Bug type

Regression (worked before, now fails)

Summary

/new injects a hardcoded reset prompt containing the phrase:

Execute your Session Startup sequence now - read the required files before responding to the user.

On Azure OpenAI deployment , this phrase can trigger 400 content_filter before any tool execution.

Trigger phrase identified

The 400 is triggered by this exact sentence injected on /new:

Execute your Session Startup sequence now - read the required files before responding to the user.

A/B check on the same environment/model:

• Fails (400): includes Execute your Session Startup sequence now ...
• Passes: replace only that sentence with
  Please start a fresh session now and read the required files before responding to the user.

Steps to reproduce

1. Configure OpenClaw to use provider azure-openai
2. Start gateway service.
3. Trigger /new on main-bot (Telegram path or equivalent gateway path).
4. Check the created session log and gateway log.

Expected behavior

/new should consistently start a fresh session and return a normal greeting.

Actual behavior

Intermittently/consistently receives:

400 The response was filtered due to the prompt triggering Azure OpenAI's content management policy...

The failure happens at startup prompt stage (before tools run).

OpenClaw version

2026.3.7 and 2026.3.8 (after upgrade, issue recurred)

Operating system

Debian GNU/Linux 11 (bullseye), x86_64

Install method

Global npm install (npm install -g openclaw) with systemd user service (openclaw-gateway.service).

Model

gpt-5.4

Provider / routing chain

Azure OpenAI

Config file / key location

No response

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Redacted evidence available:

- A failing `/new` session where injected prompt includes:
    `A new session was started via /new or /reset. Execute your Session Startup sequence now ...`
- Gateway log shows:
    `embedded run agent end ... isError=true error=400 The response was filtered ...`
- Context diagnostics show failure occurs at pre-prompt/agent-start phase for startup turn.

Impact and severity

• Affected: users relying on /new in production chat flow.
• Severity: blocks normal session reset/startup.
• Frequency: reproducible on affected provider/model combo.
• Consequence: failed onboarding/startup turns, degraded reliability, manual recovery need

Additional information

No response