bug: LocalClient sends Ollama-native POST /api/generate format to LM Studio / OpenAI-compatible servers

[JLay2026]

Description

The LocalClient provider always sends requests in the Ollama-native format — a flat prompt string to POST /api/generate. This format is specific to Ollama and breaks with any other local inference server that uses the OpenAI-compatible API (POST /v1/chat/completions with a messages array).

Affected servers: LM Studio, vLLM, LocalAI, llama.cpp server, Llamafile, and any other OpenAI-compatible backend.

---

Steps to Reproduce

1. Install LM Studio and start a local server on e.g. http://192.168.x.x:1234/v1
2. In AI Agent HA, set AI Provider → Local Model, Local API URL → http://192.168.x.x:1234/v1
3. Send any prompt from the panel

Result: LM Studio returns an error:

Unexpected endpoint or method (POST /)

HA logs show the request was sent as a flat prompt string to POST /api/generate (Ollama format) rather than the OpenAI-compatible POST /v1/chat/completions with a messages array.

Expected: Requests to /v1-style endpoints should use the OpenAI-compatible format automatically.

---

Root Cause

LocalClient.get_response unconditionally builds an Ollama payload and sends it to whatever URL the user configured, with no inspection of the URL or detection of the target server type:

# agent.py — LocalClient (current)
payload = {
    "model": self.model,
    "prompt": user_message,   # Ollama-native flat string
    "stream": False,
}
async with session.post(self.url, ...)  # blindly posts to configured URL

LM Studio and OpenAI-compatible servers expect:

payload = {
    "model": self.model,
    "messages": messages,     # OpenAI-style messages array
    "temperature": 0.7,
    "stream": False,
}
# POST to /v1/chat/completions, not /api/generate

---

Fix

URL-based auto-detection in LocalClient.__init__: if the configured URL contains /v1, use the OpenAI-compatible messages format; otherwise fall back to the existing Ollama-native prompt format. No configuration change required from users — the URL they already enter determines the path.

local_url	Request format
http://host:1234/v1	OpenAI-compatible (POST /v1/chat/completions, messages array)
http://host:11434/api/generate	Ollama-native (prompt string) — unchanged

A working implementation is available in my fork: JLay2026/ai_agent_ha PR #47

That PR also fixes a related issue: the frontend panel had a 60-second hardcoded timeout, while local models on consumer hardware commonly take 60–180 seconds — causing false Request timed out errors even when the backend was still processing. The fix raises the frontend timeout to 300 seconds, matching the aiohttp.ClientTimeout(total=300) already set on all backend clients.

---

Workaround (until merged)

Use the OpenAI provider instead of Local Model, with:

• API Key: any non-empty string (e.g. lmstudio)
• Custom Base URL: http://<lmstudio-host>:1234/v1

This routes through OpenAIClient, which already sends the correct format.

---

Environment

• AI Agent HA v1.08.7
• LM Studio 0.3.x
• Home Assistant 2025.x

[sbenodiz]

@JLay2026 can you please share your linkedin or any other method i can chat with you. it's great you contribue but your account just opened few days ago