Implement accumulate-then-act streaming for thinking models
This commit is contained in:
@@ -17,7 +17,7 @@
|
||||
import { NextResponse } from "next/server";
|
||||
import { requireWorkspacePrincipal } from "@/lib/auth/workspace-auth";
|
||||
import { query, queryOne } from "@/lib/db-postgres";
|
||||
import { callVibnChat } from "@/lib/ai/vibn-chat-model";
|
||||
import { callVibnChat, streamVibnChat } from "@/lib/ai/vibn-chat-model";
|
||||
import {
|
||||
VIBN_TOOL_DEFINITIONS,
|
||||
executeMcpTool,
|
||||
@@ -1165,8 +1165,8 @@ export async function POST(request: Request) {
|
||||
extraSystem += `\n\n[WARNING] You only have ${maxToolRounds - round} tool calls left before you are forcefully terminated. Stop exploring, make your final edits, and write your final response to the user NOW.`;
|
||||
}
|
||||
|
||||
// Execute tool calls and add results. OpenAI-compatible APIs
|
||||
const resp = await callVibnChat({
|
||||
// Execute tool calls and add results using accumulating stream.
|
||||
const stream = streamVibnChat({
|
||||
systemPrompt: systemPrompt + extraSystem,
|
||||
messages,
|
||||
tools: toolDefs,
|
||||
@@ -1175,14 +1175,48 @@ export async function POST(request: Request) {
|
||||
signal: clientSignal,
|
||||
});
|
||||
|
||||
const resp = {
|
||||
text: "",
|
||||
thoughts: "",
|
||||
toolCalls: [] as any[],
|
||||
error: undefined as string | undefined,
|
||||
};
|
||||
|
||||
for await (const chunk of stream) {
|
||||
if (aborted) break;
|
||||
|
||||
if (chunk.type === "thinking_delta" && chunk.text) {
|
||||
resp.thoughts += chunk.text;
|
||||
emit({ type: "thinking_delta", text: chunk.text });
|
||||
} else if (chunk.type === "text_delta" && chunk.text) {
|
||||
resp.text += chunk.text;
|
||||
emit({ type: "text_delta", text: chunk.text });
|
||||
} else if (chunk.type === "tool_calls" && chunk.toolCalls) {
|
||||
resp.toolCalls = chunk.toolCalls;
|
||||
} else if (chunk.type === "error" && chunk.error) {
|
||||
resp.error = chunk.error;
|
||||
}
|
||||
}
|
||||
|
||||
// If the model produced any thoughts or text, record them in the timeline once stream is complete.
|
||||
// (The UI handles the delta-rendering live, but we save the complete chunk to Postgres).
|
||||
if (resp.thoughts) {
|
||||
assistantTimeline.push({ kind: "thought", text: resp.thoughts });
|
||||
}
|
||||
if (resp.text) {
|
||||
assistantText += (assistantText ? "\n\n" : "") + resp.text;
|
||||
assistantTextSegments.push(resp.text);
|
||||
assistantTimeline.push({ kind: "text", text: resp.text });
|
||||
roundsSinceText = 0;
|
||||
toolCallsSinceText = 0;
|
||||
} else if (resp.toolCalls.length) {
|
||||
roundsSinceText++;
|
||||
toolCallsSinceText += resp.toolCalls.length;
|
||||
}
|
||||
|
||||
// When the model first reaches for a mutation, advance the phase so
|
||||
// the UI reflects "Executing Code Edits". We deliberately do NOT force
|
||||
// a separate planning round or discard the edit (the old "C-08
|
||||
// checkpoint" dance) — that made the model plan, stall on an empty
|
||||
// turn, and never execute, and it seeded scope-creep via the forced
|
||||
// "verification plan". The agent edits directly; the post-loop
|
||||
// verification layer checks the result and drives any fixes.
|
||||
const requestedMutations = resp.toolCalls.filter((tc) =>
|
||||
// the UI reflects "Executing Code Edits".
|
||||
const requestedMutations = resp.toolCalls.filter((tc: any) =>
|
||||
[
|
||||
"fs_write",
|
||||
"fs_edit",
|
||||
@@ -1197,10 +1231,7 @@ export async function POST(request: Request) {
|
||||
emit({ type: "phase", phase, label: "Executing Code Edits" });
|
||||
}
|
||||
|
||||
// A Stop click aborts the in-flight generation, which surfaces here
|
||||
// as resp.error === "aborted". Treat it as a clean user stop (break to
|
||||
// the post-loop abort handling that persists the partial reply),
|
||||
// NOT as a fatal error shown to the user.
|
||||
// A Stop click aborts the in-flight generation
|
||||
if (resp.error === "aborted" || aborted) {
|
||||
aborted = true;
|
||||
break;
|
||||
@@ -1212,28 +1243,6 @@ export async function POST(request: Request) {
|
||||
return;
|
||||
}
|
||||
|
||||
// Stream the model's reasoning narration as a separate SSE
|
||||
// event type. We pay for thinking tokens whether or not we
|
||||
// ask for them, so making them visible is free transparency
|
||||
// — and it cures the "tool tray with no narrative" feel.
|
||||
if (resp.thoughts) {
|
||||
assistantTimeline.push({ kind: "thought", text: resp.thoughts });
|
||||
emit({ type: "thinking", text: resp.thoughts });
|
||||
}
|
||||
|
||||
// Stream user-facing text to client.
|
||||
if (resp.text) {
|
||||
assistantText += (assistantText ? "\n\n" : "") + resp.text;
|
||||
assistantTextSegments.push(resp.text);
|
||||
assistantTimeline.push({ kind: "text", text: resp.text });
|
||||
emit({ type: "text", text: resp.text });
|
||||
roundsSinceText = 0;
|
||||
toolCallsSinceText = 0;
|
||||
} else if (resp.toolCalls.length) {
|
||||
roundsSinceText++;
|
||||
toolCallsSinceText += resp.toolCalls.length;
|
||||
}
|
||||
|
||||
// Announce tool calls
|
||||
for (const tc of resp.toolCalls) {
|
||||
assistantToolCalls.push(tc);
|
||||
|
||||
@@ -1281,6 +1281,7 @@ export function ChatPanel({
|
||||
setMessages((prev) => [...prev, userMsg]);
|
||||
|
||||
let assistantContent = "";
|
||||
let lastKind = "";
|
||||
const assistantMsg: Message = { role: "assistant", content: "" };
|
||||
let msgIndex = -1;
|
||||
|
||||
@@ -1402,6 +1403,43 @@ export function ChatPanel({
|
||||
}
|
||||
return next;
|
||||
});
|
||||
} else if (ev.type === "text_delta" && ev.text) {
|
||||
if (lastKind === "text") {
|
||||
assistantContent += ev.text;
|
||||
} else {
|
||||
assistantContent += (assistantContent ? "\n\n" : "") + ev.text;
|
||||
lastKind = "text";
|
||||
}
|
||||
setMessages((prev) => {
|
||||
const next = [...prev];
|
||||
if (msgIndex >= 0 && next[msgIndex]) {
|
||||
const tl = [...(next[msgIndex].timeline ?? [])];
|
||||
const last = tl[tl.length - 1];
|
||||
if (last && last.kind === "text") {
|
||||
tl[tl.length - 1] = { ...last, text: last.text + ev.text };
|
||||
} else {
|
||||
tl.push({ kind: "text", text: ev.text });
|
||||
}
|
||||
next[msgIndex] = { ...next[msgIndex], timeline: tl };
|
||||
}
|
||||
return next;
|
||||
});
|
||||
} else if (ev.type === "thinking_delta" && ev.text) {
|
||||
lastKind = "thought";
|
||||
setMessages((prev) => {
|
||||
const next = [...prev];
|
||||
if (msgIndex >= 0 && next[msgIndex]) {
|
||||
const tl = [...(next[msgIndex].timeline ?? [])];
|
||||
const last = tl[tl.length - 1];
|
||||
if (last && last.kind === "thought") {
|
||||
tl[tl.length - 1] = { ...last, text: last.text + ev.text };
|
||||
} else {
|
||||
tl.push({ kind: "thought", text: ev.text });
|
||||
}
|
||||
next[msgIndex] = { ...next[msgIndex], timeline: tl };
|
||||
}
|
||||
return next;
|
||||
});
|
||||
} else if (ev.type === "thinking" && ev.text) {
|
||||
// Each thinking event from the server is one round of the
|
||||
// model's reasoning. Push as a separate timeline entry so
|
||||
|
||||
@@ -39,10 +39,11 @@ export interface ToolDefinition {
|
||||
}
|
||||
|
||||
export interface ChatChunk {
|
||||
type: "text" | "thinking" | "tool_call" | "done" | "error";
|
||||
type: "text" | "thinking" | "text_delta" | "thinking_delta" | "tool_calls" | "done" | "error";
|
||||
text?: string;
|
||||
toolCall?: ToolCall;
|
||||
toolCalls?: ToolCall[];
|
||||
error?: string;
|
||||
finishReason?: string;
|
||||
}
|
||||
|
||||
type GeminiPart = Record<string, unknown>;
|
||||
@@ -313,17 +314,26 @@ export async function* streamGeminiChat(opts: {
|
||||
if (part.text) {
|
||||
if (isPartThought(part as Record<string, unknown>)) {
|
||||
thoughts += part.text;
|
||||
yield { type: "thinking", text: part.text };
|
||||
yield { type: "thinking_delta", text: part.text };
|
||||
} else {
|
||||
text += part.text;
|
||||
yield { type: "text", text: part.text };
|
||||
yield { type: "text_delta", text: part.text };
|
||||
}
|
||||
}
|
||||
if (part.functionCall) {
|
||||
toolCalls.push(part.functionCall);
|
||||
toolCalls.push({
|
||||
id: `tc-${Date.now()}-${Math.random().toString(36).slice(2)}`,
|
||||
name: part.functionCall.name,
|
||||
args: (part.functionCall.args as Record<string, unknown>) ?? {},
|
||||
thoughtSignature: (part as { thoughtSignature?: string }).thoughtSignature,
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (toolCalls.length > 0) {
|
||||
yield { type: "tool_calls", toolCalls };
|
||||
}
|
||||
const durationMs = Date.now() - startTime;
|
||||
|
||||
logTrainingTelemetryDb({
|
||||
|
||||
@@ -8,7 +8,7 @@
|
||||
* We normalize them to JSON Schema before sending.
|
||||
*/
|
||||
|
||||
import type { ChatMessage, ToolCall, ToolDefinition } from "./gemini-chat";
|
||||
import type { ChatMessage, ToolCall, ToolDefinition, ChatChunk } from "./gemini-chat";
|
||||
|
||||
const DEFAULT_CHAT_URL = "https://api.deepseek.com/chat/completions";
|
||||
|
||||
@@ -389,3 +389,149 @@ export async function callOpenAiCompatibleChat(opts: {
|
||||
|
||||
return { text, thoughts, toolCalls, finishReason };
|
||||
}
|
||||
|
||||
|
||||
export async function* streamOpenAiCompatibleChat(opts: {
|
||||
systemPrompt: string;
|
||||
messages: ChatMessage[];
|
||||
tools?: ToolDefinition[];
|
||||
temperature?: number;
|
||||
includeThoughts?: boolean;
|
||||
signal?: AbortSignal;
|
||||
}): AsyncGenerator<ChatChunk> {
|
||||
const apiKey = resolveApiKey();
|
||||
if (!apiKey) {
|
||||
yield {
|
||||
type: "error",
|
||||
error: "No API key: set DEEPSEEK_API_KEY or VIBN_OPENAI_COMPATIBLE_API_KEY for OpenAI-compatible chat.",
|
||||
};
|
||||
return;
|
||||
}
|
||||
|
||||
const url = resolveChatUrl();
|
||||
const model = resolveModel();
|
||||
const tools = toOpenAiTools(opts.tools);
|
||||
const oaiMessages = toOpenAiMessages(opts.systemPrompt, opts.messages);
|
||||
const body: Record<string, unknown> = {
|
||||
model,
|
||||
messages: oaiMessages,
|
||||
temperature: opts.temperature ?? 0.7,
|
||||
max_tokens: 8192,
|
||||
stream: true,
|
||||
};
|
||||
if (tools?.length) body.tools = tools;
|
||||
|
||||
let res: Response;
|
||||
try {
|
||||
res = await fetch(url, {
|
||||
method: "POST",
|
||||
headers: {
|
||||
"Content-Type": "application/json",
|
||||
Authorization: `Bearer ${apiKey}`,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
signal: opts.signal,
|
||||
});
|
||||
} catch (e) {
|
||||
const aborted = opts.signal?.aborted || (e instanceof Error && e.name === "AbortError");
|
||||
yield {
|
||||
type: "error",
|
||||
error: aborted ? "aborted" : `Network error: ${e instanceof Error ? e.message : String(e)}`,
|
||||
};
|
||||
return;
|
||||
}
|
||||
|
||||
if (!res.ok) {
|
||||
const text = await res.text().catch(() => "");
|
||||
yield { type: "error", error: `Chat API error ${res.status}: ${text}` };
|
||||
return;
|
||||
}
|
||||
|
||||
const reader = res.body?.getReader();
|
||||
if (!reader) {
|
||||
yield { type: "error", error: "No response body stream." };
|
||||
return;
|
||||
}
|
||||
|
||||
const decoder = new TextDecoder("utf-8");
|
||||
let buffer = "";
|
||||
|
||||
// Accumulated tool calls
|
||||
const toolCallsAcc: Record<number, { id: string; name: string; argsStr: string }> = {};
|
||||
|
||||
try {
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) break;
|
||||
|
||||
buffer += decoder.decode(value, { stream: true });
|
||||
const lines = buffer.split("\n");
|
||||
buffer = lines.pop() ?? "";
|
||||
|
||||
for (const line of lines) {
|
||||
const tLine = line.trim();
|
||||
if (!tLine || !tLine.startsWith("data: ")) continue;
|
||||
const dataStr = tLine.slice(6);
|
||||
if (dataStr === "[DONE]") continue;
|
||||
|
||||
try {
|
||||
const parsed = JSON.parse(dataStr);
|
||||
const delta = parsed.choices?.[0]?.delta;
|
||||
if (!delta) continue;
|
||||
|
||||
if (typeof delta.reasoning_content === "string" && delta.reasoning_content.length > 0) {
|
||||
yield { type: "thinking_delta", text: delta.reasoning_content };
|
||||
}
|
||||
if (typeof delta.content === "string" && delta.content.length > 0) {
|
||||
yield { type: "text_delta", text: delta.content };
|
||||
}
|
||||
|
||||
if (delta.tool_calls && Array.isArray(delta.tool_calls)) {
|
||||
for (const tc of delta.tool_calls) {
|
||||
const idx = tc.index;
|
||||
if (idx === undefined) continue;
|
||||
if (!toolCallsAcc[idx]) {
|
||||
toolCallsAcc[idx] = { id: "", name: "", argsStr: "" };
|
||||
}
|
||||
if (tc.id) toolCallsAcc[idx].id = tc.id;
|
||||
if (tc.function?.name) toolCallsAcc[idx].name += tc.function.name;
|
||||
if (tc.function?.arguments) toolCallsAcc[idx].argsStr += tc.function.arguments;
|
||||
}
|
||||
}
|
||||
} catch (e) {
|
||||
// ignore unparseable chunks
|
||||
}
|
||||
}
|
||||
}
|
||||
} catch (e) {
|
||||
const aborted = opts.signal?.aborted || (e instanceof Error && e.name === "AbortError");
|
||||
yield {
|
||||
type: "error",
|
||||
error: aborted ? "aborted" : `Stream read error: ${e instanceof Error ? e.message : String(e)}`,
|
||||
};
|
||||
}
|
||||
|
||||
const toolCalls: ToolCall[] = [];
|
||||
for (const idx of Object.keys(toolCallsAcc).sort((a,b) => Number(a) - Number(b))) {
|
||||
const acc = toolCallsAcc[Number(idx)];
|
||||
let args = {};
|
||||
try {
|
||||
if (acc.argsStr) args = JSON.parse(acc.argsStr);
|
||||
} catch {
|
||||
// ignore bad json
|
||||
}
|
||||
if (acc.name) {
|
||||
toolCalls.push({
|
||||
id: acc.id || `tc-${Date.now()}-${Math.random().toString(36).slice(2)}`,
|
||||
name: acc.name,
|
||||
args
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
if (toolCalls.length > 0) {
|
||||
yield { type: "tool_calls", toolCalls };
|
||||
}
|
||||
|
||||
yield { type: "done" };
|
||||
}
|
||||
|
||||
@@ -13,8 +13,8 @@
|
||||
*/
|
||||
|
||||
import type { ChatMessage, ToolDefinition } from "./gemini-chat";
|
||||
import { callGeminiChat } from "./gemini-chat";
|
||||
import { callOpenAiCompatibleChat } from "./openai-compatible-chat";
|
||||
import { callGeminiChat, streamGeminiChat } from "./gemini-chat";
|
||||
import { callOpenAiCompatibleChat, streamOpenAiCompatibleChat } from "./openai-compatible-chat";
|
||||
|
||||
export type VibnChatCallOpts = {
|
||||
systemPrompt: string;
|
||||
@@ -33,3 +33,13 @@ export async function callVibnChat(opts: VibnChatCallOpts) {
|
||||
}
|
||||
return callGeminiChat(opts);
|
||||
}
|
||||
|
||||
|
||||
export async function* streamVibnChat(opts: VibnChatCallOpts) {
|
||||
const p = (process.env.VIBN_CHAT_PROVIDER || "gemini").toLowerCase().trim();
|
||||
if (p === "deepseek" || p === "openai_compatible") {
|
||||
yield* streamOpenAiCompatibleChat(opts);
|
||||
return;
|
||||
}
|
||||
yield* streamGeminiChat(opts);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user