Response Healing | Plugins | AI Gateway

The response_healing plugin repairs malformed JSON in model output before the response reaches the caller. Models asked for structured output sometimes wrap it in a markdown code fence, add a sentence of explanation around it, or emit a trailing comma. The plugin extracts and repairs the JSON so the response parses.

This feature is in Beta.

Use cases

Getting reliable structured output from models with weaker JSON adherence.
Removing per-service JSON cleanup code from application clients.
Recovering tool calls whose arguments came back slightly malformed.

Quick start

Add a response_healing entry to the plugins array.

curl -X POST https://api.orq.ai/v3/router/chat/completions \
  -H "Authorization: Bearer $ORQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o-mini",
    "messages": [{ "role": "user", "content": "List two colors as JSON." }],
    "response_format": { "type": "json_object" },
    "plugins": [{ "id": "response_healing" }]
  }'

import OpenAI from 'openai';

const client = new OpenAI({
  apiKey: process.env.ORQ_API_KEY,
  baseURL: 'https://api.orq.ai/v3/router',
});

const response = await client.chat.completions.create({
  model: 'openai/gpt-4o-mini',
  messages: [{ role: 'user', content: 'List two colors as JSON.' }],
  response_format: { type: 'json_object' },
  // @ts-ignore - orq.ai extension
  plugins: [{ id: 'response_healing' }],
});

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ.get("ORQ_API_KEY"),
    base_url="https://api.orq.ai/v3/router",
)

response = client.chat.completions.create(
    model="openai/gpt-4o-mini",
    messages=[{"role": "user", "content": "List two colors as JSON."}],
    response_format={"type": "json_object"},
    extra_body={"plugins": [{"id": "response_healing"}]},
)

The plugin takes no configuration. response-healing, the hyphenated spelling used by OpenRouter, is accepted as an alias for response_healing. Payloads migrated from OpenRouter work without edits.

What gets repaired

Problem	Model output	Response
Markdown code fence	JSON wrapped in a ```json block	`{"a": 1}`
JSON embedded in prose	`Here you go: {"a": 1} Hope that helps.`	`{"a": 1}`
Missing closing bracket	`{"a": 1`	`{"a": 1}`
Trailing comma	`{"a": 1,}`	`{"a": 1}`
Unquoted keys	`{a: 1}`	`{"a": 1}`
Single-quoted strings	`{'a': 'b'}`	`{"a": "b"}`

When it applies

The plugin repairs two things:

Message content, only when the request asks for JSON. On /v3/router/chat/completions that means response_format is json_object or json_schema. On /v3/router/responses it means text.format declares type: json_schema, or carries a non-empty inline schema body without an explicit type. Requests that ask for plain text are never modified.
Tool call arguments, whenever the plugin is enabled. Arguments are JSON by definition, so no response_format is required.

Limits

Non-streaming requests only. Streamed chunks reach the caller before the full body is known, so there is nothing left to repair. Setting the plugin on a streaming request has no effect.
Output truncated by max_tokens may be unrecoverable. A response cut mid-string cannot be reconstructed.
Unrepairable output passes through unchanged. The plugin never fails a request, and never replaces valid JSON.

Enable for a workspace

Enable response healing for every request from Settings > Plugins, without passing a plugins array on each call. Once enabled, it applies automatically to every call that doesn’t already specify a response_healing plugin.

Get Started

Models & Providers

Configuration

Observability

AI Gateway Features

Response Healing | Plugins | AI Gateway

Use cases

Quick start

What gets repaired

When it applies

Limits

Enable for a workspace

Supported endpoints

​Use cases

​Quick start

​What gets repaired

​When it applies

​Limits

​Enable for a workspace

​Supported endpoints

Use cases

Quick start

What gets repaired

When it applies

Limits

Enable for a workspace

Supported endpoints