Appopt-chat example

AI Agent Ops App

opt-chat의 Conversation, Agent, Tool, Queue, Context, Sources를 묶고 agent memory, approval-gated EAV patch, rollback audit, replay audit export, eval suite, release gate, brandapp-sdk AI route를 전제로 한 운영 에이전트 앱입니다.

Full agent console with streaming chat, model routing, persistent agent memory, guarded EAV patch application, rollback audit packets, replay export, runnable evals, release gates, source panels, and BrandApp provider wiring.

appagenttoolauditevalsreleasebrandapp-sdk

bun add @reopt-ai/opt-chat @reopt-ai/brandapp-sdk ai

ConversationPromptInputAgentToolQueueContextSourcesModelSelectorcreateBrandappProvider

1024px

onlineanthropic/claude-sonnet-4.5

Launch reliability agent

AgentOps is connected to BrandApp AI, EAV records, deployment notes, and the approval queue. Pick a runbook or ask for a status sweep.

What is blocking the spring launches right now?

Three launches are blocked: two are waiting on legal copy and one is missing a BrandApp EAV segment mapping. Owners are assigned on all three. I drafted a guarded patch you can review in the Approval gate before any EAV write runs.

tool guarded

Launch Reliabilitybrandapp.ai

Checks release notes, BrandApp EAV records, credits, and approval state before changing customer-facing data.

지침

Never execute a destructive tool without explicit approval.

출력

Safe actions queued. No write tools executed.

Agent memory

3 facts

Guardrail

success

Writes require approvalToken; destructive tools stay draft-only.

Launch context

info

3 blockers: legal copy, owner note, missing segment.

Credit policy

warning

Cap image variants while credit load stays above 60%.

Approval gate

Owner approval required before the write tool can update EAV.

approval pending

Guarded patch preview

The agent can preview the EAV mutation but cannot write yet.

draft patch

launch.status

patch

Blocked to Ready

owner_note

patch

Needs copy review to Legal copy approved; launch can proceed

eav.segment

patch

missing to vip-spring-2026

Audit packet

Bundle approval, model route, source list, memory, EAV diff, and rollback chain.

export ready

agentops-audit-2026-06-19.json

approvalToken hash · model route · sources · rollback chain

Eval suite

Run guardrail, EAV mutation, source, and rollback replay checks.

ready

release-gate.agentops

approval policy · source grounding · EAV mutation · replay

Release gate

Export the audit packet and pass evals before release can open.

blocked

audit packetpending

eval suitepending

Run history

2 events

Guard policy loaded

success

Write tools require approvalToken before EAV mutation.

Last runbook passed

info

Blocked launch sweep completed 4 minutes ago.

Loaded BrandApp context

5 entities, 12 active launch records

Checked credit budget

64% monthly usage, no hard limit risk

Prepare guarded write

Owner approval required before the write tool can update EAV.

status:blocked owner:* launch

Spring collection launch
Referral reminder
VIP early access

매개변수

json

{
  "entity": "launches",
  "status": "blocked",
  "runbook": "Summarize blocked launches"
}

결과

json

{
  "count": 3,
  "approval": "pending",
  "riskyWrite": true,
  "patch": "draft",
  "next": "request_owner_approval"
}

Summarize blockers

Prepare owner approval

Write approved patch

Prepare rollback packet

Open release gate

33K / 128K tokens

33K 토큰128K

입력 토큰18K

출력 토큰2.2K

추론 토큰7K

캐시 토큰5.1K

BrandApp launch recordslaunches.eav

Current credit ledgerai-credits.json

Agent route

Terminal

$ POST /api/agent provider=brandapp.ai status=streaming model=anthropic/claude-sonnet-4.5 audit=idle release=blocked

테스트 결과

2 통과

tool approval policy12ms

EAV read route18ms

EAV write mutation21ms

rollback audit15ms

audit packet11ms

eval suite17ms

release gate13ms

write guard9ms

// app/api/agent/route.ts
import { createBrandappProvider } from "@reopt-ai/brandapp-sdk/ai-provider";
import { convertToModelMessages, streamText } from "ai";

let brandapp: ReturnType<typeof createBrandappProvider> | undefined;

function getBrandapp() {
  brandapp ??= createBrandappProvider({
    clientId: process.env.BRANDAPP_CLIENT_ID!,
    clientSecret: process.env.BRANDAPP_CLIENT_SECRET!,
    brandappId: process.env.BRANDAPP_ID!,
  });
  return brandapp;
}

export async function POST(req: Request) {
  const {
    messages,
    model = "anthropic/claude-sonnet-4.5",
    approvalToken,
  } = await req.json();
  const mode = approvalToken ? "approved write mode" : "read-only mode";
  const result = streamText({
    model: getBrandapp()(model),
    system:
      "You are an operations agent. Never write EAV without approvalToken. " +
      "Current mode: " + mode,
    messages: await convertToModelMessages(messages),
  });
  return result.toUIMessageStreamResponse();
}

Related Examples

Agent Surface

Agent, Queue, Artifact, Terminal, TestResults in one workspace

회귀 분석을 마친 뒤 코드 패치와 검증 결과를 함께 보여주는 agent workspace 예제입니다.

Release Investigatorgpt-5.4

rollout 이후 billing, cache, relay 경로를 우선 조사합니다.

지침

에러율, latency, failed jobs를 묶어서 root cause 후보를 좁히고 필요한 패치를 제안합니다.

출력

billing-sync latency regression identified. queue drain patch generated.

Latency diff analysed

Failing routes isolated

release-summary-card.tsx

Generated patch preview

tsx

export function ReleaseSummaryCard() {
  return (
    <section className="rounded-2xl border p-5 shadow-sm">
      <h2 className="text-lg font-semibold">Release Summary</h2>
      <p className="mt-2 text-sm text-muted-foreground">
        Billing sync latency increased after the latest rollout.
      </p>
    </section>
  );
}

Agentopt-chat

Agent Workspace

Agent, Queue, Artifact, Terminal, TestResults를 묶어 multi-step agent run surface를 구성하는 예제입니다.

agentartifactqueue

AgentQueueArtifactCodeBlock+2

Open

Approval UX

Confirmation compound around a risky tool call

프로덕션 마이그레이션을 실행하려고 합니다. 실행 전에 사용자의 명시적 승인을 요청합니다.

매개변수

json

{
  "target": "production",
  "migration": "20260410_add_release_flags",
  "dryRun": false
}

`run_migration` 도구를 실행할까요?

대상은 production database이며, 롤백 플랜 없이 바로 실행됩니다.

승인 대기 중입니다. 이 상태에서 기본 승인/거부 버튼을 그대로 사용할 수 있습니다.

Approvalopt-chat

Approval Flow

Confirmation compound로 위험한 도구 호출의 승인 대기, 승인 완료, 거부 상태를 UX로 나눈 예제입니다.

confirmationtoolworkflow

ConfirmationConfirmationActionsConfirmationRequestConfirmationAccepted+6

Open

Release Soak

Install smoke · fresh app import/render/build

릴리즈 후보 스모크 결과 요약

매개변수

json

{
  "package": "@reopt-ai/opt-chat",
  "consumers": [
    "next-app",
    "vite-app"
  ],
  "imports": [
    ".",
    "./flow",
    "./styles.css"
  ]
}

결과

json

{
  "status": "passed",
  "checks": [
    "root import renders Conversation",
    "flow subpath stays optional-peer safe",
    "production build resolves workspace dependency ranges"
  ]
}

Install smoke가 통과하면 packaging 축은 release gate를 통과합니다.

Readinessopt-chat

Release Soak

install smoke, browser matrix, streaming soak를 하나의 결정적 fixture로 묶어 릴리즈 직전 검증 흐름을 보여줍니다.

releasesoakproduction

ConversationConversationContentMessageMessageContent+5

Open

// app/api/agent/route.ts import { createBrandappProvider } from "@reopt-ai/brandapp-sdk/ai-provider"; import { convertToModelMessages, streamText } from "ai"; let brandapp: ReturnType<typeof createBrandappProvider> | undefined; function getBrandapp() { brandapp ??= createBrandappProvider({ clientId: process.env.BRANDAPP_CLIENT_ID!, clientSecret: process.env.BRANDAPP_CLIENT_SECRET!, brandappId: process.env.BRANDAPP_ID!, }); return brandapp; } export async function POST(req: Request) { const { messages, model = "anthropic/claude-sonnet-4.5", approvalToken, } = await req.json(); const mode = approvalToken ? "approved write mode" : "read-only mode"; const result = streamText({ model: getBrandapp()(model), system: "You are an operations agent. Never write EAV without approvalToken. " + "Current mode: " + mode, messages: await convertToModelMessages(messages), }); return result.toUIMessageStreamResponse(); }

Related Examples

Agent Surface

Agent, Queue, Artifact, Terminal, TestResults in one workspace

회귀 분석을 마친 뒤 코드 패치와 검증 결과를 함께 보여주는 agent workspace 예제입니다.

Release Investigatorgpt-5.4

rollout 이후 billing, cache, relay 경로를 우선 조사합니다.

지침

에러율, latency, failed jobs를 묶어서 root cause 후보를 좁히고 필요한 패치를 제안합니다.

출력

billing-sync latency regression identified. queue drain patch generated.

Latency diff analysed

Failing routes isolated

release-summary-card.tsx

Generated patch preview

tsx

export function ReleaseSummaryCard() {
  return (
    <section className="rounded-2xl border p-5 shadow-sm">
      <h2 className="text-lg font-semibold">Release Summary</h2>
      <p className="mt-2 text-sm text-muted-foreground">
        Billing sync latency increased after the latest rollout.
      </p>
    </section>
  );
}

Agentopt-chat

Agent Workspace

Agent, Queue, Artifact, Terminal, TestResults를 묶어 multi-step agent run surface를 구성하는 예제입니다.

agentartifactqueue

AgentQueueArtifactCodeBlock+2

Open

Approval UX

Confirmation compound around a risky tool call

프로덕션 마이그레이션을 실행하려고 합니다. 실행 전에 사용자의 명시적 승인을 요청합니다.

매개변수

json

{
  "target": "production",
  "migration": "20260410_add_release_flags",
  "dryRun": false
}

`run_migration` 도구를 실행할까요?

대상은 production database이며, 롤백 플랜 없이 바로 실행됩니다.

승인 대기 중입니다. 이 상태에서 기본 승인/거부 버튼을 그대로 사용할 수 있습니다.

Approvalopt-chat

Approval Flow

Confirmation compound로 위험한 도구 호출의 승인 대기, 승인 완료, 거부 상태를 UX로 나눈 예제입니다.

confirmationtoolworkflow

ConfirmationConfirmationActionsConfirmationRequestConfirmationAccepted+6

Open

Release Soak

Install smoke · fresh app import/render/build

릴리즈 후보 스모크 결과 요약

매개변수

json

{
  "package": "@reopt-ai/opt-chat",
  "consumers": [
    "next-app",
    "vite-app"
  ],
  "imports": [
    ".",
    "./flow",
    "./styles.css"
  ]
}

결과

json

{
  "status": "passed",
  "checks": [
    "root import renders Conversation",
    "flow subpath stays optional-peer safe",
    "production build resolves workspace dependency ranges"
  ]
}

Install smoke가 통과하면 packaging 축은 release gate를 통과합니다.

Readinessopt-chat

Release Soak

install smoke, browser matrix, streaming soak를 하나의 결정적 fixture로 묶어 릴리즈 직전 검증 흐름을 보여줍니다.

releasesoakproduction

ConversationConversationContentMessageMessageContent+5

Open