OpenRouter reasoningEffort injection blocks x-ai/grok-4.20-multi-agent

[KHAEntertainment]

Problem

All calls to x-ai/grok-4.20-multi-agent via OpenRouter fail with:

Model grok-4.20-multi-agent-0309 does not support parameter reasoningEffort.

Root Cause

OpenRouter's proxy infrastructure injects reasoningEffort (underscore format) at the infrastructure layer for all calls to x-ai/grok-4.20-multi-agent. This injection happens regardless of what the client sends — it is NOT triggered by extra_body or any client-side parameter.

xAI's Grok model only accepts reasoning.effort (dot notation, nested object). OpenRouter's injection uses the wrong format.

What Was Tested

Client-side approaches (all failed)

Approach	Method	Result
No reasoning param	max_tokens, temperature only	❌ OpenRouter injects reasoningEffort anyway
reasoning: { effort: "low" }	httpx raw HTTP POST	❌ OpenRouter still injects reasoningEffort
reasoningEffort: "medium"	httpx raw HTTP POST	❌ OpenRouter double-injects
extra_body: { reasoningEffort: "low" }	httpx raw HTTP POST	❌ Still fails
urllib.request (no httpx)	raw HTTP	❌ Same result
OpenAI SDK extra_body={}	SDK	❌ Same result

OpenRouter's own documentation at https://openrouter.ai/models/x-ai/grok-4.20-multi-agent claims:

• supports_reasoning_effort: true
• supported_reasoning_efforts: ["low", "medium", "high", "xhigh"]

But the actual API returns the reasoningEffort error.

Positive findings

• xAI direct (api.x.ai/v1/responses) accepts reasoning: { effort: "low" } natively — works correctly with xAI API key
• OpenRouter key valid — tested with anthropic/claude-3.5-haiku, returns 200
• grok-4 (non-multi-agent) works on OpenRouter without any reasoning parameter

Investigation Date

2026-04-12

Files Changed

The codebase was updated to use httpx for raw HTTP control (rather than OpenAI SDK) in preparation for sending reasoning.effort directly. This change is correct but insufficient — the injection happens at OpenRouter's proxy layer.

Changed files:

• src/bridge/grok_bridge.py — httpx for OpenRouter, OpenAI SDK responses API for xAI direct
• platforms/claude/src/bridge/grok_bridge.py — mirror
• skills/grok-refactor/bridge/grok_bridge.py — mirror
• .mcp.json files — absolute venv Python path

Proposed Solutions

1. OpenRouter fix: Report to OpenRouter that reasoningEffort injection is broken for x-ai/grok-4.20-multi-agent. They need to either stop injecting or support reasoning.effort (dot notation) in their injection layer.

2. Alternative model: Use x-ai/grok-4 (non-multi-agent) on OpenRouter — this works but lacks the multi-agent swarm capability.

3. xAI direct: Use xAI API key directly at api.x.ai/v1/responses — works but requires an xAI API key (not an OpenRouter key).

4. Proxy stripping (untested): A middleware/proxy that strips reasoningEffort from OpenRouter's response before it reaches the client would not help — the error comes from xAI's rejection of OpenRouter's injected parameter, which happens server-side at xAI.

Reproduction Script

import httpx, json, os

key = os.environ["OPENROUTER_API_KEY"]
client = httpx.Client(
    base_url="https://openrouter.ai/api/v1",
    headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json"},
    timeout=30,
)

body = {
    "model": "x-ai/grok-4.20-multi-agent",
    "messages": [{"role": "system", "content": "You are Grok 4.20."},
                 {"role": "user", "content": "Say hi in one word."}],
    "max_tokens": 16384,
    "temperature": 0.3,
    # No reasoning param — but OpenRouter injects reasoningEffort anyway
}
resp = client.post("/chat/completions", json=body)
# Status 400: Model grok-4.20-multi-agent-0309 does not support parameter reasoningEffort.

Environment

• Python 3.14 (venv at .venv/)
• httpx 0.28.1
• openai 1.x
• OpenRouter API key: sk-or-v1-1... (73 chars, OpenRouter format)