
||||
}
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: {} }],
|
||||
usage: { prompt_tokens: 10, completion_tokens: 20 },
|
||||
}
|
||||
},
|
||||
}
|
||||
|
||||
const mockCreate = vitest.fn().mockResolvedValue(mockStream)
|
||||
;(OpenAI as any).prototype.chat = {
|
||||
completions: { create: mockCreate },
|
||||
} as any
|
||||
|
||||
const generator = handler.createMessage("test", [])
|
||||
const chunks = []
|
||||
|
||||
for await (const chunk of generator) {
|
||||
chunks.push(chunk)
|
||||
}
|
||||
|
||||
// Should have 3 text chunks (before, tool call formatted, after) and 1 usage chunk
|
||||
expect(chunks).toHaveLength(4)
|
||||
expect(chunks[0]).toEqual({ type: "text", text: "Text before " })
|
||||
expect(chunks[1]).toEqual({ type: "text", text: "[Tool Call]: Tool call content" })
|
||||
expect(chunks[2]).toEqual({ type: "text", text: " text after" })
|
||||
expect(chunks[3]).toEqual({
|
||||
type: "usage",
|
||||
inputTokens: 10,
|
||||
outputTokens: 20,
|
||||
cacheReadTokens: undefined,
|
||||
reasoningTokens: undefined,
|
||||
totalCost: 0,
|
||||
})
|
||||
})
|
||||
|
||||
it("handles nested and multiple XML blocks", async () => {
|
||||
const handler = new OpenRouterHandler(mockOptions)
|
||||
const mockStream = {
|
||||
async *[Symbol.asyncIterator]() {
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [
|
||||
{
|
||||
delta: {
|
||||
content: "<think>First think</think> middle <tool_call>Tool usage</tool_call>",
|
||||
},
|
||||
},
|
||||
],
|
||||
}
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: { content: " <think>Second think</think> end" } }],
|
||||
}
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: {} }],
|
||||
usage: { prompt_tokens: 10, completion_tokens: 20 },
|
||||
}
|
||||
},
|
||||
}
|
||||
|
||||
const mockCreate = vitest.fn().mockResolvedValue(mockStream)
|
||||
;(OpenAI as any).prototype.chat = {
|
||||
completions: { create: mockCreate },
|
||||
} as any
|
||||
|
||||
const generator = handler.createMessage("test", [])
|
||||
const chunks = []
|
||||
|
||||
for await (const chunk of generator) {
|
||||
chunks.push(chunk)
|
||||
}
|
||||
|
||||
// Verify all chunks are parsed correctly
|
||||
expect(chunks).toContainEqual({ type: "reasoning", text: "First think" })
|
||||
expect(chunks).toContainEqual({ type: "text", text: " middle " })
|
||||
expect(chunks).toContainEqual({ type: "text", text: "[Tool Call]: Tool usage" })
|
||||
expect(chunks).toContainEqual({ type: "text", text: " " })
|
||||
expect(chunks).toContainEqual({ type: "reasoning", text: "Second think" })
|
||||
expect(chunks).toContainEqual({ type: "text", text: " end" })
|
||||
expect(chunks).toContainEqual({
|
||||
type: "usage",
|
||||
inputTokens: 10,
|
||||
outputTokens: 20,
|
||||
cacheReadTokens: undefined,
|
||||
reasoningTokens: undefined,
|
||||
totalCost: 0,
|
||||
})
|
||||
})
|
||||
|
||||
it("handles incomplete XML blocks across chunks", async () => {
|
||||
const handler = new OpenRouterHandler(mockOptions)
|
||||
const mockStream = {
|
||||
async *[Symbol.asyncIterator]() {
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: { content: "Start <thi" } }],
|
||||
}
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: { content: "nk>Thinking content</thi" } }],
|
||||
}
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: { content: "nk> End" } }],
|
||||
}
|
||||
yield {
|
||||
id: "test-id",
|
||||
choices: [{ delta: {} }],
|
||||
usage: { prompt_tokens: 10, completion_tokens: 20 },
|
||||
}
|
||||
},
|
||||
}
|
||||
|
||||
const mockCreate = vitest.fn().mockResolvedValue(mockStream)
|
||||
;(OpenAI as any).prototype.chat = {
|
||||
completions: { create: mockCreate },
|
||||
} as any
|
||||
|
||||
const generator = handler.createMessage("test", [])
|
||||
const chunks = []
|
||||
|
||||
for await (const chunk of generator) {
|
||||
chunks.push(chunk)
|
||||
}
|
||||
|
||||
// Should correctly parse the thinking block even when split across chunks
|
||||
expect(chunks).toContainEqual({ type: "text", text: "Start " })
|
||||
expect(chunks).toContainEqual({ type: "reasoning", text: "Thinking content" })
|
||||
expect(chunks).toContainEqual({ type: "text", text: " End" })
|
||||
expect(chunks).toContainEqual({
|
||||
type: "usage",
|
||||
inputTokens: 10,
|
||||
outputTokens: 20,
|
||||
cacheReadTokens: undefined,
|
||||
reasoningTokens: undefined,
|
||||
totalCost: 0,
|
||||
})
|
||||
})
|
||||
})
|
||||
|
||||
describe("completePrompt", () => {
|
||||
|
|
|
|||
|
|
@ -18,6 +18,7 @@ import { addCacheBreakpoints as addAnthropicCacheBreakpoints } from "../transfor
|
|||
import { addCacheBreakpoints as addGeminiCacheBreakpoints } from "../transform/caching/gemini"
|
||||
import type { OpenRouterReasoningParams } from "../transform/reasoning"
|
||||
import { getModelParams } from "../transform/model-params"
|
||||
import { XmlMatcher } from "../../utils/xml-matcher"
|
||||
|
||||
import { getModels } from "./fetchers/modelCache"
|
||||
import { getModelEndpoints } from "./fetchers/modelEndpointCache"
|
||||
|
|
@ -137,6 +138,7 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
|
|||
const stream = await this.client.chat.completions.create(completionParams)
|
||||
|
||||
let lastUsage: CompletionUsage | undefined = undefined
|
||||
let buffer = ""
|
||||
|
||||
for await (const chunk of stream) {
|
||||
// OpenRouter returns an error object instead of the OpenAI SDK throwing an error.
|
||||
|
|
@ -153,7 +155,65 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
|
|||
}
|
||||
|
||||
if (delta?.content) {
|
||||
yield { type: "text", text: delta.content }
|
||||
buffer += delta.content
|
||||
|
||||
// Process complete XML blocks
|
||||
let processed = true
|
||||
while (processed) {
|
||||
processed = false
|
||||
|
||||
// Check for complete <think> blocks
|
||||
const thinkMatch = buffer.match(/^(.*?)<think>([\s\S]*?)<\/think>(.*)$/s)
|
||||
if (thinkMatch) {
|
||||
const [, before, content, after] = thinkMatch
|
||||
if (before) {
|
||||
yield { type: "text", text: before }
|
||||
}
|
||||
yield { type: "reasoning", text: content }
|
||||
buffer = after
|
||||
processed = true
|
||||
continue
|
||||
}
|
||||
|
||||
// Check for complete <tool_call> blocks
|
||||
const toolMatch = buffer.match(/^(.*?)<tool_call>([\s\S]*?)<\/tool_call>(.*)$/s)
|
||||
if (toolMatch) {
|
||||
const [, before, content, after] = toolMatch
|
||||
if (before) {
|
||||
yield { type: "text", text: before }
|
||||
}
|
||||
yield { type: "text", text: `[Tool Call]: ${content}` }
|
||||
buffer = after
|
||||
processed = true
|
||||
continue
|
||||
}
|
||||
|
||||
// Check if we have an incomplete tag at the end
|
||||
const incompleteTag = buffer.match(/^(.*?)(<(?:think|tool_call)[^>]*(?:>[\s\S]*)?)?$/s)
|
||||
if (incompleteTag && incompleteTag[2]) {
|
||||
// We have an incomplete tag, yield the text before it and keep the tag in buffer
|
||||
const [, before, tag] = incompleteTag
|
||||
if (before) {
|
||||
yield { type: "text", text: before }
|
||||
buffer = tag
|
||||
}
|
||||
break
|
||||
}
|
||||
|
||||
// No tags found or incomplete, yield all content except potential start of a tag
|
||||
const tagStart = buffer.lastIndexOf("<")
|
||||
if (tagStart === -1) {
|
||||
// No < found, yield all
|
||||
if (buffer) {
|
||||
yield { type: "text", text: buffer }
|
||||
buffer = ""
|
||||
}
|
||||
} else if (tagStart > 0) {
|
||||
// Yield content before the <
|
||||
yield { type: "text", text: buffer.substring(0, tagStart) }
|
||||
buffer = buffer.substring(tagStart)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (chunk.usage) {
|
||||
|
|
@ -161,6 +221,11 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
|
|||
}
|
||||
}
|
||||
|
||||
// Process any remaining content in the buffer
|
||||
if (buffer) {
|
||||
yield { type: "text", text: buffer }
|
||||
}
|
||||
|
||||
if (lastUsage) {
|
||||
yield {
|
||||
type: "usage",
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue