Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

When using OpenClaw as an OpenAI-compatible backend for Open-WebUI, each user message creates a new session with a random UUID (agent:< agentId>:openai:<uuid>), instead of reusing a single persistent session for the chat. This results in:

• Massive session proliferation (44 sessions from a single conversation)
• No conversation continuity — each turn starts fresh
• WebUI works around this by injecting [Chat messages since your last reply - for context] into each message, but this is fragile and burn s tokens
• Session store bloat and wasted disk space

Steps to reproduce

1. Configure OpenClaw with an OpenAI-compatible provider (e.g., llama.cpp on port 11437)
2. Set up Open-WebUI to use the OpenClaw gateway as its OpenAI backend
3. Start a conversation in Open-WebUI
4. Send multiple messages
5. Check ~/.openclaw/agents/<agentId>/sessions/sessions.json — each message created a separate agent:<agentId>:openai:<uuid> session

Expected behavior

• All messages in a single WebUI chat should route to the same session
• Session key should be derived from the WebUI chat/conversation ID, not randomly generated per request
• Conversation history should persist naturally in the session transcript

Actual behavior

• Each message creates a new session with key agent:<agentId>:openai:<random-uuid>
• Each session is independent — no shared context between turns
• Sessions are marked done after each turn
• WebUI compensates by injecting prior conversation as context in each new message

OpenClaw version

2026.5.2

Operating system

EL10

Install method

npm

Model

qwen3.6:27b

Provider / routing chain

Open-WebUI -> openclaw -> llama.cpp

Additional provider/model setup details

No response

Logs, screenshots, and evidence

## Evidence

Session keys observed:

agent:sky:openai:17a95d23-8031-46cd-926c-9640e6ead873
agent:sky:openai:525e5460-bbc4-47ff-ad2e-122fe8565b0e
agent:sky:openai:db5dcd06-a0ea-4b2b-9064-e5248ad4c081
... (44 total from one conversation)


Each session's first user message contains WebUI's workaround:

[Chat messages since your last reply - for context]
User: Hello?
Assistant: Hey Wurdzy. 🦞

I'm here. Ready when you are.
User: <actual new message>


## Environment

- OpenClaw version: 2026.5.2 (observed), still present in 2026.5.4
- OS: Linux (EL10, x86_64)
- Provider: llama.cpp via OpenAI-compatible API (openai-completions)
- Frontend: Open-WebUI connecting via OpenAI API to OpenClaw gateway

## Root Cause Hypothesis

The OpenAI-compatible endpoint in the gateway generates a new session key per request instead of extracting and reusing a conversation/thread ID from the incoming request. Open-WebUI may not be sending a consistent `thread_id` or `metadata` that OpenClaw can use to correlate requests to the same conversation.

Possible fixes:
1. OpenClaw could use a stable identifier from the request (e.g., `X-Chat-Id` header, or a consistent `metadata` field)
2. OpenClaw could fall back to a single persistent `agent:<agentId>:openai:default` session when no conversation ID is provided
3. OpenClaw could hash the requesting IP + user agent as a session key for stateless clients

## Impact

- **Token waste**: System prompt + full context re-sent every turn
- **Session store bloat**: Dozens of dead sessions per conversation
- **No memory continuity**: Agent cannot maintain state between turns
- **Workaround fragility**: WebUI's context injection is lossy and doesn't include tool calls or internal state

Impact and severity

No response

Additional information

No response