Files
deer-flow/frontend/tests/unit/core/messages/utils.test.ts
T
AochenShen99 11dd5b0683 fix(frontend): strip unclosed <think> tags from streaming AI content (#3218)
* fix(frontend): strip unclosed <think> tags from streaming AI content

During streaming, an opening <think> tag may arrive in one chunk
while the matching </think> arrives in a later chunk. The existing
splitInlineReasoning regex only matched fully closed pairs, so the
mid-flight reasoning was left in message.content and rendered into
the chat bubble via the markdown pipeline's rehypeRaw plugin until
the closing tag landed.

Extend splitInlineReasoning with a second pass: after stripping every
closed <think>...</think> pair, route any remaining content from a
lone opener to the reasoning slot and leave only the preceding
preamble in content. Closed-tag behavior is unchanged.

Covers every provider whose stream emits reasoning inline as <think>
tags (MiniMax streaming path, MindIE, PatchedChatOpenAI, and any
gateway-served DeepSeek/OpenAI-compatible model).

* style(frontend): apply prettier formatting to streaming reasoning tests

* fix(frontend): skip <think> split for literal think tags in inline code

Treats a `<think>` opener immediately preceded by a backtick as part of
markdown inline code rather than a streaming reasoning marker. Prevents
permanent content truncation when an AI message documents the `<think>`
tag literally (e.g. ``Use `<think>` markers``), where the streaming-safe
fallback would otherwise route the rest of the answer into the reasoning
panel because no `</think>` ever arrives.

Adds regression tests for both the post-stream and mid-stream cases.
2026-05-26 09:35:07 +08:00

478 lines
13 KiB
TypeScript

import type { Message } from "@langchain/langgraph-sdk";
import { describe, expect, test } from "vitest";
import {
extractContentFromMessage,
extractReasoningContentFromMessage,
getAssistantTurnCopyData,
getAssistantTurnUsageMessages,
getMessageGroups,
getStreamingMessageLookup,
hasContent,
hasReasoning,
isAssistantMessageGroupStreaming,
} from "@/core/messages/utils";
function aiMessage(content: string): Message {
return {
id: "ai-1",
type: "ai",
content,
} as Message;
}
test("aggregates token usage messages once per assistant turn", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Plan a trip",
},
{
id: "ai-1",
type: "ai",
content: "",
tool_calls: [{ id: "tool-1", name: "web_search", args: {} }],
usage_metadata: { input_tokens: 10, output_tokens: 5, total_tokens: 15 },
},
{
id: "tool-1-result",
type: "tool",
name: "web_search",
tool_call_id: "tool-1",
content: "[]",
},
{
id: "ai-2",
type: "ai",
content: "Here is the itinerary",
usage_metadata: { input_tokens: 2, output_tokens: 8, total_tokens: 10 },
},
{
id: "human-2",
type: "human",
content: "Make it shorter",
},
{
id: "ai-3",
type: "ai",
content: "Short version",
usage_metadata: { input_tokens: 1, output_tokens: 1, total_tokens: 2 },
},
] as Message[];
const groups = getMessageGroups(messages);
const usageMessagesByGroupIndex = getAssistantTurnUsageMessages(groups);
expect(groups.map((group) => group.type)).toEqual([
"human",
"assistant:processing",
"assistant",
"human",
"assistant",
]);
expect(
usageMessagesByGroupIndex.map(
(groupMessages) => groupMessages?.map((message) => message.id) ?? null,
),
).toEqual([null, null, ["ai-1", "ai-2"], null, ["ai-3"]]);
});
describe("inline <think> tag splitting", () => {
test("strips a fully closed <think> block from AI content", () => {
const message = aiMessage("<think>internal reasoning</think>final answer");
expect(extractContentFromMessage(message)).toBe("final answer");
expect(extractReasoningContentFromMessage(message)).toBe(
"internal reasoning",
);
});
test("strips multiple closed <think> blocks and joins their reasoning", () => {
const message = aiMessage(
"<think>step one</think>between<think>step two</think>after",
);
expect(extractContentFromMessage(message)).toBe("betweenafter");
expect(extractReasoningContentFromMessage(message)).toBe(
"step one\n\nstep two",
);
});
test("during streaming, an unclosed <think> tag does not leak its tail into content", () => {
// Simulates accumulated content mid-stream, before </think> arrives.
const message = aiMessage(
"<think>I need to analyze the user's question step by",
);
expect(extractContentFromMessage(message)).toBe("");
expect(extractContentFromMessage(message)).not.toContain("<think>");
expect(extractReasoningContentFromMessage(message)).toBe(
"I need to analyze the user's question step by",
);
});
test("preamble before an unclosed <think> stays in content", () => {
const message = aiMessage(
"Here is part of the answer.<think>but wait, let me reconsider",
);
expect(extractContentFromMessage(message)).toBe(
"Here is part of the answer.",
);
expect(extractReasoningContentFromMessage(message)).toBe(
"but wait, let me reconsider",
);
});
test("closed <think> followed by a trailing unclosed <think> merges both into reasoning", () => {
const message = aiMessage(
"<think>first step</think>partial answer<think>second step still streaming",
);
expect(extractContentFromMessage(message)).toBe("partial answer");
expect(extractReasoningContentFromMessage(message)).toBe(
"first step\n\nsecond step still streaming",
);
});
test("hasReasoning recognises an unclosed <think> tag mid-stream", () => {
expect(hasReasoning(aiMessage("<think>thinking in progress"))).toBe(true);
});
test("hasContent excludes an unclosed <think> tail when no preamble exists", () => {
expect(hasContent(aiMessage("<think>thinking in progress"))).toBe(false);
});
test("hasContent stays true when preamble precedes an unclosed <think>", () => {
expect(hasContent(aiMessage("preamble<think>still thinking"))).toBe(true);
});
test("a lone <think> open tag with no body yields no reasoning and no content", () => {
const message = aiMessage("<think>");
expect(extractContentFromMessage(message)).toBe("");
expect(extractReasoningContentFromMessage(message)).toBeNull();
expect(hasReasoning(message)).toBe(false);
});
test("a literal <think> inside markdown inline code is not treated as reasoning", () => {
const message = aiMessage(
"Use `<think>` markers to delimit reasoning sections.",
);
expect(extractContentFromMessage(message)).toBe(
"Use `<think>` markers to delimit reasoning sections.",
);
expect(extractReasoningContentFromMessage(message)).toBeNull();
expect(hasReasoning(message)).toBe(false);
});
test("a backtick-prefixed <think> mid-stream is not split into reasoning", () => {
// Simulates the moment the model has emitted the opening backtick and
// `<think>` for a literal documentation reference, before the closing
// backtick arrives. The pre-fix behaviour would have permanently
// truncated the content here.
const message = aiMessage("Documentation: `<think>");
expect(extractContentFromMessage(message)).toBe("Documentation: `<think>");
expect(extractReasoningContentFromMessage(message)).toBeNull();
});
});
test("hides internal todo reminder messages from message groups", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Audit the middleware",
},
{
id: "todo-reminder-1",
type: "human",
name: "todo_completion_reminder",
content: "<system_reminder>finish todos</system_reminder>",
},
{
id: "todo-reminder-2",
type: "human",
name: "todo_reminder",
content: "<system_reminder>remember todos</system_reminder>",
},
{
id: "ai-1",
type: "ai",
content: "Done",
},
] as Message[];
const groups = getMessageGroups(messages);
expect(groups.map((group) => group.type)).toEqual(["human", "assistant"]);
expect(
groups.flatMap((group) => group.messages).map((message) => message.id),
).toEqual(["human-1", "ai-1"]);
});
test("hides assistant copy data while that turn is streaming", () => {
const messages = [
{
id: "ai-1",
type: "ai",
content: "Partial answer",
},
] as Message[];
expect(getAssistantTurnCopyData(messages)).toBe("Partial answer");
expect(getAssistantTurnCopyData(messages, { isStreaming: true })).toBeNull();
});
test("marks the latest assistant message as streaming", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Hello",
},
{
id: "ai-1",
type: "ai",
content: "Still generating",
},
] as Message[];
const groups = getMessageGroups(messages);
const assistantGroupIndex = groups.findIndex(
(group) => group.type === "assistant",
);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndex]?.messages ?? [],
getStreamingMessageLookup(messages, true, () => ({
streamMetadata: { langgraph_node: "agent" },
})),
),
).toBe(true);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndex]?.messages ?? [],
getStreamingMessageLookup(messages, false, () => ({
streamMetadata: { langgraph_node: "agent" },
})),
),
).toBe(false);
});
test("keeps previous assistant copyable while waiting for a new visible answer", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Hello",
},
{
id: "ai-1",
type: "ai",
content: "Completed answer",
},
{
id: "opt-human-1",
type: "human",
content: "Continue",
},
] as Message[];
const groups = getMessageGroups(messages);
const assistantGroupIndex = groups.findIndex(
(group) => group.type === "assistant",
);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndex]?.messages ?? [],
getStreamingMessageLookup(messages, true),
),
).toBe(false);
});
test("keeps previous assistant copyable while a hidden send is starting", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Hello",
},
{
id: "ai-1",
type: "ai",
content: "Completed answer",
},
] as Message[];
const groups = getMessageGroups(messages);
const assistantGroupIndex = groups.findIndex(
(group) => group.type === "assistant",
);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndex]?.messages ?? [],
getStreamingMessageLookup(messages, true),
),
).toBe(false);
});
test("keeps previous assistant copyable after a hidden send is appended", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Hello",
},
{
id: "ai-1",
type: "ai",
content: "Completed answer",
},
{
id: "human-hidden",
type: "human",
content: "Save this agent",
additional_kwargs: { hide_from_ui: true },
},
] as Message[];
const groups = getMessageGroups(messages);
const assistantGroupIndex = groups.findIndex(
(group) => group.type === "assistant",
);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndex]?.messages ?? [],
getStreamingMessageLookup(messages, true),
),
).toBe(false);
});
test("uses stream metadata to identify an assistant before optimistic input", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Hello",
},
{
id: "ai-1",
type: "ai",
content: "Completed answer",
},
{
id: "ai-2",
type: "ai",
content: "Still generating",
},
{
id: "opt-human-1",
type: "human",
content: "Continue",
},
] as Message[];
const assistantGroups = getMessageGroups(messages).filter(
(group) => group.type === "assistant",
);
const groups = getMessageGroups(messages);
const assistantGroupIndexes = groups
.map((group, index) => (group.type === "assistant" ? index : -1))
.filter((index) => index >= 0);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndexes[0] ?? -1]?.messages ?? [],
getStreamingMessageLookup(messages, true, (message) =>
message.id === "ai-2"
? { streamMetadata: { langgraph_node: "agent" } }
: undefined,
),
),
).toBe(false);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndexes[1] ?? -1]?.messages ?? [],
getStreamingMessageLookup(messages, true, (message) =>
message.id === "ai-2"
? { streamMetadata: { langgraph_node: "agent" } }
: undefined,
),
),
).toBe(true);
expect(assistantGroups.map((group) => group.id)).toEqual(["ai-1", "ai-2"]);
});
test("does not mark a completed assistant group streaming from a later processing group", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Hello",
},
{
id: "ai-1",
type: "ai",
content: "Visible answer",
},
{
id: "ai-2",
type: "ai",
content: "",
tool_calls: [{ id: "tool-1", name: "web_search", args: {} }],
},
] as Message[];
const groups = getMessageGroups(messages);
const assistantGroupIndex = groups.findIndex(
(group) => group.type === "assistant",
);
expect(groups.map((group) => group.type)).toEqual([
"human",
"assistant",
"assistant:processing",
]);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndex]?.messages ?? [],
getStreamingMessageLookup(messages, true, (message) =>
message.id === "ai-2"
? { streamMetadata: { langgraph_node: "agent" } }
: undefined,
),
),
).toBe(false);
});
test("keeps streaming assistant hidden when a hidden control message follows it", () => {
const messages = [
{
id: "human-1",
type: "human",
content: "Hello",
},
{
id: "ai-1",
type: "ai",
content: "Still generating",
},
{
id: "human-hidden",
type: "human",
content: "Save this agent",
additional_kwargs: { hide_from_ui: true },
},
] as Message[];
const groups = getMessageGroups(messages);
const assistantGroupIndex = groups.findIndex(
(group) => group.type === "assistant",
);
expect(
isAssistantMessageGroupStreaming(
groups[assistantGroupIndex]?.messages ?? [],
getStreamingMessageLookup(messages, true, (message) =>
message.id === "ai-1"
? { streamMetadata: { langgraph_node: "agent" } }
: undefined,
),
),
).toBe(true);
});