ccr-plugin-structured-outputs

A claude-code-router plugin that restores structured outputs (JSON Schema) when routing Anthropic Messages API requests to OpenAI-compatible backends (GLM, DeepSeek, Groq, vLLM, Ollama, etc.).

The problem

Anthropic's Messages API gained structured outputs in GA on 2025-11-14. A request looks like:

{
  "model": "claude-opus-4-7",
  "messages": [{ "role": "user", "content": "..." }],
  "output_config": {
    "format": {
      "type": "json_schema",
      "schema": { "type": "object", "properties": { ... }, "required": [...] }
    }
  }
}

claude-code-router (CCR) routes Anthropic-format requests to OpenAI-compatible backends, but its built-in AnthropicTransformer silently drops both output_config (GA) and output_format (beta). The backend never sees the schema, and the response degrades to free-form text — defeating the whole point of structured outputs.

This plugin closes the gap by translating either field into OpenAI's response_format, which OpenAI-compatible backends already understand.

An upstream fix is proposed in musistudio/claude-code-router#1380. Use this plugin while waiting for it (or indefinitely — the plugin is small enough that it has no real maintenance burden).

Install

1. Drop outputconfig.transformer.js into your CCR plugins directory (typically ~/.claude-code-router/plugins/):

   curl -fsSL -o ~/.claude-code-router/plugins/outputconfig.transformer.js \
     https://raw.githubusercontent.com/idemerge/ccr-plugin-structured-outputs/main/outputconfig.transformer.js

2. Register the plugin and enable it on your provider in ~/.claude-code-router/config.json:

   {
     "transformers": [
       {
         "name": "outputConfig",
         "path": "/root/.claude-code-router/plugins/outputconfig.transformer.js",
         "options": {}
       }
     ],
     "Providers": [
       {
         "name": "GLM-5",
         "api_base_url": "https://your-backend/v1/chat/completions",
         "api_key": "...",
         "models": ["glm-5"],
         "transformer": {
           "use": ["outputConfig"]
         }
       }
     ]
   }

3. Restart CCR.

Verify

curl -s http://127.0.0.1:3456/v1/messages \
  -H 'Content-Type: application/json' \
  -H 'x-api-key: YOUR_CCR_KEY' \
  -d '{
    "model": "glm-5",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "List 3 EU capitals."}],
    "output_config": {
      "format": {
        "type": "json_schema",
        "schema": {
          "type": "object",
          "additionalProperties": false,
          "required": ["capitals"],
          "properties": {
            "capitals": {
              "type": "array",
              "items": {
                "type": "object",
                "additionalProperties": false,
                "required": ["country", "capital"],
                "properties": {
                  "country": {"type": "string"},
                  "capital": {"type": "string"}
                }
              }
            }
          }
        }
      }
    }
  }'

Expected content[0].text is a JSON string parseable into the schema, e.g.:

{"capitals":[
  {"country":"France","capital":"Paris"},
  {"country":"Germany","capital":"Berlin"},
  {"country":"Italy","capital":"Rome"}
]}

Without the plugin you'd get a Markdown bullet list instead.

What the plugin does (and doesn't)

• ✅ Translates output_config.format (GA) and output_format (beta) into OpenAI response_format.
• ✅ Supports json_schema, json_object, and text types.
• ✅ Defaults strict: true to mirror Anthropic's guarantee semantics; the client can opt out with strict: false.
• ❌ Does not validate that the backend you've configured supports response_format. Most OpenAI-compatible backends do; some don't.
• ❌ Does not modify response handling. CCR's existing OpenAI → Anthropic response converter wraps the JSON string into content[0].text, which is exactly what Anthropic's GA spec returns.

How it works

CCR's built-in AnthropicTransformer.transformRequestOut is called first on the incoming /v1/messages request and produces a UnifiedChatRequest in OpenAI shape — but it strips output_config / output_format along the way. Provider-level transformRequestIn plugins run next, and CCR passes them a context.req.body reference that still contains the original Anthropic-shaped request body.

This plugin reads context.req.body.output_config?.format (or context.req.body.output_format), translates it into the OpenAI response_format shape, and attaches it to the outgoing request body. That's it — about thirty lines of code.

License

MIT