mark/vibn-agent-runner
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Implement accumulate-then-act streaming for thinking models

Browse Source
This commit is contained in:
mawkone
2026-06-15 12:13:46 -07:00
parent a0e2364481
commit 86da778721
5 changed files with 257 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

81
vibn-frontend/app/api/chat/route.ts
View File

@@ -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);