Response Healing

Not yet available. Response Healing is not yet available.

Models occasionally produce outputs that are structurally wrong — a JSON block that isn't valid JSON, a Markdown table with a missing column, a response that stops mid-sentence because max_tokens was hit. Response Healing is a response-phase plugin that catches these cases and attempts a repair before your application ever sees the output.

Enabling response healing#

{
  "model": "meta-llama/llama-4-maverick",
  "messages": [
    {
      "role": "user",
      "content": "Return a JSON array of 5 capital cities in Africa with their populations."
    }
  ],
  "plugins": [
    { "type": "opendunes.response_healing" }
  ]
}

What healing detects#

Healing runs synchronously. If repair succeeds, the corrected response is returned as normal. If repair fails after the configured retry budget, the original (broken) response is returned with an X-Healing-Failed: true header.

Configuration#

{
  "type": "opendunes.response_healing",
  "options": {
    "targets": ["json", "truncation"],
    "max_retries": 2,
    "strict": false
  }
}

Response headers#

• X-Healing-Applied: true — at least one repair was made.
• X-Healing-Failed: true — healing was attempted but the result is still invalid (only when strict: false).

Billing#

Each healing retry is a fresh model invocation billed at the original model's DA rate. Structural repairs that don't require a retry (e.g., closing a brace) have no additional token cost.