Skip to content

fix: stream reasoning_content in LM Studio provider #1175

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
daewoongoh wants to merge 1 commit into
base: main
Choose a base branch
from
+67 −1
Open

fix: stream reasoning_content in LM Studio provider #1175

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
53 changes: 53 additions & 0 deletions src/api/providers/__tests__/lmstudio.spec.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,6 +112,59 @@ describe("LmStudioHandler", () => {
expect(textChunks[0].text).toBe("Test response")
})

it("streams reasoning chunks from delta.reasoning_content", async () => {
// Regression: Qwen3 / DeepSeek-R1 style models served by LM Studio emit
// thinking via reasoning_content, not <think> tags inside content.
mockCreate.mockImplementationOnce(async () =>
asyncStreamFrom([
{ choices: [{ delta: { reasoning_content: "thinking..." }, index: 0 }] },
{ choices: [{ delta: { content: "answer" }, index: 0 }] },
{
choices: [{ delta: {}, index: 0 }],
usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
},
]),
)

const chunks = await collectStream(handler.createMessage(systemPrompt, messages))

expect(chunks).toContainEqual({ type: "reasoning", text: "thinking..." })
expect(chunks).toContainEqual({ type: "text", text: "answer" })
})

it("falls back to delta.reasoning when reasoning_content is absent", async () => {
mockCreate.mockImplementationOnce(async () =>
asyncStreamFrom([
{ choices: [{ delta: { reasoning: "router-style thought" }, index: 0 }] },
{
choices: [{ delta: {}, index: 0 }],
usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
},
]),
)

const chunks = await collectStream(handler.createMessage(systemPrompt, messages))

expect(chunks).toContainEqual({ type: "reasoning", text: "router-style thought" })
})

it("still parses <think> tags embedded in content", async () => {
mockCreate.mockImplementationOnce(async () =>
asyncStreamFrom([
{ choices: [{ delta: { content: "<think>tagged thought</think>visible" }, index: 0 }] },
{
choices: [{ delta: {}, index: 0 }],
usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
},
]),
)

const chunks = await collectStream(handler.createMessage(systemPrompt, messages))

expect(chunks).toContainEqual({ type: "reasoning", text: "tagged thought" })
expect(chunks).toContainEqual({ type: "text", text: "visible" })
})

@edelauna edelauna Aug 15, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Worth adding a case where both reasoning_content and reasoning are set in the same delta — similar to the test in openai.spec.ts at line 297? That would confirm exactly one reasoning chunk is emitted here (rather than two), which the extractReasoningFromDelta unit tests cover in isolation but this handler does not exercise end-to-end.


it("should handle API errors", async () => {
mockCreate.mockRejectedValueOnce(new Error("API Error"))

Expand Down
15 changes: 14 additions & 1 deletion src/api/providers/lm-studio.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@ import { BaseProvider } from "./base-provider"
import type { SingleCompletionHandler, ApiHandlerCreateMessageMetadata, CompletePromptOptions } from "../index"
import { getModelsFromCache } from "./fetchers/modelCache"
import { handleOpenAIError } from "./utils/error-handler"
import { extractReasoningFromDelta } from "./utils/extract-reasoning"

export class LmStudioHandler extends BaseProvider implements SingleCompletionHandler {
protected options: ApiHandlerOptions
Expand Down Expand Up @@ -80,6 +81,7 @@ export class LmStudioHandler extends BaseProvider implements SingleCompletionHan
}

let assistantText = ""
let reasoningOutput = ""

try {
const params: OpenAI.Chat.ChatCompletionCreateParamsStreaming & { draft_model?: string } = {
Expand Down Expand Up @@ -123,6 +125,15 @@ export class LmStudioHandler extends BaseProvider implements SingleCompletionHan
}
}

// Reasoning models served by LM Studio (Qwen3, DeepSeek-R1, QwQ, ...) stream
// their thinking in a dedicated `reasoning_content`/`reasoning` delta field
// rather than as <think> tags inside `content`, so TagMatcher never sees it.
const reasoningText = extractReasoningFromDelta(delta)
if (reasoningText) {
reasoningOutput += reasoningText
yield { type: "reasoning", text: reasoningText }
}

// Handle tool calls in stream - emit partial chunks for NativeToolCallParser
if (delta?.tool_calls) {
for (const toolCall of delta.tool_calls) {
Expand Down Expand Up @@ -151,7 +162,9 @@ export class LmStudioHandler extends BaseProvider implements SingleCompletionHan

let outputTokens = 0
try {
outputTokens = await this.countTokens([{ type: "text", text: assistantText }])
// Reasoning tokens are billed as output, so count them alongside the
// visible text — otherwise thinking models under-report usage entirely.
outputTokens = await this.countTokens([{ type: "text", text: reasoningOutput + assistantText }])
} catch (err) {
console.error("[LmStudio] Failed to count output tokens:", err)
outputTokens = 0
Expand Down
Loading